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 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
643 /* Start listener process (exact function is dependent on the
644 * implementation environment--kernel or user space) */
649 rx_atomic_clear_bit(&rxinit_status, 0);
656 return rx_InitHost(htonl(INADDR_ANY), port);
662 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
663 * maintaing the round trip timer.
668 * Start a new RTT timer for a given call and packet.
670 * There must be no resendEvent already listed for this call, otherwise this
671 * will leak events - intended for internal use within the RTO code only
674 * the RX call to start the timer for
675 * @param[in] lastPacket
676 * a flag indicating whether the last packet has been sent or not
678 * @pre call must be locked before calling this function
682 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
684 struct clock now, retryTime;
689 clock_Add(&retryTime, &call->rto);
691 /* If we're sending the last packet, and we're the client, then the server
692 * may wait for an additional 400ms before returning the ACK, wait for it
693 * rather than hitting a timeout */
694 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
695 clock_Addmsec(&retryTime, 400);
697 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
698 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
703 * Cancel an RTT timer for a given call.
707 * the RX call to cancel the timer for
709 * @pre call must be locked before calling this function
714 rxi_rto_cancel(struct rx_call *call)
716 rxevent_Cancel(&call->resendEvent);
717 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
721 * Tell the RTO timer that we have sent a packet.
723 * If the timer isn't already running, then start it. If the timer is running,
727 * the RX call that the packet has been sent on
728 * @param[in] lastPacket
729 * A flag which is true if this is the last packet for the call
731 * @pre The call must be locked before calling this function
736 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
738 if (call->resendEvent)
741 rxi_rto_startTimer(call, lastPacket, istack);
745 * Tell the RTO timer that we have received an new ACK message
747 * This function should be called whenever a call receives an ACK that
748 * acknowledges new packets. Whatever happens, we stop the current timer.
749 * If there are unacked packets in the queue which have been sent, then
750 * we restart the timer from now. Otherwise, we leave it stopped.
753 * the RX call that the ACK has been received on
757 rxi_rto_packet_acked(struct rx_call *call, int istack)
759 struct opr_queue *cursor;
761 rxi_rto_cancel(call);
763 if (opr_queue_IsEmpty(&call->tq))
766 for (opr_queue_Scan(&call->tq, cursor)) {
767 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
768 if (p->header.seq > call->tfirst + call->twind)
771 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
772 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
780 * Set an initial round trip timeout for a peer connection
782 * @param[in] secs The timeout to set in seconds
786 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
787 peer->rtt = secs * 8000;
791 * Enables or disables the busy call channel error (RX_CALL_BUSY).
793 * @param[in] onoff Non-zero to enable busy call channel errors.
795 * @pre Neither rx_Init nor rx_InitHost have been called yet
798 rx_SetBusyChannelError(afs_int32 onoff)
800 osi_Assert(rx_atomic_test_bit(&rxinit_status, 0));
801 rxi_busyChannelError = onoff ? 1 : 0;
805 * Set a delayed ack event on the specified call for the given time
807 * @param[in] call - the call on which to set the event
808 * @param[in] offset - the delay from now after which the event fires
811 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
813 struct clock now, when;
817 clock_Add(&when, offset);
819 if (call->delayedAckEvent && clock_Gt(&call->delayedAckTime, &when)) {
820 /* The event we're cancelling already has a reference, so we don't
822 rxevent_Cancel(&call->delayedAckEvent);
823 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
826 call->delayedAckTime = when;
827 } else if (!call->delayedAckEvent) {
828 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
829 call->delayedAckEvent = rxevent_Post(&when, &now,
832 call->delayedAckTime = when;
837 rxi_CancelDelayedAckEvent(struct rx_call *call)
839 if (call->delayedAckEvent) {
840 rxevent_Cancel(&call->delayedAckEvent);
841 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
845 /* called with unincremented nRequestsRunning to see if it is OK to start
846 * a new thread in this service. Could be "no" for two reasons: over the
847 * max quota, or would prevent others from reaching their min quota.
849 #ifdef RX_ENABLE_LOCKS
850 /* This verion of QuotaOK reserves quota if it's ok while the
851 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
854 QuotaOK(struct rx_service *aservice)
856 /* check if over max quota */
857 if (aservice->nRequestsRunning >= aservice->maxProcs) {
861 /* under min quota, we're OK */
862 /* otherwise, can use only if there are enough to allow everyone
863 * to go to their min quota after this guy starts.
866 MUTEX_ENTER(&rx_quota_mutex);
867 if ((aservice->nRequestsRunning < aservice->minProcs)
868 || (rxi_availProcs > rxi_minDeficit)) {
869 aservice->nRequestsRunning++;
870 /* just started call in minProcs pool, need fewer to maintain
872 if (aservice->nRequestsRunning <= aservice->minProcs)
875 MUTEX_EXIT(&rx_quota_mutex);
878 MUTEX_EXIT(&rx_quota_mutex);
884 ReturnToServerPool(struct rx_service *aservice)
886 aservice->nRequestsRunning--;
887 MUTEX_ENTER(&rx_quota_mutex);
888 if (aservice->nRequestsRunning < aservice->minProcs)
891 MUTEX_EXIT(&rx_quota_mutex);
894 #else /* RX_ENABLE_LOCKS */
896 QuotaOK(struct rx_service *aservice)
899 /* under min quota, we're OK */
900 if (aservice->nRequestsRunning < aservice->minProcs)
903 /* check if over max quota */
904 if (aservice->nRequestsRunning >= aservice->maxProcs)
907 /* otherwise, can use only if there are enough to allow everyone
908 * to go to their min quota after this guy starts.
910 MUTEX_ENTER(&rx_quota_mutex);
911 if (rxi_availProcs > rxi_minDeficit)
913 MUTEX_EXIT(&rx_quota_mutex);
916 #endif /* RX_ENABLE_LOCKS */
919 /* Called by rx_StartServer to start up lwp's to service calls.
920 NExistingProcs gives the number of procs already existing, and which
921 therefore needn't be created. */
923 rxi_StartServerProcs(int nExistingProcs)
925 struct rx_service *service;
930 /* For each service, reserve N processes, where N is the "minimum"
931 * number of processes that MUST be able to execute a request in parallel,
932 * at any time, for that process. Also compute the maximum difference
933 * between any service's maximum number of processes that can run
934 * (i.e. the maximum number that ever will be run, and a guarantee
935 * that this number will run if other services aren't running), and its
936 * minimum number. The result is the extra number of processes that
937 * we need in order to provide the latter guarantee */
938 for (i = 0; i < RX_MAX_SERVICES; i++) {
940 service = rx_services[i];
941 if (service == (struct rx_service *)0)
943 nProcs += service->minProcs;
944 diff = service->maxProcs - service->minProcs;
948 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
949 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
950 for (i = 0; i < nProcs; i++) {
951 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
957 /* This routine is only required on Windows */
959 rx_StartClientThread(void)
961 #ifdef AFS_PTHREAD_ENV
963 pid = pthread_self();
964 #endif /* AFS_PTHREAD_ENV */
966 #endif /* AFS_NT40_ENV */
968 /* This routine must be called if any services are exported. If the
969 * donateMe flag is set, the calling process is donated to the server
972 rx_StartServer(int donateMe)
974 struct rx_service *service;
980 /* Start server processes, if necessary (exact function is dependent
981 * on the implementation environment--kernel or user space). DonateMe
982 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
983 * case, one less new proc will be created rx_StartServerProcs.
985 rxi_StartServerProcs(donateMe);
987 /* count up the # of threads in minProcs, and add set the min deficit to
988 * be that value, too.
990 for (i = 0; i < RX_MAX_SERVICES; i++) {
991 service = rx_services[i];
992 if (service == (struct rx_service *)0)
994 MUTEX_ENTER(&rx_quota_mutex);
995 rxi_totalMin += service->minProcs;
996 /* below works even if a thread is running, since minDeficit would
997 * still have been decremented and later re-incremented.
999 rxi_minDeficit += service->minProcs;
1000 MUTEX_EXIT(&rx_quota_mutex);
1003 /* Turn on reaping of idle server connections */
1004 rxi_ReapConnections(NULL, NULL, NULL, 0);
1009 #ifndef AFS_NT40_ENV
1013 #ifdef AFS_PTHREAD_ENV
1015 pid = afs_pointer_to_int(pthread_self());
1016 #else /* AFS_PTHREAD_ENV */
1018 LWP_CurrentProcess(&pid);
1019 #endif /* AFS_PTHREAD_ENV */
1021 sprintf(name, "srv_%d", ++nProcs);
1022 if (registerProgram)
1023 (*registerProgram) (pid, name);
1025 #endif /* AFS_NT40_ENV */
1026 rx_ServerProc(NULL); /* Never returns */
1028 #ifdef RX_ENABLE_TSFPQ
1029 /* no use leaving packets around in this thread's local queue if
1030 * it isn't getting donated to the server thread pool.
1032 rxi_FlushLocalPacketsTSFPQ();
1033 #endif /* RX_ENABLE_TSFPQ */
1037 /* Create a new client connection to the specified service, using the
1038 * specified security object to implement the security model for this
1040 struct rx_connection *
1041 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1042 struct rx_securityClass *securityObject,
1043 int serviceSecurityIndex)
1047 struct rx_connection *conn;
1052 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1053 "serviceSecurityIndex %d)\n",
1054 ntohl(shost), ntohs(sport), sservice, securityObject,
1055 serviceSecurityIndex));
1057 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1058 * the case of kmem_alloc? */
1059 conn = rxi_AllocConnection();
1060 #ifdef RX_ENABLE_LOCKS
1061 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1062 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1063 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1066 MUTEX_ENTER(&rx_connHashTable_lock);
1067 cid = (rx_nextCid += RX_MAXCALLS);
1068 conn->type = RX_CLIENT_CONNECTION;
1070 conn->epoch = rx_epoch;
1071 conn->peer = rxi_FindPeer(shost, sport, 1);
1072 conn->serviceId = sservice;
1073 conn->securityObject = securityObject;
1074 conn->securityData = (void *) 0;
1075 conn->securityIndex = serviceSecurityIndex;
1076 rx_SetConnDeadTime(conn, rx_connDeadTime);
1077 rx_SetConnSecondsUntilNatPing(conn, 0);
1078 conn->ackRate = RX_FAST_ACK_RATE;
1079 conn->nSpecific = 0;
1080 conn->specific = NULL;
1081 conn->challengeEvent = NULL;
1082 conn->delayedAbortEvent = NULL;
1083 conn->abortCount = 0;
1085 for (i = 0; i < RX_MAXCALLS; i++) {
1086 conn->twind[i] = rx_initSendWindow;
1087 conn->rwind[i] = rx_initReceiveWindow;
1088 conn->lastBusy[i] = 0;
1091 RXS_NewConnection(securityObject, conn);
1093 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1095 conn->refCount++; /* no lock required since only this thread knows... */
1096 conn->next = rx_connHashTable[hashindex];
1097 rx_connHashTable[hashindex] = conn;
1098 if (rx_stats_active)
1099 rx_atomic_inc(&rx_stats.nClientConns);
1100 MUTEX_EXIT(&rx_connHashTable_lock);
1106 * Ensure a connection's timeout values are valid.
1108 * @param[in] conn The connection to check
1110 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1111 * unless idleDeadTime and/or hardDeadTime are not set
1115 rxi_CheckConnTimeouts(struct rx_connection *conn)
1117 /* a connection's timeouts must have the relationship
1118 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1119 * total loss of network to a peer may cause an idle timeout instead of a
1120 * dead timeout, simply because the idle timeout gets hit first. Also set
1121 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1122 /* this logic is slightly complicated by the fact that
1123 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1125 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1126 if (conn->idleDeadTime) {
1127 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1129 if (conn->hardDeadTime) {
1130 if (conn->idleDeadTime) {
1131 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1133 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1139 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1141 /* The idea is to set the dead time to a value that allows several
1142 * keepalives to be dropped without timing out the connection. */
1143 conn->secondsUntilDead = seconds;
1144 rxi_CheckConnTimeouts(conn);
1145 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1149 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1151 conn->hardDeadTime = seconds;
1152 rxi_CheckConnTimeouts(conn);
1156 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1158 conn->idleDeadTime = seconds;
1159 conn->idleDeadDetection = (seconds ? 1 : 0);
1160 rxi_CheckConnTimeouts(conn);
1163 int rxi_lowPeerRefCount = 0;
1164 int rxi_lowConnRefCount = 0;
1167 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1168 * NOTE: must not be called with rx_connHashTable_lock held.
1171 rxi_CleanupConnection(struct rx_connection *conn)
1173 /* Notify the service exporter, if requested, that this connection
1174 * is being destroyed */
1175 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1176 (*conn->service->destroyConnProc) (conn);
1178 /* Notify the security module that this connection is being destroyed */
1179 RXS_DestroyConnection(conn->securityObject, conn);
1181 /* If this is the last connection using the rx_peer struct, set its
1182 * idle time to now. rxi_ReapConnections will reap it if it's still
1183 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1185 MUTEX_ENTER(&rx_peerHashTable_lock);
1186 if (conn->peer->refCount < 2) {
1187 conn->peer->idleWhen = clock_Sec();
1188 if (conn->peer->refCount < 1) {
1189 conn->peer->refCount = 1;
1190 if (rx_stats_active) {
1191 MUTEX_ENTER(&rx_stats_mutex);
1192 rxi_lowPeerRefCount++;
1193 MUTEX_EXIT(&rx_stats_mutex);
1197 conn->peer->refCount--;
1198 MUTEX_EXIT(&rx_peerHashTable_lock);
1200 if (rx_stats_active)
1202 if (conn->type == RX_SERVER_CONNECTION)
1203 rx_atomic_dec(&rx_stats.nServerConns);
1205 rx_atomic_dec(&rx_stats.nClientConns);
1208 if (conn->specific) {
1210 for (i = 0; i < conn->nSpecific; i++) {
1211 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1212 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1213 conn->specific[i] = NULL;
1215 free(conn->specific);
1217 conn->specific = NULL;
1218 conn->nSpecific = 0;
1219 #endif /* !KERNEL */
1221 MUTEX_DESTROY(&conn->conn_call_lock);
1222 MUTEX_DESTROY(&conn->conn_data_lock);
1223 CV_DESTROY(&conn->conn_call_cv);
1225 rxi_FreeConnection(conn);
1228 /* Destroy the specified connection */
1230 rxi_DestroyConnection(struct rx_connection *conn)
1232 MUTEX_ENTER(&rx_connHashTable_lock);
1233 rxi_DestroyConnectionNoLock(conn);
1234 /* conn should be at the head of the cleanup list */
1235 if (conn == rx_connCleanup_list) {
1236 rx_connCleanup_list = rx_connCleanup_list->next;
1237 MUTEX_EXIT(&rx_connHashTable_lock);
1238 rxi_CleanupConnection(conn);
1240 #ifdef RX_ENABLE_LOCKS
1242 MUTEX_EXIT(&rx_connHashTable_lock);
1244 #endif /* RX_ENABLE_LOCKS */
1248 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1250 struct rx_connection **conn_ptr;
1252 struct rx_packet *packet;
1259 MUTEX_ENTER(&conn->conn_data_lock);
1260 MUTEX_ENTER(&rx_refcnt_mutex);
1261 if (conn->refCount > 0)
1264 if (rx_stats_active) {
1265 MUTEX_ENTER(&rx_stats_mutex);
1266 rxi_lowConnRefCount++;
1267 MUTEX_EXIT(&rx_stats_mutex);
1271 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1272 /* Busy; wait till the last guy before proceeding */
1273 MUTEX_EXIT(&rx_refcnt_mutex);
1274 MUTEX_EXIT(&conn->conn_data_lock);
1279 /* If the client previously called rx_NewCall, but it is still
1280 * waiting, treat this as a running call, and wait to destroy the
1281 * connection later when the call completes. */
1282 if ((conn->type == RX_CLIENT_CONNECTION)
1283 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1284 conn->flags |= RX_CONN_DESTROY_ME;
1285 MUTEX_EXIT(&conn->conn_data_lock);
1289 MUTEX_EXIT(&rx_refcnt_mutex);
1290 MUTEX_EXIT(&conn->conn_data_lock);
1292 /* Check for extant references to this connection */
1293 MUTEX_ENTER(&conn->conn_call_lock);
1294 for (i = 0; i < RX_MAXCALLS; i++) {
1295 struct rx_call *call = conn->call[i];
1298 if (conn->type == RX_CLIENT_CONNECTION) {
1299 MUTEX_ENTER(&call->lock);
1300 if (call->delayedAckEvent) {
1301 /* Push the final acknowledgment out now--there
1302 * won't be a subsequent call to acknowledge the
1303 * last reply packets */
1304 rxi_CancelDelayedAckEvent(call);
1305 if (call->state == RX_STATE_PRECALL
1306 || call->state == RX_STATE_ACTIVE) {
1307 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1312 MUTEX_EXIT(&call->lock);
1316 MUTEX_EXIT(&conn->conn_call_lock);
1318 #ifdef RX_ENABLE_LOCKS
1320 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1321 MUTEX_EXIT(&conn->conn_data_lock);
1323 /* Someone is accessing a packet right now. */
1327 #endif /* RX_ENABLE_LOCKS */
1330 /* Don't destroy the connection if there are any call
1331 * structures still in use */
1332 MUTEX_ENTER(&conn->conn_data_lock);
1333 conn->flags |= RX_CONN_DESTROY_ME;
1334 MUTEX_EXIT(&conn->conn_data_lock);
1339 if (conn->natKeepAliveEvent) {
1340 rxi_NatKeepAliveOff(conn);
1343 if (conn->delayedAbortEvent) {
1344 rxevent_Cancel(&conn->delayedAbortEvent);
1345 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1347 MUTEX_ENTER(&conn->conn_data_lock);
1348 rxi_SendConnectionAbort(conn, packet, 0, 1);
1349 MUTEX_EXIT(&conn->conn_data_lock);
1350 rxi_FreePacket(packet);
1354 /* Remove from connection hash table before proceeding */
1356 &rx_connHashTable[CONN_HASH
1357 (peer->host, peer->port, conn->cid, conn->epoch,
1359 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1360 if (*conn_ptr == conn) {
1361 *conn_ptr = conn->next;
1365 /* if the conn that we are destroying was the last connection, then we
1366 * clear rxLastConn as well */
1367 if (rxLastConn == conn)
1370 /* Make sure the connection is completely reset before deleting it. */
1371 /* get rid of pending events that could zap us later */
1372 rxevent_Cancel(&conn->challengeEvent);
1373 rxevent_Cancel(&conn->checkReachEvent);
1374 rxevent_Cancel(&conn->natKeepAliveEvent);
1376 /* Add the connection to the list of destroyed connections that
1377 * need to be cleaned up. This is necessary to avoid deadlocks
1378 * in the routines we call to inform others that this connection is
1379 * being destroyed. */
1380 conn->next = rx_connCleanup_list;
1381 rx_connCleanup_list = conn;
1384 /* Externally available version */
1386 rx_DestroyConnection(struct rx_connection *conn)
1391 rxi_DestroyConnection(conn);
1396 rx_GetConnection(struct rx_connection *conn)
1401 MUTEX_ENTER(&rx_refcnt_mutex);
1403 MUTEX_EXIT(&rx_refcnt_mutex);
1407 #ifdef RX_ENABLE_LOCKS
1408 /* Wait for the transmit queue to no longer be busy.
1409 * requires the call->lock to be held */
1411 rxi_WaitforTQBusy(struct rx_call *call) {
1412 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1413 call->flags |= RX_CALL_TQ_WAIT;
1415 MUTEX_ASSERT(&call->lock);
1416 CV_WAIT(&call->cv_tq, &call->lock);
1418 if (call->tqWaiters == 0) {
1419 call->flags &= ~RX_CALL_TQ_WAIT;
1426 rxi_WakeUpTransmitQueue(struct rx_call *call)
1428 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1429 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1430 call, call->tqWaiters, call->flags));
1431 #ifdef RX_ENABLE_LOCKS
1432 MUTEX_ASSERT(&call->lock);
1433 CV_BROADCAST(&call->cv_tq);
1434 #else /* RX_ENABLE_LOCKS */
1435 osi_rxWakeup(&call->tq);
1436 #endif /* RX_ENABLE_LOCKS */
1440 /* Start a new rx remote procedure call, on the specified connection.
1441 * If wait is set to 1, wait for a free call channel; otherwise return
1442 * 0. Maxtime gives the maximum number of seconds this call may take,
1443 * after rx_NewCall returns. After this time interval, a call to any
1444 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1445 * For fine grain locking, we hold the conn_call_lock in order to
1446 * to ensure that we don't get signalle after we found a call in an active
1447 * state and before we go to sleep.
1450 rx_NewCall(struct rx_connection *conn)
1452 int i, wait, ignoreBusy = 1;
1453 struct rx_call *call;
1454 struct clock queueTime;
1455 afs_uint32 leastBusy = 0;
1459 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1462 clock_GetTime(&queueTime);
1464 * Check if there are others waiting for a new call.
1465 * If so, let them go first to avoid starving them.
1466 * This is a fairly simple scheme, and might not be
1467 * a complete solution for large numbers of waiters.
1469 * makeCallWaiters keeps track of the number of
1470 * threads waiting to make calls and the
1471 * RX_CONN_MAKECALL_WAITING flag bit is used to
1472 * indicate that there are indeed calls waiting.
1473 * The flag is set when the waiter is incremented.
1474 * It is only cleared when makeCallWaiters is 0.
1475 * This prevents us from accidently destroying the
1476 * connection while it is potentially about to be used.
1478 MUTEX_ENTER(&conn->conn_call_lock);
1479 MUTEX_ENTER(&conn->conn_data_lock);
1480 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1481 conn->flags |= RX_CONN_MAKECALL_WAITING;
1482 conn->makeCallWaiters++;
1483 MUTEX_EXIT(&conn->conn_data_lock);
1485 #ifdef RX_ENABLE_LOCKS
1486 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1490 MUTEX_ENTER(&conn->conn_data_lock);
1491 conn->makeCallWaiters--;
1492 if (conn->makeCallWaiters == 0)
1493 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1496 /* We are now the active thread in rx_NewCall */
1497 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1498 MUTEX_EXIT(&conn->conn_data_lock);
1503 for (i = 0; i < RX_MAXCALLS; i++) {
1504 call = conn->call[i];
1506 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1507 /* we're not ignoring busy call slots; only look at the
1508 * call slot that is the "least" busy */
1512 if (call->state == RX_STATE_DALLY) {
1513 MUTEX_ENTER(&call->lock);
1514 if (call->state == RX_STATE_DALLY) {
1515 if (ignoreBusy && conn->lastBusy[i]) {
1516 /* if we're ignoring busy call slots, skip any ones that
1517 * have lastBusy set */
1518 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1519 leastBusy = conn->lastBusy[i];
1521 MUTEX_EXIT(&call->lock);
1526 * We are setting the state to RX_STATE_RESET to
1527 * ensure that no one else will attempt to use this
1528 * call once we drop the conn->conn_call_lock and
1529 * call->lock. We must drop the conn->conn_call_lock
1530 * before calling rxi_ResetCall because the process
1531 * of clearing the transmit queue can block for an
1532 * extended period of time. If we block while holding
1533 * the conn->conn_call_lock, then all rx_EndCall
1534 * processing will block as well. This has a detrimental
1535 * effect on overall system performance.
1537 call->state = RX_STATE_RESET;
1538 (*call->callNumber)++;
1539 MUTEX_EXIT(&conn->conn_call_lock);
1540 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1541 rxi_ResetCall(call, 0);
1542 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1546 * If we failed to be able to safely obtain the
1547 * conn->conn_call_lock we will have to drop the
1548 * call->lock to avoid a deadlock. When the call->lock
1549 * is released the state of the call can change. If it
1550 * is no longer RX_STATE_RESET then some other thread is
1553 MUTEX_EXIT(&call->lock);
1554 MUTEX_ENTER(&conn->conn_call_lock);
1555 MUTEX_ENTER(&call->lock);
1557 if (call->state == RX_STATE_RESET)
1561 * If we get here it means that after dropping
1562 * the conn->conn_call_lock and call->lock that
1563 * the call is no longer ours. If we can't find
1564 * a free call in the remaining slots we should
1565 * not go immediately to RX_CONN_MAKECALL_WAITING
1566 * because by dropping the conn->conn_call_lock
1567 * we have given up synchronization with rx_EndCall.
1568 * Instead, cycle through one more time to see if
1569 * we can find a call that can call our own.
1571 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1574 MUTEX_EXIT(&call->lock);
1577 if (ignoreBusy && conn->lastBusy[i]) {
1578 /* if we're ignoring busy call slots, skip any ones that
1579 * have lastBusy set */
1580 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1581 leastBusy = conn->lastBusy[i];
1586 /* rxi_NewCall returns with mutex locked */
1587 call = rxi_NewCall(conn, i);
1588 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1592 if (i < RX_MAXCALLS) {
1593 conn->lastBusy[i] = 0;
1594 call->flags &= ~RX_CALL_PEER_BUSY;
1599 if (leastBusy && ignoreBusy) {
1600 /* we didn't find a useable call slot, but we did see at least one
1601 * 'busy' slot; look again and only use a slot with the 'least
1607 MUTEX_ENTER(&conn->conn_data_lock);
1608 conn->flags |= RX_CONN_MAKECALL_WAITING;
1609 conn->makeCallWaiters++;
1610 MUTEX_EXIT(&conn->conn_data_lock);
1612 #ifdef RX_ENABLE_LOCKS
1613 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1617 MUTEX_ENTER(&conn->conn_data_lock);
1618 conn->makeCallWaiters--;
1619 if (conn->makeCallWaiters == 0)
1620 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1621 MUTEX_EXIT(&conn->conn_data_lock);
1623 /* Client is initially in send mode */
1624 call->state = RX_STATE_ACTIVE;
1625 call->error = conn->error;
1627 call->app.mode = RX_MODE_ERROR;
1629 call->app.mode = RX_MODE_SENDING;
1631 #ifdef AFS_RXERRQ_ENV
1632 /* remember how many network errors the peer has when we started, so if
1633 * more errors are encountered after the call starts, we know the other endpoint won't be
1634 * responding to us */
1635 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1638 /* remember start time for call in case we have hard dead time limit */
1639 call->queueTime = queueTime;
1640 clock_GetTime(&call->startTime);
1641 call->app.bytesSent = 0;
1642 call->app.bytesRcvd = 0;
1644 /* Turn on busy protocol. */
1645 rxi_KeepAliveOn(call);
1647 /* Attempt MTU discovery */
1648 rxi_GrowMTUOn(call);
1651 * We are no longer the active thread in rx_NewCall
1653 MUTEX_ENTER(&conn->conn_data_lock);
1654 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1655 MUTEX_EXIT(&conn->conn_data_lock);
1658 * Wake up anyone else who might be giving us a chance to
1659 * run (see code above that avoids resource starvation).
1661 #ifdef RX_ENABLE_LOCKS
1662 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1663 osi_Panic("rx_NewCall call about to be used without an empty tq");
1666 CV_BROADCAST(&conn->conn_call_cv);
1670 MUTEX_EXIT(&conn->conn_call_lock);
1671 MUTEX_EXIT(&call->lock);
1674 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1679 rxi_HasActiveCalls(struct rx_connection *aconn)
1682 struct rx_call *tcall;
1686 for (i = 0; i < RX_MAXCALLS; i++) {
1687 if ((tcall = aconn->call[i])) {
1688 if ((tcall->state == RX_STATE_ACTIVE)
1689 || (tcall->state == RX_STATE_PRECALL)) {
1700 rxi_GetCallNumberVector(struct rx_connection *aconn,
1701 afs_int32 * aint32s)
1704 struct rx_call *tcall;
1708 MUTEX_ENTER(&aconn->conn_call_lock);
1709 for (i = 0; i < RX_MAXCALLS; i++) {
1710 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1711 aint32s[i] = aconn->callNumber[i] + 1;
1713 aint32s[i] = aconn->callNumber[i];
1715 MUTEX_EXIT(&aconn->conn_call_lock);
1721 rxi_SetCallNumberVector(struct rx_connection *aconn,
1722 afs_int32 * aint32s)
1725 struct rx_call *tcall;
1729 MUTEX_ENTER(&aconn->conn_call_lock);
1730 for (i = 0; i < RX_MAXCALLS; i++) {
1731 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1732 aconn->callNumber[i] = aint32s[i] - 1;
1734 aconn->callNumber[i] = aint32s[i];
1736 MUTEX_EXIT(&aconn->conn_call_lock);
1741 /* Advertise a new service. A service is named locally by a UDP port
1742 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1745 char *serviceName; Name for identification purposes (e.g. the
1746 service name might be used for probing for
1749 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1750 char *serviceName, struct rx_securityClass **securityObjects,
1751 int nSecurityObjects,
1752 afs_int32(*serviceProc) (struct rx_call * acall))
1754 osi_socket socket = OSI_NULLSOCKET;
1755 struct rx_service *tservice;
1761 if (serviceId == 0) {
1763 "rx_NewService: service id for service %s is not non-zero.\n",
1770 "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",
1778 tservice = rxi_AllocService();
1781 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1783 for (i = 0; i < RX_MAX_SERVICES; i++) {
1784 struct rx_service *service = rx_services[i];
1786 if (port == service->servicePort && host == service->serviceHost) {
1787 if (service->serviceId == serviceId) {
1788 /* The identical service has already been
1789 * installed; if the caller was intending to
1790 * change the security classes used by this
1791 * service, he/she loses. */
1793 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1794 serviceName, serviceId, service->serviceName);
1796 rxi_FreeService(tservice);
1799 /* Different service, same port: re-use the socket
1800 * which is bound to the same port */
1801 socket = service->socket;
1804 if (socket == OSI_NULLSOCKET) {
1805 /* If we don't already have a socket (from another
1806 * service on same port) get a new one */
1807 socket = rxi_GetHostUDPSocket(host, port);
1808 if (socket == OSI_NULLSOCKET) {
1810 rxi_FreeService(tservice);
1815 service->socket = socket;
1816 service->serviceHost = host;
1817 service->servicePort = port;
1818 service->serviceId = serviceId;
1819 service->serviceName = serviceName;
1820 service->nSecurityObjects = nSecurityObjects;
1821 service->securityObjects = securityObjects;
1822 service->minProcs = 0;
1823 service->maxProcs = 1;
1824 service->idleDeadTime = 60;
1825 service->idleDeadErr = 0;
1826 service->connDeadTime = rx_connDeadTime;
1827 service->executeRequestProc = serviceProc;
1828 service->checkReach = 0;
1829 service->nSpecific = 0;
1830 service->specific = NULL;
1831 rx_services[i] = service; /* not visible until now */
1837 rxi_FreeService(tservice);
1838 (osi_Msg "rx_NewService: cannot support > %d services\n",
1843 /* Set configuration options for all of a service's security objects */
1846 rx_SetSecurityConfiguration(struct rx_service *service,
1847 rx_securityConfigVariables type,
1851 for (i = 0; i<service->nSecurityObjects; i++) {
1852 if (service->securityObjects[i]) {
1853 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1861 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1862 struct rx_securityClass **securityObjects, int nSecurityObjects,
1863 afs_int32(*serviceProc) (struct rx_call * acall))
1865 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1868 /* Generic request processing loop. This routine should be called
1869 * by the implementation dependent rx_ServerProc. If socketp is
1870 * non-null, it will be set to the file descriptor that this thread
1871 * is now listening on. If socketp is null, this routine will never
1874 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1876 struct rx_call *call;
1878 struct rx_service *tservice = NULL;
1885 call = rx_GetCall(threadID, tservice, socketp);
1886 if (socketp && *socketp != OSI_NULLSOCKET) {
1887 /* We are now a listener thread */
1893 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1894 #ifdef RX_ENABLE_LOCKS
1896 #endif /* RX_ENABLE_LOCKS */
1897 afs_termState = AFSOP_STOP_AFS;
1898 afs_osi_Wakeup(&afs_termState);
1899 #ifdef RX_ENABLE_LOCKS
1901 #endif /* RX_ENABLE_LOCKS */
1906 /* if server is restarting( typically smooth shutdown) then do not
1907 * allow any new calls.
1910 if (rx_tranquil && (call != NULL)) {
1914 MUTEX_ENTER(&call->lock);
1916 rxi_CallError(call, RX_RESTARTING);
1917 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1919 MUTEX_EXIT(&call->lock);
1924 tservice = call->conn->service;
1926 if (tservice->beforeProc)
1927 (*tservice->beforeProc) (call);
1929 code = tservice->executeRequestProc(call);
1931 if (tservice->afterProc)
1932 (*tservice->afterProc) (call, code);
1934 rx_EndCall(call, code);
1936 if (tservice->postProc)
1937 (*tservice->postProc) (code);
1939 if (rx_stats_active) {
1940 MUTEX_ENTER(&rx_stats_mutex);
1942 MUTEX_EXIT(&rx_stats_mutex);
1949 rx_WakeupServerProcs(void)
1951 struct rx_serverQueueEntry *np, *tqp;
1952 struct opr_queue *cursor;
1956 MUTEX_ENTER(&rx_serverPool_lock);
1958 #ifdef RX_ENABLE_LOCKS
1959 if (rx_waitForPacket)
1960 CV_BROADCAST(&rx_waitForPacket->cv);
1961 #else /* RX_ENABLE_LOCKS */
1962 if (rx_waitForPacket)
1963 osi_rxWakeup(rx_waitForPacket);
1964 #endif /* RX_ENABLE_LOCKS */
1965 MUTEX_ENTER(&freeSQEList_lock);
1966 for (np = rx_FreeSQEList; np; np = tqp) {
1967 tqp = *(struct rx_serverQueueEntry **)np;
1968 #ifdef RX_ENABLE_LOCKS
1969 CV_BROADCAST(&np->cv);
1970 #else /* RX_ENABLE_LOCKS */
1972 #endif /* RX_ENABLE_LOCKS */
1974 MUTEX_EXIT(&freeSQEList_lock);
1975 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1976 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1977 #ifdef RX_ENABLE_LOCKS
1978 CV_BROADCAST(&np->cv);
1979 #else /* RX_ENABLE_LOCKS */
1981 #endif /* RX_ENABLE_LOCKS */
1983 MUTEX_EXIT(&rx_serverPool_lock);
1988 * One thing that seems to happen is that all the server threads get
1989 * tied up on some empty or slow call, and then a whole bunch of calls
1990 * arrive at once, using up the packet pool, so now there are more
1991 * empty calls. The most critical resources here are server threads
1992 * and the free packet pool. The "doreclaim" code seems to help in
1993 * general. I think that eventually we arrive in this state: there
1994 * are lots of pending calls which do have all their packets present,
1995 * so they won't be reclaimed, are multi-packet calls, so they won't
1996 * be scheduled until later, and thus are tying up most of the free
1997 * packet pool for a very long time.
1999 * 1. schedule multi-packet calls if all the packets are present.
2000 * Probably CPU-bound operation, useful to return packets to pool.
2001 * Do what if there is a full window, but the last packet isn't here?
2002 * 3. preserve one thread which *only* runs "best" calls, otherwise
2003 * it sleeps and waits for that type of call.
2004 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2005 * the current dataquota business is badly broken. The quota isn't adjusted
2006 * to reflect how many packets are presently queued for a running call.
2007 * So, when we schedule a queued call with a full window of packets queued
2008 * up for it, that *should* free up a window full of packets for other 2d-class
2009 * calls to be able to use from the packet pool. But it doesn't.
2011 * NB. Most of the time, this code doesn't run -- since idle server threads
2012 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2013 * as a new call arrives.
2015 /* Sleep until a call arrives. Returns a pointer to the call, ready
2016 * for an rx_Read. */
2017 #ifdef RX_ENABLE_LOCKS
2019 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2021 struct rx_serverQueueEntry *sq;
2022 struct rx_call *call = (struct rx_call *)0;
2023 struct rx_service *service = NULL;
2025 MUTEX_ENTER(&freeSQEList_lock);
2027 if ((sq = rx_FreeSQEList)) {
2028 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2029 MUTEX_EXIT(&freeSQEList_lock);
2030 } else { /* otherwise allocate a new one and return that */
2031 MUTEX_EXIT(&freeSQEList_lock);
2032 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2033 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2034 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2037 MUTEX_ENTER(&rx_serverPool_lock);
2038 if (cur_service != NULL) {
2039 ReturnToServerPool(cur_service);
2042 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2043 struct rx_call *tcall, *choice2 = NULL;
2044 struct opr_queue *cursor;
2046 /* Scan for eligible incoming calls. A call is not eligible
2047 * if the maximum number of calls for its service type are
2048 * already executing */
2049 /* One thread will process calls FCFS (to prevent starvation),
2050 * while the other threads may run ahead looking for calls which
2051 * have all their input data available immediately. This helps
2052 * keep threads from blocking, waiting for data from the client. */
2053 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2054 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2056 service = tcall->conn->service;
2057 if (!QuotaOK(service)) {
2060 MUTEX_ENTER(&rx_pthread_mutex);
2061 if (tno == rxi_fcfs_thread_num
2062 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2063 MUTEX_EXIT(&rx_pthread_mutex);
2064 /* If we're the fcfs thread , then we'll just use
2065 * this call. If we haven't been able to find an optimal
2066 * choice, and we're at the end of the list, then use a
2067 * 2d choice if one has been identified. Otherwise... */
2068 call = (choice2 ? choice2 : tcall);
2069 service = call->conn->service;
2071 MUTEX_EXIT(&rx_pthread_mutex);
2072 if (!opr_queue_IsEmpty(&tcall->rq)) {
2073 struct rx_packet *rp;
2074 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2076 if (rp->header.seq == 1) {
2078 || (rp->header.flags & RX_LAST_PACKET)) {
2080 } else if (rxi_2dchoice && !choice2
2081 && !(tcall->flags & RX_CALL_CLEARED)
2082 && (tcall->rprev > rxi_HardAckRate)) {
2092 ReturnToServerPool(service);
2098 opr_queue_Remove(&call->entry);
2099 MUTEX_EXIT(&rx_serverPool_lock);
2100 MUTEX_ENTER(&call->lock);
2102 if (call->flags & RX_CALL_WAIT_PROC) {
2103 call->flags &= ~RX_CALL_WAIT_PROC;
2104 rx_atomic_dec(&rx_nWaiting);
2107 if (call->state != RX_STATE_PRECALL || call->error) {
2108 MUTEX_EXIT(&call->lock);
2109 MUTEX_ENTER(&rx_serverPool_lock);
2110 ReturnToServerPool(service);
2115 if (opr_queue_IsEmpty(&call->rq)
2116 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2117 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2119 CLEAR_CALL_QUEUE_LOCK(call);
2122 /* If there are no eligible incoming calls, add this process
2123 * to the idle server queue, to wait for one */
2127 *socketp = OSI_NULLSOCKET;
2129 sq->socketp = socketp;
2130 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2131 #ifndef AFS_AIX41_ENV
2132 rx_waitForPacket = sq;
2134 rx_waitingForPacket = sq;
2135 #endif /* AFS_AIX41_ENV */
2137 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2139 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2140 MUTEX_EXIT(&rx_serverPool_lock);
2141 return (struct rx_call *)0;
2144 } while (!(call = sq->newcall)
2145 && !(socketp && *socketp != OSI_NULLSOCKET));
2146 MUTEX_EXIT(&rx_serverPool_lock);
2148 MUTEX_ENTER(&call->lock);
2154 MUTEX_ENTER(&freeSQEList_lock);
2155 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2156 rx_FreeSQEList = sq;
2157 MUTEX_EXIT(&freeSQEList_lock);
2160 clock_GetTime(&call->startTime);
2161 call->state = RX_STATE_ACTIVE;
2162 call->app.mode = RX_MODE_RECEIVING;
2163 #ifdef RX_KERNEL_TRACE
2164 if (ICL_SETACTIVE(afs_iclSetp)) {
2165 int glockOwner = ISAFS_GLOCK();
2168 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2169 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2176 rxi_calltrace(RX_CALL_START, call);
2177 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2178 call->conn->service->servicePort, call->conn->service->serviceId,
2181 MUTEX_EXIT(&call->lock);
2182 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2184 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2189 #else /* RX_ENABLE_LOCKS */
2191 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2193 struct rx_serverQueueEntry *sq;
2194 struct rx_call *call = (struct rx_call *)0, *choice2;
2195 struct rx_service *service = NULL;
2199 MUTEX_ENTER(&freeSQEList_lock);
2201 if ((sq = rx_FreeSQEList)) {
2202 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2203 MUTEX_EXIT(&freeSQEList_lock);
2204 } else { /* otherwise allocate a new one and return that */
2205 MUTEX_EXIT(&freeSQEList_lock);
2206 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2207 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2208 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2210 MUTEX_ENTER(&sq->lock);
2212 if (cur_service != NULL) {
2213 cur_service->nRequestsRunning--;
2214 MUTEX_ENTER(&rx_quota_mutex);
2215 if (cur_service->nRequestsRunning < cur_service->minProcs)
2218 MUTEX_EXIT(&rx_quota_mutex);
2220 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2221 struct rx_call *tcall;
2222 struct opr_queue *cursor;
2223 /* Scan for eligible incoming calls. A call is not eligible
2224 * if the maximum number of calls for its service type are
2225 * already executing */
2226 /* One thread will process calls FCFS (to prevent starvation),
2227 * while the other threads may run ahead looking for calls which
2228 * have all their input data available immediately. This helps
2229 * keep threads from blocking, waiting for data from the client. */
2230 choice2 = (struct rx_call *)0;
2231 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2232 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2233 service = tcall->conn->service;
2234 if (QuotaOK(service)) {
2235 MUTEX_ENTER(&rx_pthread_mutex);
2236 /* XXX - If tcall->entry.next is NULL, then we're no longer
2237 * on a queue at all. This shouldn't happen. */
2238 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2239 MUTEX_EXIT(&rx_pthread_mutex);
2240 /* If we're the fcfs thread, then we'll just use
2241 * this call. If we haven't been able to find an optimal
2242 * choice, and we're at the end of the list, then use a
2243 * 2d choice if one has been identified. Otherwise... */
2244 call = (choice2 ? choice2 : tcall);
2245 service = call->conn->service;
2247 MUTEX_EXIT(&rx_pthread_mutex);
2248 if (!opr_queue_IsEmpty(&tcall->rq)) {
2249 struct rx_packet *rp;
2250 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2252 if (rp->header.seq == 1
2254 || (rp->header.flags & RX_LAST_PACKET))) {
2256 } else if (rxi_2dchoice && !choice2
2257 && !(tcall->flags & RX_CALL_CLEARED)
2258 && (tcall->rprev > rxi_HardAckRate)) {
2271 opr_queue_Remove(&call->entry);
2272 /* we can't schedule a call if there's no data!!! */
2273 /* send an ack if there's no data, if we're missing the
2274 * first packet, or we're missing something between first
2275 * and last -- there's a "hole" in the incoming data. */
2276 if (opr_queue_IsEmpty(&call->rq)
2277 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2278 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2279 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2281 call->flags &= (~RX_CALL_WAIT_PROC);
2282 service->nRequestsRunning++;
2283 /* just started call in minProcs pool, need fewer to maintain
2285 MUTEX_ENTER(&rx_quota_mutex);
2286 if (service->nRequestsRunning <= service->minProcs)
2289 MUTEX_EXIT(&rx_quota_mutex);
2290 rx_atomic_dec(&rx_nWaiting);
2291 /* MUTEX_EXIT(&call->lock); */
2293 /* If there are no eligible incoming calls, add this process
2294 * to the idle server queue, to wait for one */
2297 *socketp = OSI_NULLSOCKET;
2299 sq->socketp = socketp;
2300 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2304 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2306 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2307 return (struct rx_call *)0;
2310 } while (!(call = sq->newcall)
2311 && !(socketp && *socketp != OSI_NULLSOCKET));
2313 MUTEX_EXIT(&sq->lock);
2315 MUTEX_ENTER(&freeSQEList_lock);
2316 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2317 rx_FreeSQEList = sq;
2318 MUTEX_EXIT(&freeSQEList_lock);
2321 clock_GetTime(&call->startTime);
2322 call->state = RX_STATE_ACTIVE;
2323 call->app.mode = RX_MODE_RECEIVING;
2324 #ifdef RX_KERNEL_TRACE
2325 if (ICL_SETACTIVE(afs_iclSetp)) {
2326 int glockOwner = ISAFS_GLOCK();
2329 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2330 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2337 rxi_calltrace(RX_CALL_START, call);
2338 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2339 call->conn->service->servicePort, call->conn->service->serviceId,
2342 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2349 #endif /* RX_ENABLE_LOCKS */
2353 /* Establish a procedure to be called when a packet arrives for a
2354 * call. This routine will be called at most once after each call,
2355 * and will also be called if there is an error condition on the or
2356 * the call is complete. Used by multi rx to build a selection
2357 * function which determines which of several calls is likely to be a
2358 * good one to read from.
2359 * NOTE: the way this is currently implemented it is probably only a
2360 * good idea to (1) use it immediately after a newcall (clients only)
2361 * and (2) only use it once. Other uses currently void your warranty
2364 rx_SetArrivalProc(struct rx_call *call,
2365 void (*proc) (struct rx_call * call,
2368 void * handle, int arg)
2370 call->arrivalProc = proc;
2371 call->arrivalProcHandle = handle;
2372 call->arrivalProcArg = arg;
2375 /* Call is finished (possibly prematurely). Return rc to the peer, if
2376 * appropriate, and return the final error code from the conversation
2380 rx_EndCall(struct rx_call *call, afs_int32 rc)
2382 struct rx_connection *conn = call->conn;
2386 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2387 call, rc, call->error, call->abortCode));
2390 MUTEX_ENTER(&call->lock);
2392 if (rc == 0 && call->error == 0) {
2393 call->abortCode = 0;
2394 call->abortCount = 0;
2397 call->arrivalProc = (void (*)())0;
2398 if (rc && call->error == 0) {
2399 rxi_CallError(call, rc);
2400 call->app.mode = RX_MODE_ERROR;
2401 /* Send an abort message to the peer if this error code has
2402 * only just been set. If it was set previously, assume the
2403 * peer has already been sent the error code or will request it
2405 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2407 if (conn->type == RX_SERVER_CONNECTION) {
2408 /* Make sure reply or at least dummy reply is sent */
2409 if (call->app.mode == RX_MODE_RECEIVING) {
2410 MUTEX_EXIT(&call->lock);
2411 rxi_WriteProc(call, 0, 0);
2412 MUTEX_ENTER(&call->lock);
2414 if (call->app.mode == RX_MODE_SENDING) {
2415 MUTEX_EXIT(&call->lock);
2416 rxi_FlushWrite(call);
2417 MUTEX_ENTER(&call->lock);
2419 rxi_calltrace(RX_CALL_END, call);
2420 /* Call goes to hold state until reply packets are acknowledged */
2421 if (call->tfirst + call->nSoftAcked < call->tnext) {
2422 call->state = RX_STATE_HOLD;
2424 call->state = RX_STATE_DALLY;
2425 rxi_ClearTransmitQueue(call, 0);
2426 rxi_rto_cancel(call);
2427 rxi_CancelKeepAliveEvent(call);
2429 } else { /* Client connection */
2431 /* Make sure server receives input packets, in the case where
2432 * no reply arguments are expected */
2434 if ((call->app.mode == RX_MODE_SENDING)
2435 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2436 MUTEX_EXIT(&call->lock);
2437 (void)rxi_ReadProc(call, &dummy, 1);
2438 MUTEX_ENTER(&call->lock);
2441 /* If we had an outstanding delayed ack, be nice to the server
2442 * and force-send it now.
2444 if (call->delayedAckEvent) {
2445 rxi_CancelDelayedAckEvent(call);
2446 rxi_SendDelayedAck(NULL, call, NULL, 0);
2449 /* We need to release the call lock since it's lower than the
2450 * conn_call_lock and we don't want to hold the conn_call_lock
2451 * over the rx_ReadProc call. The conn_call_lock needs to be held
2452 * here for the case where rx_NewCall is perusing the calls on
2453 * the connection structure. We don't want to signal until
2454 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2455 * have checked this call, found it active and by the time it
2456 * goes to sleep, will have missed the signal.
2458 MUTEX_EXIT(&call->lock);
2459 MUTEX_ENTER(&conn->conn_call_lock);
2460 MUTEX_ENTER(&call->lock);
2462 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2463 conn->lastBusy[call->channel] = 0;
2466 MUTEX_ENTER(&conn->conn_data_lock);
2467 conn->flags |= RX_CONN_BUSY;
2468 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2469 MUTEX_EXIT(&conn->conn_data_lock);
2470 #ifdef RX_ENABLE_LOCKS
2471 CV_BROADCAST(&conn->conn_call_cv);
2476 #ifdef RX_ENABLE_LOCKS
2478 MUTEX_EXIT(&conn->conn_data_lock);
2480 #endif /* RX_ENABLE_LOCKS */
2481 call->state = RX_STATE_DALLY;
2483 error = call->error;
2485 /* currentPacket, nLeft, and NFree must be zeroed here, because
2486 * ResetCall cannot: ResetCall may be called at splnet(), in the
2487 * kernel version, and may interrupt the macros rx_Read or
2488 * rx_Write, which run at normal priority for efficiency. */
2489 if (call->app.currentPacket) {
2490 #ifdef RX_TRACK_PACKETS
2491 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2493 rxi_FreePacket(call->app.currentPacket);
2494 call->app.currentPacket = (struct rx_packet *)0;
2497 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2499 /* Free any packets from the last call to ReadvProc/WritevProc */
2500 #ifdef RXDEBUG_PACKET
2502 #endif /* RXDEBUG_PACKET */
2503 rxi_FreePackets(0, &call->app.iovq);
2504 MUTEX_EXIT(&call->lock);
2506 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2507 if (conn->type == RX_CLIENT_CONNECTION) {
2508 MUTEX_ENTER(&conn->conn_data_lock);
2509 conn->flags &= ~RX_CONN_BUSY;
2510 MUTEX_EXIT(&conn->conn_data_lock);
2511 MUTEX_EXIT(&conn->conn_call_lock);
2515 * Map errors to the local host's errno.h format.
2517 error = ntoh_syserr_conv(error);
2521 #if !defined(KERNEL)
2523 /* Call this routine when shutting down a server or client (especially
2524 * clients). This will allow Rx to gracefully garbage collect server
2525 * connections, and reduce the number of retries that a server might
2526 * make to a dead client.
2527 * This is not quite right, since some calls may still be ongoing and
2528 * we can't lock them to destroy them. */
2532 struct rx_connection **conn_ptr, **conn_end;
2535 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2536 return; /* Already shutdown. */
2538 rxi_DeleteCachedConnections();
2539 if (rx_connHashTable) {
2540 MUTEX_ENTER(&rx_connHashTable_lock);
2541 for (conn_ptr = &rx_connHashTable[0], conn_end =
2542 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2544 struct rx_connection *conn, *next;
2545 for (conn = *conn_ptr; conn; conn = next) {
2547 if (conn->type == RX_CLIENT_CONNECTION) {
2548 MUTEX_ENTER(&rx_refcnt_mutex);
2550 MUTEX_EXIT(&rx_refcnt_mutex);
2551 #ifdef RX_ENABLE_LOCKS
2552 rxi_DestroyConnectionNoLock(conn);
2553 #else /* RX_ENABLE_LOCKS */
2554 rxi_DestroyConnection(conn);
2555 #endif /* RX_ENABLE_LOCKS */
2559 #ifdef RX_ENABLE_LOCKS
2560 while (rx_connCleanup_list) {
2561 struct rx_connection *conn;
2562 conn = rx_connCleanup_list;
2563 rx_connCleanup_list = rx_connCleanup_list->next;
2564 MUTEX_EXIT(&rx_connHashTable_lock);
2565 rxi_CleanupConnection(conn);
2566 MUTEX_ENTER(&rx_connHashTable_lock);
2568 MUTEX_EXIT(&rx_connHashTable_lock);
2569 #endif /* RX_ENABLE_LOCKS */
2574 afs_winsockCleanup();
2580 /* if we wakeup packet waiter too often, can get in loop with two
2581 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2583 rxi_PacketsUnWait(void)
2585 if (!rx_waitingForPackets) {
2589 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2590 return; /* still over quota */
2593 rx_waitingForPackets = 0;
2594 #ifdef RX_ENABLE_LOCKS
2595 CV_BROADCAST(&rx_waitingForPackets_cv);
2597 osi_rxWakeup(&rx_waitingForPackets);
2603 /* ------------------Internal interfaces------------------------- */
2605 /* Return this process's service structure for the
2606 * specified socket and service */
2607 static struct rx_service *
2608 rxi_FindService(osi_socket socket, u_short serviceId)
2610 struct rx_service **sp;
2611 for (sp = &rx_services[0]; *sp; sp++) {
2612 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2618 #ifdef RXDEBUG_PACKET
2619 #ifdef KDUMP_RX_LOCK
2620 static struct rx_call_rx_lock *rx_allCallsp = 0;
2622 static struct rx_call *rx_allCallsp = 0;
2624 #endif /* RXDEBUG_PACKET */
2626 /* Allocate a call structure, for the indicated channel of the
2627 * supplied connection. The mode and state of the call must be set by
2628 * the caller. Returns the call with mutex locked. */
2629 static struct rx_call *
2630 rxi_NewCall(struct rx_connection *conn, int channel)
2632 struct rx_call *call;
2633 #ifdef RX_ENABLE_LOCKS
2634 struct rx_call *cp; /* Call pointer temp */
2635 struct opr_queue *cursor;
2638 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2640 /* Grab an existing call structure, or allocate a new one.
2641 * Existing call structures are assumed to have been left reset by
2643 MUTEX_ENTER(&rx_freeCallQueue_lock);
2645 #ifdef RX_ENABLE_LOCKS
2647 * EXCEPT that the TQ might not yet be cleared out.
2648 * Skip over those with in-use TQs.
2651 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2652 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2653 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2659 #else /* RX_ENABLE_LOCKS */
2660 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2661 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2662 #endif /* RX_ENABLE_LOCKS */
2663 opr_queue_Remove(&call->entry);
2664 if (rx_stats_active)
2665 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2666 MUTEX_EXIT(&rx_freeCallQueue_lock);
2667 MUTEX_ENTER(&call->lock);
2668 CLEAR_CALL_QUEUE_LOCK(call);
2669 #ifdef RX_ENABLE_LOCKS
2670 /* Now, if TQ wasn't cleared earlier, do it now. */
2671 rxi_WaitforTQBusy(call);
2672 if (call->flags & RX_CALL_TQ_CLEARME) {
2673 rxi_ClearTransmitQueue(call, 1);
2674 /*queue_Init(&call->tq);*/
2676 #endif /* RX_ENABLE_LOCKS */
2677 /* Bind the call to its connection structure */
2679 rxi_ResetCall(call, 1);
2682 call = rxi_Alloc(sizeof(struct rx_call));
2683 #ifdef RXDEBUG_PACKET
2684 call->allNextp = rx_allCallsp;
2685 rx_allCallsp = call;
2687 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2688 #else /* RXDEBUG_PACKET */
2689 rx_atomic_inc(&rx_stats.nCallStructs);
2690 #endif /* RXDEBUG_PACKET */
2692 MUTEX_EXIT(&rx_freeCallQueue_lock);
2693 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2694 MUTEX_ENTER(&call->lock);
2695 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2696 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2697 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2699 /* Initialize once-only items */
2700 opr_queue_Init(&call->tq);
2701 opr_queue_Init(&call->rq);
2702 opr_queue_Init(&call->app.iovq);
2703 #ifdef RXDEBUG_PACKET
2704 call->rqc = call->tqc = call->iovqc = 0;
2705 #endif /* RXDEBUG_PACKET */
2706 /* Bind the call to its connection structure (prereq for reset) */
2708 rxi_ResetCall(call, 1);
2710 call->channel = channel;
2711 call->callNumber = &conn->callNumber[channel];
2712 call->rwind = conn->rwind[channel];
2713 call->twind = conn->twind[channel];
2714 /* Note that the next expected call number is retained (in
2715 * conn->callNumber[i]), even if we reallocate the call structure
2717 conn->call[channel] = call;
2718 /* if the channel's never been used (== 0), we should start at 1, otherwise
2719 * the call number is valid from the last time this channel was used */
2720 if (*call->callNumber == 0)
2721 *call->callNumber = 1;
2726 /* A call has been inactive long enough that so we can throw away
2727 * state, including the call structure, which is placed on the call
2730 * call->lock amd rx_refcnt_mutex are held upon entry.
2731 * haveCTLock is set when called from rxi_ReapConnections.
2733 * return 1 if the call is freed, 0 if not.
2736 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2738 int channel = call->channel;
2739 struct rx_connection *conn = call->conn;
2740 u_char state = call->state;
2743 * We are setting the state to RX_STATE_RESET to
2744 * ensure that no one else will attempt to use this
2745 * call once we drop the refcnt lock. We must drop
2746 * the refcnt lock before calling rxi_ResetCall
2747 * because it cannot be held across acquiring the
2748 * freepktQ lock. NewCall does the same.
2750 call->state = RX_STATE_RESET;
2751 MUTEX_EXIT(&rx_refcnt_mutex);
2752 rxi_ResetCall(call, 0);
2754 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2756 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2757 (*call->callNumber)++;
2759 if (call->conn->call[channel] == call)
2760 call->conn->call[channel] = 0;
2761 MUTEX_EXIT(&conn->conn_call_lock);
2764 * We couldn't obtain the conn_call_lock so we can't
2765 * disconnect the call from the connection. Set the
2766 * call state to dally so that the call can be reused.
2768 MUTEX_ENTER(&rx_refcnt_mutex);
2769 call->state = RX_STATE_DALLY;
2773 MUTEX_ENTER(&rx_freeCallQueue_lock);
2774 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2775 #ifdef RX_ENABLE_LOCKS
2776 /* A call may be free even though its transmit queue is still in use.
2777 * Since we search the call list from head to tail, put busy calls at
2778 * the head of the list, and idle calls at the tail.
2780 if (call->flags & RX_CALL_TQ_BUSY)
2781 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2783 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2784 #else /* RX_ENABLE_LOCKS */
2785 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2786 #endif /* RX_ENABLE_LOCKS */
2787 if (rx_stats_active)
2788 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2789 MUTEX_EXIT(&rx_freeCallQueue_lock);
2791 /* Destroy the connection if it was previously slated for
2792 * destruction, i.e. the Rx client code previously called
2793 * rx_DestroyConnection (client connections), or
2794 * rxi_ReapConnections called the same routine (server
2795 * connections). Only do this, however, if there are no
2796 * outstanding calls. Note that for fine grain locking, there appears
2797 * to be a deadlock in that rxi_FreeCall has a call locked and
2798 * DestroyConnectionNoLock locks each call in the conn. But note a
2799 * few lines up where we have removed this call from the conn.
2800 * If someone else destroys a connection, they either have no
2801 * call lock held or are going through this section of code.
2803 MUTEX_ENTER(&conn->conn_data_lock);
2804 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2805 MUTEX_ENTER(&rx_refcnt_mutex);
2807 MUTEX_EXIT(&rx_refcnt_mutex);
2808 MUTEX_EXIT(&conn->conn_data_lock);
2809 #ifdef RX_ENABLE_LOCKS
2811 rxi_DestroyConnectionNoLock(conn);
2813 rxi_DestroyConnection(conn);
2814 #else /* RX_ENABLE_LOCKS */
2815 rxi_DestroyConnection(conn);
2816 #endif /* RX_ENABLE_LOCKS */
2818 MUTEX_EXIT(&conn->conn_data_lock);
2820 MUTEX_ENTER(&rx_refcnt_mutex);
2824 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2825 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2828 rxi_Alloc(size_t size)
2832 if (rx_stats_active) {
2833 rx_atomic_add(&rxi_Allocsize, (int) size);
2834 rx_atomic_inc(&rxi_Alloccnt);
2838 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2839 afs_osi_Alloc_NoSleep(size);
2844 osi_Panic("rxi_Alloc error");
2850 rxi_Free(void *addr, size_t size)
2852 if (rx_stats_active) {
2853 rx_atomic_sub(&rxi_Allocsize, (int) size);
2854 rx_atomic_dec(&rxi_Alloccnt);
2856 osi_Free(addr, size);
2860 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2862 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2863 struct rx_peer *next = NULL;
2867 MUTEX_ENTER(&rx_peerHashTable_lock);
2869 peer_ptr = &rx_peerHashTable[0];
2870 peer_end = &rx_peerHashTable[rx_hashTableSize];
2873 for ( ; peer_ptr < peer_end; peer_ptr++) {
2876 for ( ; peer; peer = next) {
2878 if (host == peer->host)
2883 hashIndex = PEER_HASH(host, port);
2884 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2885 if ((peer->host == host) && (peer->port == port))
2890 MUTEX_ENTER(&rx_peerHashTable_lock);
2895 MUTEX_EXIT(&rx_peerHashTable_lock);
2897 MUTEX_ENTER(&peer->peer_lock);
2898 /* We don't handle dropping below min, so don't */
2899 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2900 peer->ifMTU=MIN(mtu, peer->ifMTU);
2901 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2902 /* if we tweaked this down, need to tune our peer MTU too */
2903 peer->MTU = MIN(peer->MTU, peer->natMTU);
2904 /* if we discovered a sub-1500 mtu, degrade */
2905 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2906 peer->maxDgramPackets = 1;
2907 /* We no longer have valid peer packet information */
2908 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2909 peer->maxPacketSize = 0;
2910 MUTEX_EXIT(&peer->peer_lock);
2912 MUTEX_ENTER(&rx_peerHashTable_lock);
2914 if (host && !port) {
2916 /* pick up where we left off */
2920 MUTEX_EXIT(&rx_peerHashTable_lock);
2923 #ifdef AFS_RXERRQ_ENV
2925 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2927 int hashIndex = PEER_HASH(host, port);
2928 struct rx_peer *peer;
2930 MUTEX_ENTER(&rx_peerHashTable_lock);
2932 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2933 if (peer->host == host && peer->port == port) {
2939 MUTEX_EXIT(&rx_peerHashTable_lock);
2942 rx_atomic_inc(&peer->neterrs);
2943 MUTEX_ENTER(&peer->peer_lock);
2944 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2945 peer->last_err_type = err->ee_type;
2946 peer->last_err_code = err->ee_code;
2947 MUTEX_EXIT(&peer->peer_lock);
2949 MUTEX_ENTER(&rx_peerHashTable_lock);
2951 MUTEX_EXIT(&rx_peerHashTable_lock);
2956 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2958 # ifdef AFS_ADAPT_PMTU
2959 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2960 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2964 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2965 switch (err->ee_code) {
2966 case ICMP_NET_UNREACH:
2967 case ICMP_HOST_UNREACH:
2968 case ICMP_PORT_UNREACH:
2971 rxi_SetPeerDead(err, addr, port);
2978 rxi_TranslateICMP(int type, int code)
2981 case ICMP_DEST_UNREACH:
2983 case ICMP_NET_UNREACH:
2984 return "Destination Net Unreachable";
2985 case ICMP_HOST_UNREACH:
2986 return "Destination Host Unreachable";
2987 case ICMP_PROT_UNREACH:
2988 return "Destination Protocol Unreachable";
2989 case ICMP_PORT_UNREACH:
2990 return "Destination Port Unreachable";
2992 return "Destination Net Prohibited";
2994 return "Destination Host Prohibited";
3000 #endif /* AFS_RXERRQ_ENV */
3003 * Get the last network error for a connection
3005 * A "network error" here means an error retrieved from ICMP, or some other
3006 * mechanism outside of Rx that informs us of errors in network reachability.
3008 * If a peer associated with the given Rx connection has received a network
3009 * error recently, this function allows the caller to know what error
3010 * specifically occurred. This can be useful to know, since e.g. ICMP errors
3011 * can cause calls to that peer to be quickly aborted. So, this function can
3012 * help see why a call was aborted due to network errors.
3014 * If we have received traffic from a peer since the last network error, we
3015 * treat that peer as if we had not received an network error for it.
3017 * @param[in] conn The Rx connection to examine
3018 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
3019 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
3020 * @param[out] err_type The type of the last error
3021 * @param[out] err_code The code of the last error
3022 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
3024 * @return If we have an error
3025 * @retval -1 No error to get; 'out' params are undefined
3026 * @retval 0 We have an error; 'out' params contain the last error
3029 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
3030 int *err_code, const char **msg)
3032 #ifdef AFS_RXERRQ_ENV
3033 struct rx_peer *peer = conn->peer;
3034 if (rx_atomic_read(&peer->neterrs)) {
3035 MUTEX_ENTER(&peer->peer_lock);
3036 *err_origin = peer->last_err_origin;
3037 *err_type = peer->last_err_type;
3038 *err_code = peer->last_err_code;
3039 MUTEX_EXIT(&peer->peer_lock);
3042 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3043 *msg = rxi_TranslateICMP(*err_type, *err_code);
3052 /* Find the peer process represented by the supplied (host,port)
3053 * combination. If there is no appropriate active peer structure, a
3054 * new one will be allocated and initialized
3057 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3061 hashIndex = PEER_HASH(host, port);
3062 MUTEX_ENTER(&rx_peerHashTable_lock);
3063 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3064 if ((pp->host == host) && (pp->port == port))
3069 pp = rxi_AllocPeer(); /* This bzero's *pp */
3070 pp->host = host; /* set here or in InitPeerParams is zero */
3072 #ifdef AFS_RXERRQ_ENV
3073 rx_atomic_set(&pp->neterrs, 0);
3075 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3076 opr_queue_Init(&pp->rpcStats);
3077 pp->next = rx_peerHashTable[hashIndex];
3078 rx_peerHashTable[hashIndex] = pp;
3079 rxi_InitPeerParams(pp);
3080 if (rx_stats_active)
3081 rx_atomic_inc(&rx_stats.nPeerStructs);
3087 MUTEX_EXIT(&rx_peerHashTable_lock);
3092 /* Find the connection at (host, port) started at epoch, and with the
3093 * given connection id. Creates the server connection if necessary.
3094 * The type specifies whether a client connection or a server
3095 * connection is desired. In both cases, (host, port) specify the
3096 * peer's (host, pair) pair. Client connections are not made
3097 * automatically by this routine. The parameter socket gives the
3098 * socket descriptor on which the packet was received. This is used,
3099 * in the case of server connections, to check that *new* connections
3100 * come via a valid (port, serviceId). Finally, the securityIndex
3101 * parameter must match the existing index for the connection. If a
3102 * server connection is created, it will be created using the supplied
3103 * index, if the index is valid for this service */
3104 static struct rx_connection *
3105 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3106 u_short port, u_short serviceId, afs_uint32 cid,
3107 afs_uint32 epoch, int type, u_int securityIndex,
3108 int *unknownService)
3110 int hashindex, flag, i;
3111 struct rx_connection *conn;
3112 *unknownService = 0;
3113 hashindex = CONN_HASH(host, port, cid, epoch, type);
3114 MUTEX_ENTER(&rx_connHashTable_lock);
3115 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3116 rx_connHashTable[hashindex],
3119 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3120 && (epoch == conn->epoch)) {
3121 struct rx_peer *pp = conn->peer;
3122 if (securityIndex != conn->securityIndex) {
3123 /* this isn't supposed to happen, but someone could forge a packet
3124 * like this, and there seems to be some CM bug that makes this
3125 * happen from time to time -- in which case, the fileserver
3127 MUTEX_EXIT(&rx_connHashTable_lock);
3128 return (struct rx_connection *)0;
3130 if (pp->host == host && pp->port == port)
3132 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3134 /* So what happens when it's a callback connection? */
3135 if ( /*type == RX_CLIENT_CONNECTION && */
3136 (conn->epoch & 0x80000000))
3140 /* the connection rxLastConn that was used the last time is not the
3141 ** one we are looking for now. Hence, start searching in the hash */
3143 conn = rx_connHashTable[hashindex];
3148 struct rx_service *service;
3149 if (type == RX_CLIENT_CONNECTION) {
3150 MUTEX_EXIT(&rx_connHashTable_lock);
3151 return (struct rx_connection *)0;
3153 service = rxi_FindService(socket, serviceId);
3154 if (!service || (securityIndex >= service->nSecurityObjects)
3155 || (service->securityObjects[securityIndex] == 0)) {
3156 MUTEX_EXIT(&rx_connHashTable_lock);
3157 *unknownService = 1;
3158 return (struct rx_connection *)0;
3160 conn = rxi_AllocConnection(); /* This bzero's the connection */
3161 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3162 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3163 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3164 conn->next = rx_connHashTable[hashindex];
3165 rx_connHashTable[hashindex] = conn;
3166 conn->peer = rxi_FindPeer(host, port, 1);
3167 conn->type = RX_SERVER_CONNECTION;
3168 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3169 conn->epoch = epoch;
3170 conn->cid = cid & RX_CIDMASK;
3171 conn->ackRate = RX_FAST_ACK_RATE;
3172 conn->service = service;
3173 conn->serviceId = serviceId;
3174 conn->securityIndex = securityIndex;
3175 conn->securityObject = service->securityObjects[securityIndex];
3176 conn->nSpecific = 0;
3177 conn->specific = NULL;
3178 rx_SetConnDeadTime(conn, service->connDeadTime);
3179 conn->idleDeadTime = service->idleDeadTime;
3180 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3181 for (i = 0; i < RX_MAXCALLS; i++) {
3182 conn->twind[i] = rx_initSendWindow;
3183 conn->rwind[i] = rx_initReceiveWindow;
3185 /* Notify security object of the new connection */
3186 RXS_NewConnection(conn->securityObject, conn);
3187 /* XXXX Connection timeout? */
3188 if (service->newConnProc)
3189 (*service->newConnProc) (conn);
3190 if (rx_stats_active)
3191 rx_atomic_inc(&rx_stats.nServerConns);
3194 MUTEX_ENTER(&rx_refcnt_mutex);
3196 MUTEX_EXIT(&rx_refcnt_mutex);
3198 rxLastConn = conn; /* store this connection as the last conn used */
3199 MUTEX_EXIT(&rx_connHashTable_lock);
3204 * Timeout a call on a busy call channel if appropriate.
3206 * @param[in] call The busy call.
3208 * @pre 'call' is marked as busy (namely,
3209 * call->conn->lastBusy[call->channel] != 0)
3211 * @pre call->lock is held
3212 * @pre rxi_busyChannelError is nonzero
3214 * @note call->lock is dropped and reacquired
3217 rxi_CheckBusy(struct rx_call *call)
3219 struct rx_connection *conn = call->conn;
3220 int channel = call->channel;
3221 int freechannel = 0;
3224 MUTEX_EXIT(&call->lock);
3226 MUTEX_ENTER(&conn->conn_call_lock);
3228 /* Are there any other call slots on this conn that we should try? Look for
3229 * slots that are empty and are either non-busy, or were marked as busy
3230 * longer than conn->secondsUntilDead seconds before this call started. */
3232 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3234 /* only look at channels that aren't us */
3238 if (conn->lastBusy[i]) {
3239 /* if this channel looked busy too recently, don't look at it */
3240 if (conn->lastBusy[i] >= call->startTime.sec) {
3243 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3248 if (conn->call[i]) {
3249 struct rx_call *tcall = conn->call[i];
3250 MUTEX_ENTER(&tcall->lock);
3251 if (tcall->state == RX_STATE_DALLY) {
3254 MUTEX_EXIT(&tcall->lock);
3260 MUTEX_ENTER(&call->lock);
3262 /* Since the call->lock has been released it is possible that the call may
3263 * no longer be busy (the call channel cannot have been reallocated as we
3264 * haven't dropped the conn_call_lock) Therefore, we must confirm
3265 * that the call state has not changed when deciding whether or not to
3266 * force this application thread to retry by forcing a Timeout error. */
3268 if (freechannel && (call->flags & RX_CALL_PEER_BUSY)) {
3269 /* Since 'freechannel' is set, there exists another channel in this
3270 * rx_conn that the application thread might be able to use. We know
3271 * that we have the correct call since callNumber is unchanged, and we
3272 * know that the call is still busy. So, set the call error state to
3273 * rxi_busyChannelError so the application can retry the request,
3274 * presumably on a less-busy call channel. */
3276 rxi_CallError(call, RX_CALL_BUSY);
3278 MUTEX_EXIT(&conn->conn_call_lock);
3282 * Abort the call if the server is over the busy threshold. This
3283 * can be used without requiring a call structure be initialised,
3284 * or connected to a particular channel
3287 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3288 struct rx_packet *np)
3290 if ((rx_BusyThreshold > 0) &&
3291 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3292 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3293 rx_BusyError, np, 0);
3294 if (rx_stats_active)
3295 rx_atomic_inc(&rx_stats.nBusies);
3302 static_inline struct rx_call *
3303 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3306 struct rx_call *call;
3308 channel = np->header.cid & RX_CHANNELMASK;
3309 MUTEX_ENTER(&conn->conn_call_lock);
3310 call = conn->call[channel];
3311 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3312 MUTEX_EXIT(&conn->conn_call_lock);
3313 if (rx_stats_active)
3314 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3318 MUTEX_ENTER(&call->lock);
3319 MUTEX_EXIT(&conn->conn_call_lock);
3321 if ((call->state == RX_STATE_DALLY)
3322 && np->header.type == RX_PACKET_TYPE_ACK) {
3323 if (rx_stats_active)
3324 rx_atomic_inc(&rx_stats.ignorePacketDally);
3325 MUTEX_EXIT(&call->lock);
3332 static_inline struct rx_call *
3333 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3334 struct rx_connection *conn)
3337 struct rx_call *call;
3339 channel = np->header.cid & RX_CHANNELMASK;
3340 MUTEX_ENTER(&conn->conn_call_lock);
3341 call = conn->call[channel];
3344 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3345 MUTEX_EXIT(&conn->conn_call_lock);
3349 call = rxi_NewCall(conn, channel); /* returns locked call */
3350 *call->callNumber = np->header.callNumber;
3351 MUTEX_EXIT(&conn->conn_call_lock);
3353 call->state = RX_STATE_PRECALL;
3354 clock_GetTime(&call->queueTime);
3355 call->app.bytesSent = 0;
3356 call->app.bytesRcvd = 0;
3357 rxi_KeepAliveOn(call);
3362 if (np->header.callNumber == conn->callNumber[channel]) {
3363 MUTEX_ENTER(&call->lock);
3364 MUTEX_EXIT(&conn->conn_call_lock);
3368 if (np->header.callNumber < conn->callNumber[channel]) {
3369 MUTEX_EXIT(&conn->conn_call_lock);
3370 if (rx_stats_active)
3371 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3375 MUTEX_ENTER(&call->lock);
3376 MUTEX_EXIT(&conn->conn_call_lock);
3378 /* Wait until the transmit queue is idle before deciding
3379 * whether to reset the current call. Chances are that the
3380 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3383 #ifdef RX_ENABLE_LOCKS
3384 if (call->state == RX_STATE_ACTIVE) {
3385 int old_error = call->error;
3386 rxi_WaitforTQBusy(call);
3387 /* If we entered error state while waiting,
3388 * must call rxi_CallError to permit rxi_ResetCall
3389 * to processed when the tqWaiter count hits zero.
3391 if (call->error && call->error != old_error) {
3392 rxi_CallError(call, call->error);
3393 MUTEX_EXIT(&call->lock);
3397 #endif /* RX_ENABLE_LOCKS */
3398 /* If the new call cannot be taken right now send a busy and set
3399 * the error condition in this call, so that it terminates as
3400 * quickly as possible */
3401 if (call->state == RX_STATE_ACTIVE) {
3402 rxi_CallError(call, RX_CALL_DEAD);
3403 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3405 MUTEX_EXIT(&call->lock);
3409 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3410 MUTEX_EXIT(&call->lock);
3414 rxi_ResetCall(call, 0);
3415 /* The conn_call_lock is not held but no one else should be
3416 * using this call channel while we are processing this incoming
3417 * packet. This assignment should be safe.
3419 *call->callNumber = np->header.callNumber;
3420 call->state = RX_STATE_PRECALL;
3421 clock_GetTime(&call->queueTime);
3422 call->app.bytesSent = 0;
3423 call->app.bytesRcvd = 0;
3424 rxi_KeepAliveOn(call);
3430 /* There are two packet tracing routines available for testing and monitoring
3431 * Rx. One is called just after every packet is received and the other is
3432 * called just before every packet is sent. Received packets, have had their
3433 * headers decoded, and packets to be sent have not yet had their headers
3434 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3435 * containing the network address. Both can be modified. The return value, if
3436 * non-zero, indicates that the packet should be dropped. */
3438 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3439 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3441 /* A packet has been received off the interface. Np is the packet, socket is
3442 * the socket number it was received from (useful in determining which service
3443 * this packet corresponds to), and (host, port) reflect the host,port of the
3444 * sender. This call returns the packet to the caller if it is finished with
3445 * it, rather than de-allocating it, just as a small performance hack */
3448 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3449 afs_uint32 host, u_short port, int *tnop,
3450 struct rx_call **newcallp)
3452 struct rx_call *call;
3453 struct rx_connection *conn;
3455 int unknownService = 0;
3459 struct rx_packet *tnp;
3462 /* We don't print out the packet until now because (1) the time may not be
3463 * accurate enough until now in the lwp implementation (rx_Listener only gets
3464 * the time after the packet is read) and (2) from a protocol point of view,
3465 * this is the first time the packet has been seen */
3466 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3467 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3468 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3469 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3470 np->header.epoch, np->header.cid, np->header.callNumber,
3471 np->header.seq, np->header.flags, np));
3474 /* Account for connectionless packets */
3475 if (rx_stats_active &&
3476 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3477 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3478 struct rx_peer *peer;
3480 /* Try to look up the peer structure, but don't create one */
3481 peer = rxi_FindPeer(host, port, 0);
3483 /* Since this may not be associated with a connection, it may have
3484 * no refCount, meaning we could race with ReapConnections
3487 if (peer && (peer->refCount > 0)) {
3488 #ifdef AFS_RXERRQ_ENV
3489 if (rx_atomic_read(&peer->neterrs)) {
3490 rx_atomic_set(&peer->neterrs, 0);
3493 MUTEX_ENTER(&peer->peer_lock);
3494 peer->bytesReceived += np->length;
3495 MUTEX_EXIT(&peer->peer_lock);
3499 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3500 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3503 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3504 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3507 /* If an input tracer function is defined, call it with the packet and
3508 * network address. Note this function may modify its arguments. */
3509 if (rx_justReceived) {
3510 struct sockaddr_in addr;
3512 addr.sin_family = AF_INET;
3513 addr.sin_port = port;
3514 addr.sin_addr.s_addr = host;
3515 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3516 addr.sin_len = sizeof(addr);
3517 #endif /* AFS_OSF_ENV */
3518 drop = (*rx_justReceived) (np, &addr);
3519 /* drop packet if return value is non-zero */
3522 port = addr.sin_port; /* in case fcn changed addr */
3523 host = addr.sin_addr.s_addr;
3527 /* If packet was not sent by the client, then *we* must be the client */
3528 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3529 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3531 /* Find the connection (or fabricate one, if we're the server & if
3532 * necessary) associated with this packet */
3534 rxi_FindConnection(socket, host, port, np->header.serviceId,
3535 np->header.cid, np->header.epoch, type,
3536 np->header.securityIndex, &unknownService);
3538 /* To avoid having 2 connections just abort at each other,
3539 don't abort an abort. */
3541 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3542 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3547 #ifdef AFS_RXERRQ_ENV
3548 if (rx_atomic_read(&conn->peer->neterrs)) {
3549 rx_atomic_set(&conn->peer->neterrs, 0);
3553 /* If we're doing statistics, then account for the incoming packet */
3554 if (rx_stats_active) {
3555 MUTEX_ENTER(&conn->peer->peer_lock);
3556 conn->peer->bytesReceived += np->length;
3557 MUTEX_EXIT(&conn->peer->peer_lock);
3560 /* If the connection is in an error state, send an abort packet and ignore
3561 * the incoming packet */
3563 /* Don't respond to an abort packet--we don't want loops! */
3564 MUTEX_ENTER(&conn->conn_data_lock);
3565 if (np->header.type != RX_PACKET_TYPE_ABORT)
3566 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3567 putConnection(conn);
3568 MUTEX_EXIT(&conn->conn_data_lock);
3572 /* Check for connection-only requests (i.e. not call specific). */
3573 if (np->header.callNumber == 0) {
3574 switch (np->header.type) {
3575 case RX_PACKET_TYPE_ABORT: {
3576 /* What if the supplied error is zero? */
3577 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3578 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3579 rxi_ConnectionError(conn, errcode);
3580 putConnection(conn);
3583 case RX_PACKET_TYPE_CHALLENGE:
3584 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3585 putConnection(conn);
3587 case RX_PACKET_TYPE_RESPONSE:
3588 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3589 putConnection(conn);
3591 case RX_PACKET_TYPE_PARAMS:
3592 case RX_PACKET_TYPE_PARAMS + 1:
3593 case RX_PACKET_TYPE_PARAMS + 2:
3594 /* ignore these packet types for now */
3595 putConnection(conn);
3599 /* Should not reach here, unless the peer is broken: send an
3601 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3602 MUTEX_ENTER(&conn->conn_data_lock);
3603 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3604 putConnection(conn);
3605 MUTEX_EXIT(&conn->conn_data_lock);
3610 if (type == RX_SERVER_CONNECTION)
3611 call = rxi_ReceiveServerCall(socket, np, conn);
3613 call = rxi_ReceiveClientCall(np, conn);
3616 putConnection(conn);
3620 MUTEX_ASSERT(&call->lock);
3621 /* Set remote user defined status from packet */
3622 call->remoteStatus = np->header.userStatus;
3624 /* Now do packet type-specific processing */
3625 switch (np->header.type) {
3626 case RX_PACKET_TYPE_DATA:
3627 /* If we're a client, and receiving a response, then all the packets
3628 * we transmitted packets are implicitly acknowledged. */
3629 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3630 rxi_AckAllInTransmitQueue(call);
3632 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3635 case RX_PACKET_TYPE_ACK:
3636 /* Respond immediately to ack packets requesting acknowledgement
3638 if (np->header.flags & RX_REQUEST_ACK) {
3640 (void)rxi_SendCallAbort(call, 0, 1, 0);
3642 (void)rxi_SendAck(call, 0, np->header.serial,
3643 RX_ACK_PING_RESPONSE, 1);
3645 np = rxi_ReceiveAckPacket(call, np, 1);
3647 case RX_PACKET_TYPE_ABORT: {
3648 /* An abort packet: reset the call, passing the error up to the user. */
3649 /* What if error is zero? */
3650 /* What if the error is -1? the application will treat it as a timeout. */
3651 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3652 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3653 rxi_CallError(call, errdata);
3654 MUTEX_EXIT(&call->lock);
3655 putConnection(conn);
3656 return np; /* xmitting; drop packet */
3658 case RX_PACKET_TYPE_BUSY: {
3659 struct clock busyTime;
3661 clock_GetTime(&busyTime);
3663 MUTEX_EXIT(&call->lock);
3665 MUTEX_ENTER(&conn->conn_call_lock);
3666 MUTEX_ENTER(&call->lock);
3667 conn->lastBusy[call->channel] = busyTime.sec;
3668 call->flags |= RX_CALL_PEER_BUSY;
3669 MUTEX_EXIT(&call->lock);
3670 MUTEX_EXIT(&conn->conn_call_lock);
3672 putConnection(conn);
3676 case RX_PACKET_TYPE_ACKALL:
3677 /* All packets acknowledged, so we can drop all packets previously
3678 * readied for sending */
3679 rxi_AckAllInTransmitQueue(call);
3682 /* Should not reach here, unless the peer is broken: send an abort
3684 rxi_CallError(call, RX_PROTOCOL_ERROR);
3685 np = rxi_SendCallAbort(call, np, 1, 0);
3688 /* Note when this last legitimate packet was received, for keep-alive
3689 * processing. Note, we delay getting the time until now in the hope that
3690 * the packet will be delivered to the user before any get time is required
3691 * (if not, then the time won't actually be re-evaluated here). */
3692 call->lastReceiveTime = clock_Sec();
3693 /* we've received a legit packet, so the channel is not busy */
3694 call->flags &= ~RX_CALL_PEER_BUSY;
3695 MUTEX_EXIT(&call->lock);
3696 putConnection(conn);
3700 /* return true if this is an "interesting" connection from the point of view
3701 of someone trying to debug the system */
3703 rxi_IsConnInteresting(struct rx_connection *aconn)
3706 struct rx_call *tcall;
3708 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3711 for (i = 0; i < RX_MAXCALLS; i++) {
3712 tcall = aconn->call[i];
3714 if ((tcall->state == RX_STATE_PRECALL)
3715 || (tcall->state == RX_STATE_ACTIVE))
3717 if ((tcall->app.mode == RX_MODE_SENDING)
3718 || (tcall->app.mode == RX_MODE_RECEIVING))
3726 /* if this is one of the last few packets AND it wouldn't be used by the
3727 receiving call to immediately satisfy a read request, then drop it on
3728 the floor, since accepting it might prevent a lock-holding thread from
3729 making progress in its reading. If a call has been cleared while in
3730 the precall state then ignore all subsequent packets until the call
3731 is assigned to a thread. */
3734 TooLow(struct rx_packet *ap, struct rx_call *acall)
3738 MUTEX_ENTER(&rx_quota_mutex);
3739 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3740 && (acall->state == RX_STATE_PRECALL))
3741 || ((rx_nFreePackets < rxi_dataQuota + 2)
3742 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3743 && (acall->flags & RX_CALL_READER_WAIT)))) {
3746 MUTEX_EXIT(&rx_quota_mutex);
3752 * Clear the attach wait flag on a connection and proceed.
3754 * Any processing waiting for a connection to be attached should be
3755 * unblocked. We clear the flag and do any other needed tasks.
3758 * the conn to unmark waiting for attach
3760 * @pre conn's conn_data_lock must be locked before calling this function
3764 rxi_ConnClearAttachWait(struct rx_connection *conn)
3766 /* Indicate that rxi_CheckReachEvent is no longer running by
3767 * clearing the flag. Must be atomic under conn_data_lock to
3768 * avoid a new call slipping by: rxi_CheckConnReach holds
3769 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3771 conn->flags &= ~RX_CONN_ATTACHWAIT;
3772 if (conn->flags & RX_CONN_NAT_PING) {
3773 conn->flags &= ~RX_CONN_NAT_PING;
3774 rxi_ScheduleNatKeepAliveEvent(conn);
3779 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3781 struct rx_connection *conn = arg1;
3782 struct rx_call *acall = arg2;
3783 struct rx_call *call = acall;
3784 struct clock when, now;
3787 MUTEX_ENTER(&conn->conn_data_lock);
3790 rxevent_Put(&conn->checkReachEvent);
3792 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3794 putConnection(conn);
3796 MUTEX_EXIT(&conn->conn_data_lock);
3800 MUTEX_ENTER(&conn->conn_call_lock);
3801 MUTEX_ENTER(&conn->conn_data_lock);
3802 for (i = 0; i < RX_MAXCALLS; i++) {
3803 struct rx_call *tc = conn->call[i];
3804 if (tc && tc->state == RX_STATE_PRECALL) {
3810 rxi_ConnClearAttachWait(conn);
3811 MUTEX_EXIT(&conn->conn_data_lock);
3812 MUTEX_EXIT(&conn->conn_call_lock);
3817 MUTEX_ENTER(&call->lock);
3818 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3820 MUTEX_EXIT(&call->lock);
3822 clock_GetTime(&now);
3824 when.sec += RX_CHECKREACH_TIMEOUT;
3825 MUTEX_ENTER(&conn->conn_data_lock);
3826 if (!conn->checkReachEvent) {
3827 MUTEX_ENTER(&rx_refcnt_mutex);
3829 MUTEX_EXIT(&rx_refcnt_mutex);
3830 conn->checkReachEvent = rxevent_Post(&when, &now,
3831 rxi_CheckReachEvent, conn,
3834 MUTEX_EXIT(&conn->conn_data_lock);
3840 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3842 struct rx_service *service = conn->service;
3843 struct rx_peer *peer = conn->peer;
3844 afs_uint32 now, lastReach;
3846 if (service->checkReach == 0)
3850 MUTEX_ENTER(&peer->peer_lock);
3851 lastReach = peer->lastReachTime;
3852 MUTEX_EXIT(&peer->peer_lock);
3853 if (now - lastReach < RX_CHECKREACH_TTL)
3856 MUTEX_ENTER(&conn->conn_data_lock);
3857 if (conn->flags & RX_CONN_ATTACHWAIT) {
3858 MUTEX_EXIT(&conn->conn_data_lock);
3861 conn->flags |= RX_CONN_ATTACHWAIT;
3862 MUTEX_EXIT(&conn->conn_data_lock);
3863 if (!conn->checkReachEvent)
3864 rxi_CheckReachEvent(NULL, conn, call, 0);
3869 /* try to attach call, if authentication is complete */
3871 TryAttach(struct rx_call *acall, osi_socket socket,
3872 int *tnop, struct rx_call **newcallp,
3875 struct rx_connection *conn = acall->conn;
3877 if (conn->type == RX_SERVER_CONNECTION
3878 && acall->state == RX_STATE_PRECALL) {
3879 /* Don't attach until we have any req'd. authentication. */
3880 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3881 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3882 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3883 /* Note: this does not necessarily succeed; there
3884 * may not any proc available
3887 rxi_ChallengeOn(acall->conn);
3892 /* A data packet has been received off the interface. This packet is
3893 * appropriate to the call (the call is in the right state, etc.). This
3894 * routine can return a packet to the caller, for re-use */
3896 static struct rx_packet *
3897 rxi_ReceiveDataPacket(struct rx_call *call,
3898 struct rx_packet *np, int istack,
3899 osi_socket socket, afs_uint32 host, u_short port,
3900 int *tnop, struct rx_call **newcallp)
3902 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3907 afs_uint32 serial=0, flags=0;
3909 struct rx_packet *tnp;
3910 if (rx_stats_active)
3911 rx_atomic_inc(&rx_stats.dataPacketsRead);
3914 /* If there are no packet buffers, drop this new packet, unless we can find
3915 * packet buffers from inactive calls */
3917 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3918 MUTEX_ENTER(&rx_freePktQ_lock);
3919 rxi_NeedMorePackets = TRUE;
3920 MUTEX_EXIT(&rx_freePktQ_lock);
3921 if (rx_stats_active)
3922 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3923 rxi_calltrace(RX_TRACE_DROP, call);
3924 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3925 /* We used to clear the receive queue here, in an attempt to free
3926 * packets. However this is unsafe if the queue has received a
3927 * soft ACK for the final packet */
3928 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3934 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3935 * packet is one of several packets transmitted as a single
3936 * datagram. Do not send any soft or hard acks until all packets
3937 * in a jumbogram have been processed. Send negative acks right away.
3939 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3940 /* tnp is non-null when there are more packets in the
3941 * current jumbo gram */
3948 seq = np->header.seq;
3949 serial = np->header.serial;
3950 flags = np->header.flags;
3952 /* If the call is in an error state, send an abort message */
3954 return rxi_SendCallAbort(call, np, istack, 0);
3956 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3957 * AFS 3.5 jumbogram. */
3958 if (flags & RX_JUMBO_PACKET) {
3959 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3964 if (np->header.spare != 0) {
3965 MUTEX_ENTER(&call->conn->conn_data_lock);
3966 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3967 MUTEX_EXIT(&call->conn->conn_data_lock);
3970 /* The usual case is that this is the expected next packet */
3971 if (seq == call->rnext) {
3973 /* Check to make sure it is not a duplicate of one already queued */
3974 if (!opr_queue_IsEmpty(&call->rq)
3975 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3976 if (rx_stats_active)
3977 rx_atomic_inc(&rx_stats.dupPacketsRead);
3978 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3979 rxi_CancelDelayedAckEvent(call);
3980 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3986 /* It's the next packet. Stick it on the receive queue
3987 * for this call. Set newPackets to make sure we wake
3988 * the reader once all packets have been processed */
3989 #ifdef RX_TRACK_PACKETS
3990 np->flags |= RX_PKTFLAG_RQ;
3992 opr_queue_Prepend(&call->rq, &np->entry);
3993 #ifdef RXDEBUG_PACKET
3995 #endif /* RXDEBUG_PACKET */
3997 np = NULL; /* We can't use this anymore */
4000 /* If an ack is requested then set a flag to make sure we
4001 * send an acknowledgement for this packet */
4002 if (flags & RX_REQUEST_ACK) {
4003 ackNeeded = RX_ACK_REQUESTED;
4006 /* Keep track of whether we have received the last packet */
4007 if (flags & RX_LAST_PACKET) {
4008 call->flags |= RX_CALL_HAVE_LAST;
4012 /* Check whether we have all of the packets for this call */
4013 if (call->flags & RX_CALL_HAVE_LAST) {
4014 afs_uint32 tseq; /* temporary sequence number */
4015 struct opr_queue *cursor;
4017 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
4018 struct rx_packet *tp;
4020 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4021 if (tseq != tp->header.seq)
4023 if (tp->header.flags & RX_LAST_PACKET) {
4024 call->flags |= RX_CALL_RECEIVE_DONE;
4031 /* Provide asynchronous notification for those who want it
4032 * (e.g. multi rx) */
4033 if (call->arrivalProc) {
4034 (*call->arrivalProc) (call, call->arrivalProcHandle,
4035 call->arrivalProcArg);
4036 call->arrivalProc = (void (*)())0;
4039 /* Update last packet received */
4042 /* If there is no server process serving this call, grab
4043 * one, if available. We only need to do this once. If a
4044 * server thread is available, this thread becomes a server
4045 * thread and the server thread becomes a listener thread. */
4047 TryAttach(call, socket, tnop, newcallp, 0);
4050 /* This is not the expected next packet. */
4052 /* Determine whether this is a new or old packet, and if it's
4053 * a new one, whether it fits into the current receive window.
4054 * Also figure out whether the packet was delivered in sequence.
4055 * We use the prev variable to determine whether the new packet
4056 * is the successor of its immediate predecessor in the
4057 * receive queue, and the missing flag to determine whether
4058 * any of this packets predecessors are missing. */
4060 afs_uint32 prev; /* "Previous packet" sequence number */
4061 struct opr_queue *cursor;
4062 int missing; /* Are any predecessors missing? */
4064 /* If the new packet's sequence number has been sent to the
4065 * application already, then this is a duplicate */
4066 if (seq < call->rnext) {
4067 if (rx_stats_active)
4068 rx_atomic_inc(&rx_stats.dupPacketsRead);
4069 rxi_CancelDelayedAckEvent(call);
4070 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4076 /* If the sequence number is greater than what can be
4077 * accomodated by the current window, then send a negative
4078 * acknowledge and drop the packet */
4079 if ((call->rnext + call->rwind) <= seq) {
4080 rxi_CancelDelayedAckEvent(call);
4081 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4088 /* Look for the packet in the queue of old received packets */
4089 prev = call->rnext - 1;
4091 for (opr_queue_Scan(&call->rq, cursor)) {
4092 struct rx_packet *tp
4093 = opr_queue_Entry(cursor, struct rx_packet, entry);
4095 /*Check for duplicate packet */
4096 if (seq == tp->header.seq) {
4097 if (rx_stats_active)
4098 rx_atomic_inc(&rx_stats.dupPacketsRead);
4099 rxi_CancelDelayedAckEvent(call);
4100 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4106 /* If we find a higher sequence packet, break out and
4107 * insert the new packet here. */
4108 if (seq < tp->header.seq)
4110 /* Check for missing packet */
4111 if (tp->header.seq != prev + 1) {
4115 prev = tp->header.seq;
4118 /* Keep track of whether we have received the last packet. */
4119 if (flags & RX_LAST_PACKET) {
4120 call->flags |= RX_CALL_HAVE_LAST;
4123 /* It's within the window: add it to the the receive queue.
4124 * tp is left by the previous loop either pointing at the
4125 * packet before which to insert the new packet, or at the
4126 * queue head if the queue is empty or the packet should be
4128 #ifdef RX_TRACK_PACKETS
4129 np->flags |= RX_PKTFLAG_RQ;
4131 #ifdef RXDEBUG_PACKET
4133 #endif /* RXDEBUG_PACKET */
4134 opr_queue_InsertBefore(cursor, &np->entry);
4138 /* Check whether we have all of the packets for this call */
4139 if ((call->flags & RX_CALL_HAVE_LAST)
4140 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4141 afs_uint32 tseq; /* temporary sequence number */
4144 for (opr_queue_Scan(&call->rq, cursor)) {
4145 struct rx_packet *tp
4146 = opr_queue_Entry(cursor, struct rx_packet, entry);
4147 if (tseq != tp->header.seq)
4149 if (tp->header.flags & RX_LAST_PACKET) {
4150 call->flags |= RX_CALL_RECEIVE_DONE;
4157 /* We need to send an ack of the packet is out of sequence,
4158 * or if an ack was requested by the peer. */
4159 if (seq != prev + 1 || missing) {
4160 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4161 } else if (flags & RX_REQUEST_ACK) {
4162 ackNeeded = RX_ACK_REQUESTED;
4165 /* Acknowledge the last packet for each call */
4166 if (flags & RX_LAST_PACKET) {
4177 * If the receiver is waiting for an iovec, fill the iovec
4178 * using the data from the receive queue */
4179 if (call->flags & RX_CALL_IOVEC_WAIT) {
4180 didHardAck = rxi_FillReadVec(call, serial);
4181 /* the call may have been aborted */
4190 /* Wakeup the reader if any */
4191 if ((call->flags & RX_CALL_READER_WAIT)
4192 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4193 || (call->iovNext >= call->iovMax)
4194 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4195 call->flags &= ~RX_CALL_READER_WAIT;
4196 #ifdef RX_ENABLE_LOCKS
4197 CV_BROADCAST(&call->cv_rq);
4199 osi_rxWakeup(&call->rq);
4205 * Send an ack when requested by the peer, or once every
4206 * rxi_SoftAckRate packets until the last packet has been
4207 * received. Always send a soft ack for the last packet in
4208 * the server's reply. */
4210 rxi_CancelDelayedAckEvent(call);
4211 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4212 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4213 rxi_CancelDelayedAckEvent(call);
4214 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4215 } else if (call->nSoftAcks) {
4216 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4217 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4219 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4220 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4221 rxi_CancelDelayedAckEvent(call);
4228 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4230 struct rx_peer *peer = conn->peer;
4232 MUTEX_ENTER(&peer->peer_lock);
4233 peer->lastReachTime = clock_Sec();
4234 MUTEX_EXIT(&peer->peer_lock);
4236 MUTEX_ENTER(&conn->conn_data_lock);
4237 if (conn->flags & RX_CONN_ATTACHWAIT) {
4240 rxi_ConnClearAttachWait(conn);
4241 MUTEX_EXIT(&conn->conn_data_lock);
4243 for (i = 0; i < RX_MAXCALLS; i++) {
4244 struct rx_call *call = conn->call[i];
4247 MUTEX_ENTER(&call->lock);
4248 /* tnop can be null if newcallp is null */
4249 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4251 MUTEX_EXIT(&call->lock);
4255 MUTEX_EXIT(&conn->conn_data_lock);
4258 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4260 rx_ack_reason(int reason)
4263 case RX_ACK_REQUESTED:
4265 case RX_ACK_DUPLICATE:
4267 case RX_ACK_OUT_OF_SEQUENCE:
4269 case RX_ACK_EXCEEDS_WINDOW:
4271 case RX_ACK_NOSPACE:
4275 case RX_ACK_PING_RESPONSE:
4288 /* The real smarts of the whole thing. */
4289 static struct rx_packet *
4290 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4293 struct rx_ackPacket *ap;
4295 struct rx_packet *tp;
4296 struct rx_connection *conn = call->conn;
4297 struct rx_peer *peer = conn->peer;
4298 struct opr_queue *cursor;
4299 struct clock now; /* Current time, for RTT calculations */
4307 int newAckCount = 0;
4308 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4309 int pktsize = 0; /* Set if we need to update the peer mtu */
4310 int conn_data_locked = 0;
4312 if (rx_stats_active)
4313 rx_atomic_inc(&rx_stats.ackPacketsRead);
4314 ap = (struct rx_ackPacket *)rx_DataOf(np);
4315 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4317 return np; /* truncated ack packet */
4319 /* depends on ack packet struct */
4320 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4321 first = ntohl(ap->firstPacket);
4322 prev = ntohl(ap->previousPacket);
4323 serial = ntohl(ap->serial);
4326 * Ignore ack packets received out of order while protecting
4327 * against peers that set the previousPacket field to a packet
4328 * serial number instead of a sequence number.
4330 if (first < call->tfirst ||
4331 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4338 if (np->header.flags & RX_SLOW_START_OK) {
4339 call->flags |= RX_CALL_SLOW_START_OK;
4342 if (ap->reason == RX_ACK_PING_RESPONSE)
4343 rxi_UpdatePeerReach(conn, call);
4345 if (conn->lastPacketSizeSeq) {
4346 MUTEX_ENTER(&conn->conn_data_lock);
4347 conn_data_locked = 1;
4348 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4349 pktsize = conn->lastPacketSize;
4350 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4353 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4354 if (!conn_data_locked) {
4355 MUTEX_ENTER(&conn->conn_data_lock);
4356 conn_data_locked = 1;
4358 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4359 /* process mtu ping ack */
4360 pktsize = conn->lastPingSize;
4361 conn->lastPingSizeSer = conn->lastPingSize = 0;
4365 if (conn_data_locked) {
4366 MUTEX_EXIT(&conn->conn_data_lock);
4367 conn_data_locked = 0;
4371 if (rxdebug_active) {
4375 len = _snprintf(msg, sizeof(msg),
4376 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4377 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4378 ntohl(ap->serial), ntohl(ap->previousPacket),
4379 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4380 ap->nAcks, ntohs(ap->bufferSpace) );
4384 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4385 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4389 OutputDebugString(msg);
4391 #else /* AFS_NT40_ENV */
4394 "RACK: reason %x previous %u seq %u serial %u first %u",
4395 ap->reason, ntohl(ap->previousPacket),
4396 (unsigned int)np->header.seq, (unsigned int)serial,
4397 ntohl(ap->firstPacket));
4400 for (offset = 0; offset < nAcks; offset++)
4401 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4406 #endif /* AFS_NT40_ENV */
4409 MUTEX_ENTER(&peer->peer_lock);
4412 * Start somewhere. Can't assume we can send what we can receive,
4413 * but we are clearly receiving.
4415 if (!peer->maxPacketSize)
4416 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4418 if (pktsize > peer->maxPacketSize) {
4419 peer->maxPacketSize = pktsize;
4420 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4421 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4422 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4423 rxi_ScheduleGrowMTUEvent(call, 1);
4428 clock_GetTime(&now);
4430 /* The transmit queue splits into 4 sections.
4432 * The first section is packets which have now been acknowledged
4433 * by a window size change in the ack. These have reached the
4434 * application layer, and may be discarded. These are packets
4435 * with sequence numbers < ap->firstPacket.
4437 * The second section is packets which have sequence numbers in
4438 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4439 * contents of the packet's ack array determines whether these
4440 * packets are acknowledged or not.
4442 * The third section is packets which fall above the range
4443 * addressed in the ack packet. These have not yet been received
4446 * The four section is packets which have not yet been transmitted.
4447 * These packets will have a header.serial of 0.
4450 /* First section - implicitly acknowledged packets that can be
4454 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4455 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4456 struct rx_packet *next;
4458 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4459 call->tfirst = tp->header.seq + 1;
4461 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4463 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4466 #ifdef RX_ENABLE_LOCKS
4467 /* XXX Hack. Because we have to release the global call lock when sending
4468 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4469 * in rxi_Start sending packets out because packets may move to the
4470 * freePacketQueue as result of being here! So we drop these packets until
4471 * we're safely out of the traversing. Really ugly!
4472 * To make it even uglier, if we're using fine grain locking, we can
4473 * set the ack bits in the packets and have rxi_Start remove the packets
4474 * when it's done transmitting.
4476 if (call->flags & RX_CALL_TQ_BUSY) {
4477 tp->flags |= RX_PKTFLAG_ACKED;
4478 call->flags |= RX_CALL_TQ_SOME_ACKED;
4480 #endif /* RX_ENABLE_LOCKS */
4482 opr_queue_Remove(&tp->entry);
4483 #ifdef RX_TRACK_PACKETS
4484 tp->flags &= ~RX_PKTFLAG_TQ;
4486 #ifdef RXDEBUG_PACKET
4488 #endif /* RXDEBUG_PACKET */
4489 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4494 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4496 /* Second section of the queue - packets for which we are receiving
4499 * Go through the explicit acks/nacks and record the results in
4500 * the waiting packets. These are packets that can't be released
4501 * yet, even with a positive acknowledge. This positive
4502 * acknowledge only means the packet has been received by the
4503 * peer, not that it will be retained long enough to be sent to
4504 * the peer's upper level. In addition, reset the transmit timers
4505 * of any missing packets (those packets that must be missing
4506 * because this packet was out of sequence) */
4508 call->nSoftAcked = 0;
4510 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4511 && tp->header.seq < first + nAcks) {
4512 /* Set the acknowledge flag per packet based on the
4513 * information in the ack packet. An acknowlegded packet can
4514 * be downgraded when the server has discarded a packet it
4515 * soacked previously, or when an ack packet is received
4516 * out of sequence. */
4517 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4518 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4520 tp->flags |= RX_PKTFLAG_ACKED;
4521 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4528 } else /* RX_ACK_TYPE_NACK */ {
4529 tp->flags &= ~RX_PKTFLAG_ACKED;
4533 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4536 /* We don't need to take any action with the 3rd or 4th section in the
4537 * queue - they're not addressed by the contents of this ACK packet.
4540 /* If the window has been extended by this acknowledge packet,
4541 * then wakeup a sender waiting in alloc for window space, or try
4542 * sending packets now, if he's been sitting on packets due to
4543 * lack of window space */
4544 if (call->tnext < (call->tfirst + call->twind)) {
4545 #ifdef RX_ENABLE_LOCKS
4546 CV_SIGNAL(&call->cv_twind);
4548 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4549 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4550 osi_rxWakeup(&call->twind);
4553 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4554 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4558 /* if the ack packet has a receivelen field hanging off it,
4559 * update our state */
4560 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4563 /* If the ack packet has a "recommended" size that is less than
4564 * what I am using now, reduce my size to match */
4565 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4566 (int)sizeof(afs_int32), &tSize);
4567 tSize = (afs_uint32) ntohl(tSize);
4568 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4570 /* Get the maximum packet size to send to this peer */
4571 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4573 tSize = (afs_uint32) ntohl(tSize);
4574 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4575 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4577 /* sanity check - peer might have restarted with different params.
4578 * If peer says "send less", dammit, send less... Peer should never
4579 * be unable to accept packets of the size that prior AFS versions would
4580 * send without asking. */
4581 if (peer->maxMTU != tSize) {
4582 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4584 peer->maxMTU = tSize;
4585 peer->MTU = MIN(tSize, peer->MTU);
4586 call->MTU = MIN(call->MTU, tSize);
4589 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4592 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4593 (int)sizeof(afs_int32), &tSize);
4594 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4595 if (tSize < call->twind) { /* smaller than our send */
4596 call->twind = tSize; /* window, we must send less... */
4597 call->ssthresh = MIN(call->twind, call->ssthresh);
4598 call->conn->twind[call->channel] = call->twind;
4601 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4602 * network MTU confused with the loopback MTU. Calculate the
4603 * maximum MTU here for use in the slow start code below.
4605 /* Did peer restart with older RX version? */
4606 if (peer->maxDgramPackets > 1) {
4607 peer->maxDgramPackets = 1;
4609 } else if (np->length >=
4610 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4613 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4614 sizeof(afs_int32), &tSize);
4615 tSize = (afs_uint32) ntohl(tSize);
4617 * As of AFS 3.5 we set the send window to match the receive window.
4619 if (tSize < call->twind) {
4620 call->twind = tSize;
4621 call->conn->twind[call->channel] = call->twind;
4622 call->ssthresh = MIN(call->twind, call->ssthresh);
4623 } else if (tSize > call->twind) {
4624 call->twind = tSize;
4625 call->conn->twind[call->channel] = call->twind;
4629 * As of AFS 3.5, a jumbogram is more than one fixed size
4630 * packet transmitted in a single UDP datagram. If the remote
4631 * MTU is smaller than our local MTU then never send a datagram
4632 * larger than the natural MTU.
4635 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4636 (int)sizeof(afs_int32), &tSize);
4637 maxDgramPackets = (afs_uint32) ntohl(tSize);
4638 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4640 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4641 if (maxDgramPackets > 1) {
4642 peer->maxDgramPackets = maxDgramPackets;
4643 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4645 peer->maxDgramPackets = 1;
4646 call->MTU = peer->natMTU;
4648 } else if (peer->maxDgramPackets > 1) {
4649 /* Restarted with lower version of RX */
4650 peer->maxDgramPackets = 1;
4652 } else if (peer->maxDgramPackets > 1
4653 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4654 /* Restarted with lower version of RX */
4655 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4656 peer->natMTU = OLD_MAX_PACKET_SIZE;
4657 peer->MTU = OLD_MAX_PACKET_SIZE;
4658 peer->maxDgramPackets = 1;
4659 peer->nDgramPackets = 1;
4661 call->MTU = OLD_MAX_PACKET_SIZE;
4666 * Calculate how many datagrams were successfully received after
4667 * the first missing packet and adjust the negative ack counter
4672 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4673 if (call->nNacks < nNacked) {
4674 call->nNacks = nNacked;
4677 call->nAcks += newAckCount;
4681 /* If the packet contained new acknowledgements, rather than just
4682 * being a duplicate of one we have previously seen, then we can restart
4685 if (newAckCount > 0)
4686 rxi_rto_packet_acked(call, istack);
4688 if (call->flags & RX_CALL_FAST_RECOVER) {
4689 if (newAckCount == 0) {
4690 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4692 call->flags &= ~RX_CALL_FAST_RECOVER;
4693 call->cwind = call->nextCwind;
4694 call->nextCwind = 0;
4697 call->nCwindAcks = 0;
4698 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4699 /* Three negative acks in a row trigger congestion recovery */
4700 call->flags |= RX_CALL_FAST_RECOVER;
4701 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4703 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4704 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4705 call->nextCwind = call->ssthresh;
4708 peer->MTU = call->MTU;
4709 peer->cwind = call->nextCwind;
4710 peer->nDgramPackets = call->nDgramPackets;
4712 call->congestSeq = peer->congestSeq;
4714 /* Reset the resend times on the packets that were nacked
4715 * so we will retransmit as soon as the window permits
4719 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4720 struct rx_packet *tp =
4721 opr_queue_Entry(cursor, struct rx_packet, entry);
4723 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4724 tp->flags &= ~RX_PKTFLAG_SENT;
4726 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4731 /* If cwind is smaller than ssthresh, then increase
4732 * the window one packet for each ack we receive (exponential
4734 * If cwind is greater than or equal to ssthresh then increase
4735 * the congestion window by one packet for each cwind acks we
4736 * receive (linear growth). */
4737 if (call->cwind < call->ssthresh) {
4739 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4740 call->nCwindAcks = 0;
4742 call->nCwindAcks += newAckCount;
4743 if (call->nCwindAcks >= call->cwind) {
4744 call->nCwindAcks = 0;
4745 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4749 * If we have received several acknowledgements in a row then
4750 * it is time to increase the size of our datagrams
4752 if ((int)call->nAcks > rx_nDgramThreshold) {
4753 if (peer->maxDgramPackets > 1) {
4754 if (call->nDgramPackets < peer->maxDgramPackets) {
4755 call->nDgramPackets++;
4757 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4758 } else if (call->MTU < peer->maxMTU) {
4759 /* don't upgrade if we can't handle it */
4760 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4761 call->MTU = peer->ifMTU;
4763 call->MTU += peer->natMTU;
4764 call->MTU = MIN(call->MTU, peer->maxMTU);
4771 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4773 /* Servers need to hold the call until all response packets have
4774 * been acknowledged. Soft acks are good enough since clients
4775 * are not allowed to clear their receive queues. */
4776 if (call->state == RX_STATE_HOLD
4777 && call->tfirst + call->nSoftAcked >= call->tnext) {
4778 call->state = RX_STATE_DALLY;
4779 rxi_ClearTransmitQueue(call, 0);
4780 rxi_CancelKeepAliveEvent(call);
4781 } else if (!opr_queue_IsEmpty(&call->tq)) {
4782 rxi_Start(call, istack);
4787 /* Received a response to a challenge packet */
4788 static struct rx_packet *
4789 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4790 struct rx_packet *np, int istack)
4794 /* Ignore the packet if we're the client */
4795 if (conn->type == RX_CLIENT_CONNECTION)
4798 /* If already authenticated, ignore the packet (it's probably a retry) */
4799 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4802 /* Otherwise, have the security object evaluate the response packet */
4803 error = RXS_CheckResponse(conn->securityObject, conn, np);
4805 /* If the response is invalid, reset the connection, sending
4806 * an abort to the peer */
4810 rxi_ConnectionError(conn, error);
4811 MUTEX_ENTER(&conn->conn_data_lock);
4812 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4813 MUTEX_EXIT(&conn->conn_data_lock);
4816 /* If the response is valid, any calls waiting to attach
4817 * servers can now do so */
4820 for (i = 0; i < RX_MAXCALLS; i++) {
4821 struct rx_call *call = conn->call[i];
4823 MUTEX_ENTER(&call->lock);
4824 if (call->state == RX_STATE_PRECALL)
4825 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4826 /* tnop can be null if newcallp is null */
4827 MUTEX_EXIT(&call->lock);
4831 /* Update the peer reachability information, just in case
4832 * some calls went into attach-wait while we were waiting
4833 * for authentication..
4835 rxi_UpdatePeerReach(conn, NULL);
4840 /* A client has received an authentication challenge: the security
4841 * object is asked to cough up a respectable response packet to send
4842 * back to the server. The server is responsible for retrying the
4843 * challenge if it fails to get a response. */
4845 static struct rx_packet *
4846 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4847 struct rx_packet *np, int istack)
4851 /* Ignore the challenge if we're the server */
4852 if (conn->type == RX_SERVER_CONNECTION)
4855 /* Ignore the challenge if the connection is otherwise idle; someone's
4856 * trying to use us as an oracle. */
4857 if (!rxi_HasActiveCalls(conn))
4860 /* Send the security object the challenge packet. It is expected to fill
4861 * in the response. */
4862 error = RXS_GetResponse(conn->securityObject, conn, np);
4864 /* If the security object is unable to return a valid response, reset the
4865 * connection and send an abort to the peer. Otherwise send the response
4866 * packet to the peer connection. */
4868 rxi_ConnectionError(conn, error);
4869 MUTEX_ENTER(&conn->conn_data_lock);
4870 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4871 MUTEX_EXIT(&conn->conn_data_lock);
4873 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4874 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4880 /* Find an available server process to service the current request in
4881 * the given call structure. If one isn't available, queue up this
4882 * call so it eventually gets one */
4884 rxi_AttachServerProc(struct rx_call *call,
4885 osi_socket socket, int *tnop,
4886 struct rx_call **newcallp)
4888 struct rx_serverQueueEntry *sq;
4889 struct rx_service *service = call->conn->service;
4892 /* May already be attached */
4893 if (call->state == RX_STATE_ACTIVE)
4896 MUTEX_ENTER(&rx_serverPool_lock);
4898 haveQuota = QuotaOK(service);
4899 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4900 /* If there are no processes available to service this call,
4901 * put the call on the incoming call queue (unless it's
4902 * already on the queue).
4904 #ifdef RX_ENABLE_LOCKS
4906 ReturnToServerPool(service);
4907 #endif /* RX_ENABLE_LOCKS */
4909 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4910 call->flags |= RX_CALL_WAIT_PROC;
4911 rx_atomic_inc(&rx_nWaiting);
4912 rx_atomic_inc(&rx_nWaited);
4913 rxi_calltrace(RX_CALL_ARRIVAL, call);
4914 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4915 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4918 sq = opr_queue_Last(&rx_idleServerQueue,
4919 struct rx_serverQueueEntry, entry);
4921 /* If hot threads are enabled, and both newcallp and sq->socketp
4922 * are non-null, then this thread will process the call, and the
4923 * idle server thread will start listening on this threads socket.
4925 opr_queue_Remove(&sq->entry);
4927 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4930 *sq->socketp = socket;
4931 clock_GetTime(&call->startTime);
4932 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4936 if (call->flags & RX_CALL_WAIT_PROC) {
4937 /* Conservative: I don't think this should happen */
4938 call->flags &= ~RX_CALL_WAIT_PROC;
4939 rx_atomic_dec(&rx_nWaiting);
4940 if (opr_queue_IsOnQueue(&call->entry)) {
4941 opr_queue_Remove(&call->entry);
4944 call->state = RX_STATE_ACTIVE;
4945 call->app.mode = RX_MODE_RECEIVING;
4946 #ifdef RX_KERNEL_TRACE
4948 int glockOwner = ISAFS_GLOCK();
4951 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4952 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4958 if (call->flags & RX_CALL_CLEARED) {
4959 /* send an ack now to start the packet flow up again */
4960 call->flags &= ~RX_CALL_CLEARED;
4961 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4963 #ifdef RX_ENABLE_LOCKS
4966 service->nRequestsRunning++;
4967 MUTEX_ENTER(&rx_quota_mutex);
4968 if (service->nRequestsRunning <= service->minProcs)
4971 MUTEX_EXIT(&rx_quota_mutex);
4975 MUTEX_EXIT(&rx_serverPool_lock);
4978 /* Delay the sending of an acknowledge event for a short while, while
4979 * a new call is being prepared (in the case of a client) or a reply
4980 * is being prepared (in the case of a server). Rather than sending
4981 * an ack packet, an ACKALL packet is sent. */
4983 rxi_AckAll(struct rx_call *call)
4985 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4987 call->flags |= RX_CALL_ACKALL_SENT;
4991 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4994 struct rx_call *call = arg1;
4995 #ifdef RX_ENABLE_LOCKS
4997 MUTEX_ENTER(&call->lock);
4998 if (event == call->delayedAckEvent)
4999 rxevent_Put(&call->delayedAckEvent);
5000 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
5002 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5004 MUTEX_EXIT(&call->lock);
5005 #else /* RX_ENABLE_LOCKS */
5007 rxevent_Put(&call->delayedAckEvent);
5008 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5009 #endif /* RX_ENABLE_LOCKS */
5012 #ifdef RX_ENABLE_LOCKS
5013 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5014 * clearing them out.
5017 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5019 struct opr_queue *cursor;
5022 for (opr_queue_Scan(&call->tq, cursor)) {
5024 = opr_queue_Entry(cursor, struct rx_packet, entry);
5026 p->flags |= RX_PKTFLAG_ACKED;
5031 call->flags |= RX_CALL_TQ_CLEARME;
5032 call->flags |= RX_CALL_TQ_SOME_ACKED;
5035 rxi_rto_cancel(call);
5037 call->tfirst = call->tnext;
5038 call->nSoftAcked = 0;
5040 if (call->flags & RX_CALL_FAST_RECOVER) {
5041 call->flags &= ~RX_CALL_FAST_RECOVER;
5042 call->cwind = call->nextCwind;
5043 call->nextCwind = 0;
5046 CV_SIGNAL(&call->cv_twind);
5048 #endif /* RX_ENABLE_LOCKS */
5051 * Acknowledge the whole transmit queue.
5053 * If we're running without locks, or the transmit queue isn't busy, then
5054 * we can just clear the queue now. Otherwise, we have to mark all of the
5055 * packets as acknowledged, and let rxi_Start clear it later on
5058 rxi_AckAllInTransmitQueue(struct rx_call *call)
5060 #ifdef RX_ENABLE_LOCKS
5061 if (call->flags & RX_CALL_TQ_BUSY) {
5062 rxi_SetAcksInTransmitQueue(call);
5066 rxi_ClearTransmitQueue(call, 0);
5068 /* Clear out the transmit queue for the current call (all packets have
5069 * been received by peer) */
5071 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5073 #ifdef RX_ENABLE_LOCKS
5074 struct opr_queue *cursor;
5075 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5077 for (opr_queue_Scan(&call->tq, cursor)) {
5079 = opr_queue_Entry(cursor, struct rx_packet, entry);
5081 p->flags |= RX_PKTFLAG_ACKED;
5085 call->flags |= RX_CALL_TQ_CLEARME;
5086 call->flags |= RX_CALL_TQ_SOME_ACKED;
5089 #endif /* RX_ENABLE_LOCKS */
5090 #ifdef RXDEBUG_PACKET
5092 #endif /* RXDEBUG_PACKET */
5093 rxi_FreePackets(0, &call->tq);
5094 rxi_WakeUpTransmitQueue(call);
5095 #ifdef RX_ENABLE_LOCKS
5096 call->flags &= ~RX_CALL_TQ_CLEARME;
5100 rxi_rto_cancel(call);
5101 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5102 call->nSoftAcked = 0;
5104 if (call->flags & RX_CALL_FAST_RECOVER) {
5105 call->flags &= ~RX_CALL_FAST_RECOVER;
5106 call->cwind = call->nextCwind;
5108 #ifdef RX_ENABLE_LOCKS
5109 CV_SIGNAL(&call->cv_twind);
5111 osi_rxWakeup(&call->twind);
5116 rxi_ClearReceiveQueue(struct rx_call *call)
5118 if (!opr_queue_IsEmpty(&call->rq)) {
5121 count = rxi_FreePackets(0, &call->rq);
5122 rx_packetReclaims += count;
5123 #ifdef RXDEBUG_PACKET
5125 if ( call->rqc != 0 )
5126 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5128 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5130 if (call->state == RX_STATE_PRECALL) {
5131 call->flags |= RX_CALL_CLEARED;
5135 /* Send an abort packet for the specified call */
5136 static struct rx_packet *
5137 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5138 int istack, int force)
5140 afs_int32 error, cerror;
5141 struct clock when, now;
5146 switch (call->error) {
5149 cerror = RX_CALL_TIMEOUT;
5152 cerror = call->error;
5155 /* Clients should never delay abort messages */
5156 if (rx_IsClientConn(call->conn))
5159 if (call->abortCode != cerror) {
5160 call->abortCode = cerror;
5161 call->abortCount = 0;
5164 if (force || rxi_callAbortThreshhold == 0
5165 || call->abortCount < rxi_callAbortThreshhold) {
5166 rxi_CancelDelayedAbortEvent(call);
5167 error = htonl(cerror);
5170 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5171 (char *)&error, sizeof(error), istack);
5172 } else if (!call->delayedAbortEvent) {
5173 clock_GetTime(&now);
5175 clock_Addmsec(&when, rxi_callAbortDelay);
5176 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5177 call->delayedAbortEvent =
5178 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5184 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5186 if (call->delayedAbortEvent) {
5187 rxevent_Cancel(&call->delayedAbortEvent);
5188 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5192 /* Send an abort packet for the specified connection. Packet is an
5193 * optional pointer to a packet that can be used to send the abort.
5194 * Once the number of abort messages reaches the threshhold, an
5195 * event is scheduled to send the abort. Setting the force flag
5196 * overrides sending delayed abort messages.
5198 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5199 * to send the abort packet.
5202 rxi_SendConnectionAbort(struct rx_connection *conn,
5203 struct rx_packet *packet, int istack, int force)
5206 struct clock when, now;
5211 /* Clients should never delay abort messages */
5212 if (rx_IsClientConn(conn))
5215 if (force || rxi_connAbortThreshhold == 0
5216 || conn->abortCount < rxi_connAbortThreshhold) {
5218 rxevent_Cancel(&conn->delayedAbortEvent);
5219 error = htonl(conn->error);
5221 MUTEX_EXIT(&conn->conn_data_lock);
5223 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5224 RX_PACKET_TYPE_ABORT, (char *)&error,
5225 sizeof(error), istack);
5226 MUTEX_ENTER(&conn->conn_data_lock);
5227 } else if (!conn->delayedAbortEvent) {
5228 clock_GetTime(&now);
5230 clock_Addmsec(&when, rxi_connAbortDelay);
5231 conn->delayedAbortEvent =
5232 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5237 /* Associate an error all of the calls owned by a connection. Called
5238 * with error non-zero. This is only for really fatal things, like
5239 * bad authentication responses. The connection itself is set in
5240 * error at this point, so that future packets received will be
5243 rxi_ConnectionError(struct rx_connection *conn,
5249 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5251 MUTEX_ENTER(&conn->conn_data_lock);
5252 rxevent_Cancel(&conn->challengeEvent);
5253 rxevent_Cancel(&conn->natKeepAliveEvent);
5254 if (conn->checkReachEvent) {
5255 rxevent_Cancel(&conn->checkReachEvent);
5256 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5257 putConnection(conn);
5259 MUTEX_EXIT(&conn->conn_data_lock);
5260 for (i = 0; i < RX_MAXCALLS; i++) {
5261 struct rx_call *call = conn->call[i];
5263 MUTEX_ENTER(&call->lock);
5264 rxi_CallError(call, error);
5265 MUTEX_EXIT(&call->lock);
5268 conn->error = error;
5269 if (rx_stats_active)
5270 rx_atomic_inc(&rx_stats.fatalErrors);
5275 * Interrupt an in-progress call with the specified error and wakeup waiters.
5277 * @param[in] call The call to interrupt
5278 * @param[in] error The error code to send to the peer
5281 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5283 MUTEX_ENTER(&call->lock);
5284 rxi_CallError(call, error);
5285 rxi_SendCallAbort(call, NULL, 0, 1);
5286 MUTEX_EXIT(&call->lock);
5290 rxi_CallError(struct rx_call *call, afs_int32 error)
5292 MUTEX_ASSERT(&call->lock);
5293 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5295 error = call->error;
5297 #ifdef RX_ENABLE_LOCKS
5298 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5299 rxi_ResetCall(call, 0);
5302 rxi_ResetCall(call, 0);
5304 call->error = error;
5307 /* Reset various fields in a call structure, and wakeup waiting
5308 * processes. Some fields aren't changed: state & mode are not
5309 * touched (these must be set by the caller), and bufptr, nLeft, and
5310 * nFree are not reset, since these fields are manipulated by
5311 * unprotected macros, and may only be reset by non-interrupting code.
5315 rxi_ResetCall(struct rx_call *call, int newcall)
5318 struct rx_peer *peer;
5319 struct rx_packet *packet;
5321 MUTEX_ASSERT(&call->lock);
5322 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5324 /* Notify anyone who is waiting for asynchronous packet arrival */
5325 if (call->arrivalProc) {
5326 (*call->arrivalProc) (call, call->arrivalProcHandle,
5327 call->arrivalProcArg);
5328 call->arrivalProc = (void (*)())0;
5332 rxi_CancelGrowMTUEvent(call);
5334 if (call->delayedAbortEvent) {
5335 rxi_CancelDelayedAbortEvent(call);
5336 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5338 rxi_SendCallAbort(call, packet, 0, 1);
5339 rxi_FreePacket(packet);
5344 * Update the peer with the congestion information in this call
5345 * so other calls on this connection can pick up where this call
5346 * left off. If the congestion sequence numbers don't match then
5347 * another call experienced a retransmission.
5349 peer = call->conn->peer;
5350 MUTEX_ENTER(&peer->peer_lock);
5352 if (call->congestSeq == peer->congestSeq) {
5353 peer->cwind = MAX(peer->cwind, call->cwind);
5354 peer->MTU = MAX(peer->MTU, call->MTU);
5355 peer->nDgramPackets =
5356 MAX(peer->nDgramPackets, call->nDgramPackets);
5359 call->abortCode = 0;
5360 call->abortCount = 0;
5362 if (peer->maxDgramPackets > 1) {
5363 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5365 call->MTU = peer->MTU;
5367 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5368 call->ssthresh = rx_maxSendWindow;
5369 call->nDgramPackets = peer->nDgramPackets;
5370 call->congestSeq = peer->congestSeq;
5371 call->rtt = peer->rtt;
5372 call->rtt_dev = peer->rtt_dev;
5373 clock_Zero(&call->rto);
5374 clock_Addmsec(&call->rto,
5375 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5376 MUTEX_EXIT(&peer->peer_lock);
5378 flags = call->flags;
5379 rxi_WaitforTQBusy(call);
5381 rxi_ClearTransmitQueue(call, 1);
5382 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5383 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5387 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5388 /* The call channel is still busy; resetting the call doesn't change
5389 * that. However, if 'newcall' is set, we are processing a call
5390 * structure that has either been recycled from the free list, or has
5391 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5392 * 'newcall' is set, since it describes a completely different call
5393 * channel which we do not care about. */
5394 call->flags |= RX_CALL_PEER_BUSY;
5397 rxi_ClearReceiveQueue(call);
5398 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5402 call->twind = call->conn->twind[call->channel];
5403 call->rwind = call->conn->rwind[call->channel];
5404 call->nSoftAcked = 0;
5405 call->nextCwind = 0;
5408 call->nCwindAcks = 0;
5409 call->nSoftAcks = 0;
5410 call->nHardAcks = 0;
5412 call->tfirst = call->rnext = call->tnext = 1;
5415 call->lastAcked = 0;
5416 call->localStatus = call->remoteStatus = 0;
5418 if (flags & RX_CALL_READER_WAIT) {
5419 #ifdef RX_ENABLE_LOCKS
5420 CV_BROADCAST(&call->cv_rq);
5422 osi_rxWakeup(&call->rq);
5425 if (flags & RX_CALL_WAIT_PACKETS) {
5426 MUTEX_ENTER(&rx_freePktQ_lock);
5427 rxi_PacketsUnWait(); /* XXX */
5428 MUTEX_EXIT(&rx_freePktQ_lock);
5430 #ifdef RX_ENABLE_LOCKS
5431 CV_SIGNAL(&call->cv_twind);
5433 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5434 osi_rxWakeup(&call->twind);
5437 if (flags & RX_CALL_WAIT_PROC) {
5438 rx_atomic_dec(&rx_nWaiting);
5440 #ifdef RX_ENABLE_LOCKS
5441 /* The following ensures that we don't mess with any queue while some
5442 * other thread might also be doing so. The call_queue_lock field is
5443 * is only modified under the call lock. If the call is in the process
5444 * of being removed from a queue, the call is not locked until the
5445 * the queue lock is dropped and only then is the call_queue_lock field
5446 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5447 * Note that any other routine which removes a call from a queue has to
5448 * obtain the queue lock before examing the queue and removing the call.
5450 if (call->call_queue_lock) {
5451 MUTEX_ENTER(call->call_queue_lock);
5452 if (opr_queue_IsOnQueue(&call->entry)) {
5453 opr_queue_Remove(&call->entry);
5455 MUTEX_EXIT(call->call_queue_lock);
5456 CLEAR_CALL_QUEUE_LOCK(call);
5458 #else /* RX_ENABLE_LOCKS */
5459 if (opr_queue_IsOnQueue(&call->entry)) {
5460 opr_queue_Remove(&call->entry);
5462 #endif /* RX_ENABLE_LOCKS */
5464 rxi_CancelKeepAliveEvent(call);
5465 rxi_CancelDelayedAckEvent(call);
5468 /* Send an acknowledge for the indicated packet (seq,serial) of the
5469 * indicated call, for the indicated reason (reason). This
5470 * acknowledge will specifically acknowledge receiving the packet, and
5471 * will also specify which other packets for this call have been
5472 * received. This routine returns the packet that was used to the
5473 * caller. The caller is responsible for freeing it or re-using it.
5474 * This acknowledgement also returns the highest sequence number
5475 * actually read out by the higher level to the sender; the sender
5476 * promises to keep around packets that have not been read by the
5477 * higher level yet (unless, of course, the sender decides to abort
5478 * the call altogether). Any of p, seq, serial, pflags, or reason may
5479 * be set to zero without ill effect. That is, if they are zero, they
5480 * will not convey any information.
5481 * NOW there is a trailer field, after the ack where it will safely be
5482 * ignored by mundanes, which indicates the maximum size packet this
5483 * host can swallow. */
5485 struct rx_packet *optionalPacket; use to send ack (or null)
5486 int seq; Sequence number of the packet we are acking
5487 int serial; Serial number of the packet
5488 int pflags; Flags field from packet header
5489 int reason; Reason an acknowledge was prompted
5493 rxi_SendAck(struct rx_call *call,
5494 struct rx_packet *optionalPacket, int serial, int reason,
5497 struct rx_ackPacket *ap;
5498 struct rx_packet *p;
5499 struct opr_queue *cursor;
5502 afs_uint32 padbytes = 0;
5503 #ifdef RX_ENABLE_TSFPQ
5504 struct rx_ts_info_t * rx_ts_info;
5508 * Open the receive window once a thread starts reading packets
5510 if (call->rnext > 1) {
5511 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5514 /* Don't attempt to grow MTU if this is a critical ping */
5515 if (reason == RX_ACK_MTU) {
5516 /* keep track of per-call attempts, if we're over max, do in small
5517 * otherwise in larger? set a size to increment by, decrease
5520 if (call->conn->peer->maxPacketSize &&
5521 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5523 padbytes = call->conn->peer->maxPacketSize+16;
5525 padbytes = call->conn->peer->maxMTU + 128;
5527 /* do always try a minimum size ping */
5528 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5530 /* subtract the ack payload */
5531 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5532 reason = RX_ACK_PING;
5535 call->nHardAcks = 0;
5536 call->nSoftAcks = 0;
5537 if (call->rnext > call->lastAcked)
5538 call->lastAcked = call->rnext;
5542 rx_computelen(p, p->length); /* reset length, you never know */
5543 } /* where that's been... */
5544 #ifdef RX_ENABLE_TSFPQ
5546 RX_TS_INFO_GET(rx_ts_info);
5547 if ((p = rx_ts_info->local_special_packet)) {
5548 rx_computelen(p, p->length);
5549 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5550 rx_ts_info->local_special_packet = p;
5551 } else { /* We won't send the ack, but don't panic. */
5552 return optionalPacket;
5556 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5557 /* We won't send the ack, but don't panic. */
5558 return optionalPacket;
5563 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5566 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5567 #ifndef RX_ENABLE_TSFPQ
5568 if (!optionalPacket)
5571 return optionalPacket;
5573 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5574 if (rx_Contiguous(p) < templ) {
5575 #ifndef RX_ENABLE_TSFPQ
5576 if (!optionalPacket)
5579 return optionalPacket;
5584 /* MTUXXX failing to send an ack is very serious. We should */
5585 /* try as hard as possible to send even a partial ack; it's */
5586 /* better than nothing. */
5587 ap = (struct rx_ackPacket *)rx_DataOf(p);
5588 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5589 ap->reason = reason;
5591 /* The skew computation used to be bogus, I think it's better now. */
5592 /* We should start paying attention to skew. XXX */
5593 ap->serial = htonl(serial);
5594 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5597 * First packet not yet forwarded to reader. When ACKALL has been
5598 * sent the peer has been told that all received packets will be
5599 * delivered to the reader. The value 'rnext' is used internally
5600 * to refer to the next packet in the receive queue that must be
5601 * delivered to the reader. From the perspective of the peer it
5602 * already has so report the last sequence number plus one if there
5603 * are packets in the receive queue awaiting processing.
5605 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5606 !opr_queue_IsEmpty(&call->rq)) {
5607 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5609 ap->firstPacket = htonl(call->rnext);
5611 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5613 /* No fear of running out of ack packet here because there can only
5614 * be at most one window full of unacknowledged packets. The window
5615 * size must be constrained to be less than the maximum ack size,
5616 * of course. Also, an ack should always fit into a single packet
5617 * -- it should not ever be fragmented. */
5619 for (opr_queue_Scan(&call->rq, cursor)) {
5620 struct rx_packet *rqp
5621 = opr_queue_Entry(cursor, struct rx_packet, entry);
5623 if (!rqp || !call->rq.next
5624 || (rqp->header.seq > (call->rnext + call->rwind))) {
5625 #ifndef RX_ENABLE_TSFPQ
5626 if (!optionalPacket)
5629 rxi_CallError(call, RX_CALL_DEAD);
5630 return optionalPacket;
5633 while (rqp->header.seq > call->rnext + offset)
5634 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5635 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5637 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5638 #ifndef RX_ENABLE_TSFPQ
5639 if (!optionalPacket)
5642 rxi_CallError(call, RX_CALL_DEAD);
5643 return optionalPacket;
5649 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5651 /* these are new for AFS 3.3 */
5652 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5653 templ = htonl(templ);
5654 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5655 templ = htonl(call->conn->peer->ifMTU);
5656 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5657 sizeof(afs_int32), &templ);
5659 /* new for AFS 3.4 */
5660 templ = htonl(call->rwind);
5661 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5662 sizeof(afs_int32), &templ);
5664 /* new for AFS 3.5 */
5665 templ = htonl(call->conn->peer->ifDgramPackets);
5666 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5667 sizeof(afs_int32), &templ);
5669 p->header.serviceId = call->conn->serviceId;
5670 p->header.cid = (call->conn->cid | call->channel);
5671 p->header.callNumber = *call->callNumber;
5673 p->header.securityIndex = call->conn->securityIndex;
5674 p->header.epoch = call->conn->epoch;
5675 p->header.type = RX_PACKET_TYPE_ACK;
5676 p->header.flags = RX_SLOW_START_OK;
5677 if (reason == RX_ACK_PING) {
5678 p->header.flags |= RX_REQUEST_ACK;
5680 p->length = padbytes +
5681 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5684 /* not fast but we can potentially use this if truncated
5685 * fragments are delivered to figure out the mtu.
5687 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5688 sizeof(afs_int32), sizeof(afs_int32),
5692 if (call->conn->type == RX_CLIENT_CONNECTION)
5693 p->header.flags |= RX_CLIENT_INITIATED;
5697 if (rxdebug_active) {
5701 len = _snprintf(msg, sizeof(msg),
5702 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5703 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5704 ntohl(ap->serial), ntohl(ap->previousPacket),
5705 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5706 ap->nAcks, ntohs(ap->bufferSpace) );
5710 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5711 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5715 OutputDebugString(msg);
5717 #else /* AFS_NT40_ENV */
5719 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5720 ap->reason, ntohl(ap->previousPacket),
5721 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5723 for (offset = 0; offset < ap->nAcks; offset++)
5724 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5729 #endif /* AFS_NT40_ENV */
5732 int i, nbytes = p->length;
5734 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5735 if (nbytes <= p->wirevec[i].iov_len) {
5738 savelen = p->wirevec[i].iov_len;
5740 p->wirevec[i].iov_len = nbytes;
5742 rxi_Send(call, p, istack);
5743 p->wirevec[i].iov_len = savelen;
5747 nbytes -= p->wirevec[i].iov_len;
5750 if (rx_stats_active)
5751 rx_atomic_inc(&rx_stats.ackPacketsSent);
5752 #ifndef RX_ENABLE_TSFPQ
5753 if (!optionalPacket)
5756 return optionalPacket; /* Return packet for re-use by caller */
5760 struct rx_packet **list;
5765 /* Send all of the packets in the list in single datagram */
5767 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5768 int istack, int moreFlag)
5774 struct rx_connection *conn = call->conn;
5775 struct rx_peer *peer = conn->peer;
5777 MUTEX_ENTER(&peer->peer_lock);
5778 peer->nSent += xmit->len;
5779 if (xmit->resending)
5780 peer->reSends += xmit->len;
5781 MUTEX_EXIT(&peer->peer_lock);
5783 if (rx_stats_active) {
5784 if (xmit->resending)
5785 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5787 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5790 clock_GetTime(&now);
5792 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5796 /* Set the packet flags and schedule the resend events */
5797 /* Only request an ack for the last packet in the list */
5798 for (i = 0; i < xmit->len; i++) {
5799 struct rx_packet *packet = xmit->list[i];
5801 /* Record the time sent */
5802 packet->timeSent = now;
5803 packet->flags |= RX_PKTFLAG_SENT;
5805 /* Ask for an ack on retransmitted packets, on every other packet
5806 * if the peer doesn't support slow start. Ask for an ack on every
5807 * packet until the congestion window reaches the ack rate. */
5808 if (packet->header.serial) {
5811 packet->firstSent = now;
5812 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5813 || (!(call->flags & RX_CALL_SLOW_START_OK)
5814 && (packet->header.seq & 1)))) {
5819 /* Tag this packet as not being the last in this group,
5820 * for the receiver's benefit */
5821 if (i < xmit->len - 1 || moreFlag) {
5822 packet->header.flags |= RX_MORE_PACKETS;
5827 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5830 /* Since we're about to send a data packet to the peer, it's
5831 * safe to nuke any scheduled end-of-packets ack */
5832 rxi_CancelDelayedAckEvent(call);
5834 MUTEX_EXIT(&call->lock);
5835 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5836 if (xmit->len > 1) {
5837 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5839 rxi_SendPacket(call, conn, xmit->list[0], istack);
5841 MUTEX_ENTER(&call->lock);
5842 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5844 /* Tell the RTO calculation engine that we have sent a packet, and
5845 * if it was the last one */
5846 rxi_rto_packet_sent(call, lastPacket, istack);
5848 /* Update last send time for this call (for keep-alive
5849 * processing), and for the connection (so that we can discover
5850 * idle connections) */
5851 conn->lastSendTime = call->lastSendTime = clock_Sec();
5852 /* Let a set of retransmits trigger an idle timeout */
5853 if (!xmit->resending)
5854 call->lastSendData = call->lastSendTime;
5857 /* When sending packets we need to follow these rules:
5858 * 1. Never send more than maxDgramPackets in a jumbogram.
5859 * 2. Never send a packet with more than two iovecs in a jumbogram.
5860 * 3. Never send a retransmitted packet in a jumbogram.
5861 * 4. Never send more than cwind/4 packets in a jumbogram
5862 * We always keep the last list we should have sent so we
5863 * can set the RX_MORE_PACKETS flags correctly.
5867 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5872 struct xmitlist working;
5873 struct xmitlist last;
5875 struct rx_peer *peer = call->conn->peer;
5876 int morePackets = 0;
5878 memset(&last, 0, sizeof(struct xmitlist));
5879 working.list = &list[0];
5881 working.resending = 0;
5883 recovery = call->flags & RX_CALL_FAST_RECOVER;
5885 for (i = 0; i < len; i++) {
5886 /* Does the current packet force us to flush the current list? */
5888 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5889 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5891 /* This sends the 'last' list and then rolls the current working
5892 * set into the 'last' one, and resets the working set */
5895 rxi_SendList(call, &last, istack, 1);
5896 /* If the call enters an error state stop sending, or if
5897 * we entered congestion recovery mode, stop sending */
5899 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5904 working.resending = 0;
5905 working.list = &list[i];
5907 /* Add the current packet to the list if it hasn't been acked.
5908 * Otherwise adjust the list pointer to skip the current packet. */
5909 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5912 if (list[i]->header.serial)
5913 working.resending = 1;
5915 /* Do we need to flush the list? */
5916 if (working.len >= (int)peer->maxDgramPackets
5917 || working.len >= (int)call->nDgramPackets
5918 || working.len >= (int)call->cwind
5919 || list[i]->header.serial
5920 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5922 rxi_SendList(call, &last, istack, 1);
5923 /* If the call enters an error state stop sending, or if
5924 * we entered congestion recovery mode, stop sending */
5926 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5931 working.resending = 0;
5932 working.list = &list[i + 1];
5935 if (working.len != 0) {
5936 osi_Panic("rxi_SendList error");
5938 working.list = &list[i + 1];
5942 /* Send the whole list when the call is in receive mode, when
5943 * the call is in eof mode, when we are in fast recovery mode,
5944 * and when we have the last packet */
5945 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5946 * the listener or event threads
5948 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5949 || (call->flags & RX_CALL_FLUSH)
5950 || (call->flags & RX_CALL_FAST_RECOVER)) {
5951 /* Check for the case where the current list contains
5952 * an acked packet. Since we always send retransmissions
5953 * in a separate packet, we only need to check the first
5954 * packet in the list */
5955 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5959 rxi_SendList(call, &last, istack, morePackets);
5960 /* If the call enters an error state stop sending, or if
5961 * we entered congestion recovery mode, stop sending */
5963 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5967 rxi_SendList(call, &working, istack, 0);
5969 } else if (last.len > 0) {
5970 rxi_SendList(call, &last, istack, 0);
5971 /* Packets which are in 'working' are not sent by this call */
5976 * Check if the peer for the given call is known to be dead
5978 * If the call's peer appears dead (it has encountered fatal network errors
5979 * since the call started) the call is killed with RX_CALL_DEAD if the call
5980 * is active. Otherwise, we do nothing.
5982 * @param[in] call The call to check
5985 * @retval 0 The call is fine, and we haven't done anything to the call
5986 * @retval nonzero The call's peer appears dead, and the call has been
5987 * terminated if it was active
5989 * @pre call->lock must be locked
5992 rxi_CheckPeerDead(struct rx_call *call)
5994 #ifdef AFS_RXERRQ_ENV
5997 if (call->state == RX_STATE_DALLY) {
6001 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
6002 if (call->neterr_gen < peererrs) {
6003 /* we have received network errors since this call started; kill
6005 if (call->state == RX_STATE_ACTIVE) {
6006 rxi_CallError(call, RX_CALL_DEAD);
6010 if (call->neterr_gen > peererrs) {
6011 /* someone has reset the number of peer errors; set the call error gen
6012 * so we can detect if more errors are encountered */
6013 call->neterr_gen = peererrs;
6020 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
6022 struct rx_call *call = arg0;
6023 struct rx_peer *peer;
6024 struct opr_queue *cursor;
6025 struct clock maxTimeout = { 60, 0 };
6027 MUTEX_ENTER(&call->lock);
6029 peer = call->conn->peer;
6031 /* Make sure that the event pointer is removed from the call
6032 * structure, since there is no longer a per-call retransmission
6034 if (event == call->resendEvent) {
6035 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6036 rxevent_Put(&call->resendEvent);
6039 rxi_CheckPeerDead(call);
6041 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6042 rxi_CheckBusy(call);
6045 if (opr_queue_IsEmpty(&call->tq)) {
6046 /* Nothing to do. This means that we've been raced, and that an
6047 * ACK has come in between when we were triggered, and when we
6048 * actually got to run. */
6052 /* We're in loss recovery */
6053 call->flags |= RX_CALL_FAST_RECOVER;
6055 /* Mark all of the pending packets in the queue as being lost */
6056 for (opr_queue_Scan(&call->tq, cursor)) {
6057 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6058 if (!(p->flags & RX_PKTFLAG_ACKED))
6059 p->flags &= ~RX_PKTFLAG_SENT;
6062 /* We're resending, so we double the timeout of the call. This will be
6063 * dropped back down by the first successful ACK that we receive.
6065 * We apply a maximum value here of 60 seconds
6067 clock_Add(&call->rto, &call->rto);
6068 if (clock_Gt(&call->rto, &maxTimeout))
6069 call->rto = maxTimeout;
6071 /* Packet loss is most likely due to congestion, so drop our window size
6072 * and start again from the beginning */
6073 if (peer->maxDgramPackets >1) {
6074 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6075 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6077 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6078 call->nDgramPackets = 1;
6080 call->nextCwind = 1;
6083 MUTEX_ENTER(&peer->peer_lock);
6084 peer->MTU = call->MTU;
6085 peer->cwind = call->cwind;
6086 peer->nDgramPackets = 1;
6088 call->congestSeq = peer->congestSeq;
6089 MUTEX_EXIT(&peer->peer_lock);
6091 rxi_Start(call, istack);
6094 MUTEX_EXIT(&call->lock);
6097 /* This routine is called when new packets are readied for
6098 * transmission and when retransmission may be necessary, or when the
6099 * transmission window or burst count are favourable. This should be
6100 * better optimized for new packets, the usual case, now that we've
6101 * got rid of queues of send packets. XXXXXXXXXXX */
6103 rxi_Start(struct rx_call *call, int istack)
6105 struct opr_queue *cursor;
6106 #ifdef RX_ENABLE_LOCKS
6107 struct opr_queue *store;
6113 #ifdef RX_ENABLE_LOCKS
6114 if (rx_stats_active)
6115 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6120 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6121 /* Send (or resend) any packets that need it, subject to
6122 * window restrictions and congestion burst control
6123 * restrictions. Ask for an ack on the last packet sent in
6124 * this burst. For now, we're relying upon the window being
6125 * considerably bigger than the largest number of packets that
6126 * are typically sent at once by one initial call to
6127 * rxi_Start. This is probably bogus (perhaps we should ask
6128 * for an ack when we're half way through the current
6129 * window?). Also, for non file transfer applications, this
6130 * may end up asking for an ack for every packet. Bogus. XXXX
6133 * But check whether we're here recursively, and let the other guy
6136 #ifdef RX_ENABLE_LOCKS
6137 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6138 call->flags |= RX_CALL_TQ_BUSY;
6140 #endif /* RX_ENABLE_LOCKS */
6142 #ifdef RX_ENABLE_LOCKS
6143 call->flags &= ~RX_CALL_NEED_START;
6144 #endif /* RX_ENABLE_LOCKS */
6146 maxXmitPackets = MIN(call->twind, call->cwind);
6147 for (opr_queue_Scan(&call->tq, cursor)) {
6149 = opr_queue_Entry(cursor, struct rx_packet, entry);
6151 if (p->flags & RX_PKTFLAG_ACKED) {
6152 /* Since we may block, don't trust this */
6153 if (rx_stats_active)
6154 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6155 continue; /* Ignore this packet if it has been acknowledged */
6158 /* Turn off all flags except these ones, which are the same
6159 * on each transmission */
6160 p->header.flags &= RX_PRESET_FLAGS;
6162 if (p->header.seq >=
6163 call->tfirst + MIN((int)call->twind,
6164 (int)(call->nSoftAcked +
6166 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6167 /* Note: if we're waiting for more window space, we can
6168 * still send retransmits; hence we don't return here, but
6169 * break out to schedule a retransmit event */
6170 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6171 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6176 /* Transmit the packet if it needs to be sent. */
6177 if (!(p->flags & RX_PKTFLAG_SENT)) {
6178 if (nXmitPackets == maxXmitPackets) {
6179 rxi_SendXmitList(call, call->xmitList,
6180 nXmitPackets, istack);
6183 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6184 *(call->callNumber), p));
6185 call->xmitList[nXmitPackets++] = p;
6187 } /* end of the queue_Scan */
6189 /* xmitList now hold pointers to all of the packets that are
6190 * ready to send. Now we loop to send the packets */
6191 if (nXmitPackets > 0) {
6192 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6196 #ifdef RX_ENABLE_LOCKS
6198 /* We went into the error state while sending packets. Now is
6199 * the time to reset the call. This will also inform the using
6200 * process that the call is in an error state.
6202 if (rx_stats_active)
6203 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6204 call->flags &= ~RX_CALL_TQ_BUSY;
6205 rxi_WakeUpTransmitQueue(call);
6206 rxi_CallError(call, call->error);
6210 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6212 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6213 /* Some packets have received acks. If they all have, we can clear
6214 * the transmit queue.
6217 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6219 = opr_queue_Entry(cursor, struct rx_packet, entry);
6221 if (p->header.seq < call->tfirst
6222 && (p->flags & RX_PKTFLAG_ACKED)) {
6223 opr_queue_Remove(&p->entry);
6224 #ifdef RX_TRACK_PACKETS
6225 p->flags &= ~RX_PKTFLAG_TQ;
6227 #ifdef RXDEBUG_PACKET
6235 call->flags |= RX_CALL_TQ_CLEARME;
6237 if (call->flags & RX_CALL_TQ_CLEARME)
6238 rxi_ClearTransmitQueue(call, 1);
6239 } while (call->flags & RX_CALL_NEED_START);
6241 * TQ references no longer protected by this flag; they must remain
6242 * protected by the call lock.
6244 call->flags &= ~RX_CALL_TQ_BUSY;
6245 rxi_WakeUpTransmitQueue(call);
6247 call->flags |= RX_CALL_NEED_START;
6249 #endif /* RX_ENABLE_LOCKS */
6251 rxi_rto_cancel(call);
6255 /* Also adjusts the keep alive parameters for the call, to reflect
6256 * that we have just sent a packet (so keep alives aren't sent
6259 rxi_Send(struct rx_call *call, struct rx_packet *p,
6262 struct rx_connection *conn = call->conn;
6264 /* Stamp each packet with the user supplied status */
6265 p->header.userStatus = call->localStatus;
6267 /* Allow the security object controlling this call's security to
6268 * make any last-minute changes to the packet */
6269 RXS_SendPacket(conn->securityObject, call, p);
6271 /* Since we're about to send SOME sort of packet to the peer, it's
6272 * safe to nuke any scheduled end-of-packets ack */
6273 rxi_CancelDelayedAckEvent(call);
6275 /* Actually send the packet, filling in more connection-specific fields */
6276 MUTEX_EXIT(&call->lock);
6277 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6278 rxi_SendPacket(call, conn, p, istack);
6279 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6280 MUTEX_ENTER(&call->lock);
6282 /* Update last send time for this call (for keep-alive
6283 * processing), and for the connection (so that we can discover
6284 * idle connections) */
6285 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6286 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6287 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6289 conn->lastSendTime = call->lastSendTime = clock_Sec();
6290 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6291 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6292 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6293 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6294 RX_ACK_PING_RESPONSE)))
6295 call->lastSendData = call->lastSendTime;
6299 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6300 * that things are fine. Also called periodically to guarantee that nothing
6301 * falls through the cracks (e.g. (error + dally) connections have keepalive
6302 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6304 * haveCTLock Set if calling from rxi_ReapConnections
6307 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6309 struct rx_connection *conn = call->conn;
6311 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6312 afs_uint32 fudgeFactor;
6315 int idle_timeout = 0;
6316 afs_int32 clock_diff = 0;
6318 if (rxi_CheckPeerDead(call)) {
6324 /* Large swings in the clock can have a significant impact on
6325 * the performance of RX call processing. Forward clock shifts
6326 * will result in premature event triggering or timeouts.
6327 * Backward shifts can result in calls not completing until
6328 * the clock catches up with the original start clock value.
6330 * If a backward clock shift of more than five minutes is noticed,
6331 * just fail the call.
6333 if (now < call->lastSendTime)
6334 clock_diff = call->lastSendTime - now;
6335 if (now < call->startWait)
6336 clock_diff = MAX(clock_diff, call->startWait - now);
6337 if (now < call->lastReceiveTime)
6338 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6339 if (clock_diff > 5 * 60)
6341 if (call->state == RX_STATE_ACTIVE)
6342 rxi_CallError(call, RX_CALL_TIMEOUT);
6346 #ifdef RX_ENABLE_LOCKS
6347 if (call->flags & RX_CALL_TQ_BUSY) {
6348 /* Call is active and will be reset by rxi_Start if it's
6349 * in an error state.
6354 /* RTT + 8*MDEV, rounded up to the next second. */
6355 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6356 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6358 deadTime = conn->secondsUntilDead + fudgeFactor;
6359 /* These are computed to the second (+- 1 second). But that's
6360 * good enough for these values, which should be a significant
6361 * number of seconds. */
6362 if (now > (call->lastReceiveTime + deadTime)) {
6363 if (call->state == RX_STATE_ACTIVE) {
6364 #ifdef AFS_ADAPT_PMTU
6365 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6367 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6368 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6369 ip_stack_t *ipst = ns->netstack_ip;
6371 ire = ire_cache_lookup(conn->peer->host
6372 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6374 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6376 # if defined(GLOBAL_NETSTACKID)
6383 if (ire && ire->ire_max_frag > 0)
6384 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6386 # if defined(GLOBAL_NETSTACKID)
6390 #endif /* AFS_ADAPT_PMTU */
6391 cerror = RX_CALL_DEAD;
6394 #ifdef RX_ENABLE_LOCKS
6395 /* Cancel pending events */
6396 rxi_CancelDelayedAckEvent(call);
6397 rxi_rto_cancel(call);
6398 rxi_CancelKeepAliveEvent(call);
6399 rxi_CancelGrowMTUEvent(call);
6400 MUTEX_ENTER(&rx_refcnt_mutex);
6401 /* if rxi_FreeCall returns 1 it has freed the call */
6402 if (call->refCount == 0 &&
6403 rxi_FreeCall(call, haveCTLock))
6405 MUTEX_EXIT(&rx_refcnt_mutex);
6408 MUTEX_EXIT(&rx_refcnt_mutex);
6410 #else /* RX_ENABLE_LOCKS */
6411 rxi_FreeCall(call, 0);
6413 #endif /* RX_ENABLE_LOCKS */
6415 /* Non-active calls are destroyed if they are not responding
6416 * to pings; active calls are simply flagged in error, so the
6417 * attached process can die reasonably gracefully. */
6420 if (conn->idleDeadDetection) {
6421 if (conn->idleDeadTime) {
6422 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6426 /* see if we have a non-activity timeout */
6427 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6428 (call->flags & RX_CALL_READER_WAIT)) {
6429 if (call->state == RX_STATE_ACTIVE) {
6430 cerror = RX_CALL_TIMEOUT;
6435 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6436 if (call->state == RX_STATE_ACTIVE) {
6437 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6445 if (conn->hardDeadTime) {
6446 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6449 /* see if we have a hard timeout */
6451 && (now > (hardDeadTime + call->startTime.sec))) {
6452 if (call->state == RX_STATE_ACTIVE)
6453 rxi_CallError(call, RX_CALL_TIMEOUT);
6458 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6459 call->lastReceiveTime) {
6460 int oldMTU = conn->peer->ifMTU;
6462 /* if we thought we could send more, perhaps things got worse */
6463 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6464 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6465 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6466 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6468 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6470 /* minimum capped in SetPeerMtu */
6471 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6474 conn->lastPacketSize = 0;
6476 /* needed so ResetCall doesn't clobber us. */
6477 call->MTU = conn->peer->ifMTU;
6479 /* if we never succeeded, let the error pass out as-is */
6480 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6481 cerror = conn->msgsizeRetryErr;
6484 rxi_CallError(call, cerror);
6489 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6490 void *dummy, int dummy2)
6492 struct rx_connection *conn = arg1;
6493 struct rx_header theader;
6494 char tbuffer[1 + sizeof(struct rx_header)];
6495 struct sockaddr_in taddr;
6498 struct iovec tmpiov[2];
6501 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6504 tp = &tbuffer[sizeof(struct rx_header)];
6505 taddr.sin_family = AF_INET;
6506 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6507 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6508 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6509 taddr.sin_len = sizeof(struct sockaddr_in);
6511 memset(&theader, 0, sizeof(theader));
6512 theader.epoch = htonl(999);
6514 theader.callNumber = 0;
6517 theader.type = RX_PACKET_TYPE_VERSION;
6518 theader.flags = RX_LAST_PACKET;
6519 theader.serviceId = 0;
6521 memcpy(tbuffer, &theader, sizeof(theader));
6522 memcpy(tp, &a, sizeof(a));
6523 tmpiov[0].iov_base = tbuffer;
6524 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6526 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6528 MUTEX_ENTER(&conn->conn_data_lock);
6529 MUTEX_ENTER(&rx_refcnt_mutex);
6530 /* Only reschedule ourselves if the connection would not be destroyed */
6531 if (conn->refCount <= 1) {
6532 rxevent_Put(&conn->natKeepAliveEvent);
6533 MUTEX_EXIT(&rx_refcnt_mutex);
6534 MUTEX_EXIT(&conn->conn_data_lock);
6535 rx_DestroyConnection(conn); /* drop the reference for this */
6537 conn->refCount--; /* drop the reference for this */
6538 MUTEX_EXIT(&rx_refcnt_mutex);
6539 rxevent_Put(&conn->natKeepAliveEvent);
6540 rxi_ScheduleNatKeepAliveEvent(conn);
6541 MUTEX_EXIT(&conn->conn_data_lock);
6546 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6548 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6549 struct clock when, now;
6550 clock_GetTime(&now);
6552 when.sec += conn->secondsUntilNatPing;
6553 MUTEX_ENTER(&rx_refcnt_mutex);
6554 conn->refCount++; /* hold a reference for this */
6555 MUTEX_EXIT(&rx_refcnt_mutex);
6556 conn->natKeepAliveEvent =
6557 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6562 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6564 MUTEX_ENTER(&conn->conn_data_lock);
6565 conn->secondsUntilNatPing = seconds;
6567 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6568 rxi_ScheduleNatKeepAliveEvent(conn);
6570 conn->flags |= RX_CONN_NAT_PING;
6572 MUTEX_EXIT(&conn->conn_data_lock);
6575 /* When a call is in progress, this routine is called occasionally to
6576 * make sure that some traffic has arrived (or been sent to) the peer.
6577 * If nothing has arrived in a reasonable amount of time, the call is
6578 * declared dead; if nothing has been sent for a while, we send a
6579 * keep-alive packet (if we're actually trying to keep the call alive)
6582 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6585 struct rx_call *call = arg1;
6586 struct rx_connection *conn;
6589 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6590 MUTEX_ENTER(&call->lock);
6592 if (event == call->keepAliveEvent)
6593 rxevent_Put(&call->keepAliveEvent);
6597 if (rxi_CheckCall(call, 0)) {
6598 MUTEX_EXIT(&call->lock);
6602 /* Don't try to keep alive dallying calls */
6603 if (call->state == RX_STATE_DALLY) {
6604 MUTEX_EXIT(&call->lock);
6609 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6610 /* Don't try to send keepalives if there is unacknowledged data */
6611 /* the rexmit code should be good enough, this little hack
6612 * doesn't quite work XXX */
6613 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6615 rxi_ScheduleKeepAliveEvent(call);
6616 MUTEX_EXIT(&call->lock);
6619 /* Does what's on the nameplate. */
6621 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6623 struct rx_call *call = arg1;
6624 struct rx_connection *conn;
6626 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6627 MUTEX_ENTER(&call->lock);
6629 if (event == call->growMTUEvent)
6630 rxevent_Put(&call->growMTUEvent);
6632 if (rxi_CheckCall(call, 0)) {
6633 MUTEX_EXIT(&call->lock);
6637 /* Don't bother with dallying calls */
6638 if (call->state == RX_STATE_DALLY) {
6639 MUTEX_EXIT(&call->lock);
6646 * keep being scheduled, just don't do anything if we're at peak,
6647 * or we're not set up to be properly handled (idle timeout required)
6649 if ((conn->peer->maxPacketSize != 0) &&
6650 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6651 conn->idleDeadDetection)
6652 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6653 rxi_ScheduleGrowMTUEvent(call, 0);
6654 MUTEX_EXIT(&call->lock);
6658 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6660 if (!call->keepAliveEvent) {
6661 struct clock when, now;
6662 clock_GetTime(&now);
6664 when.sec += call->conn->secondsUntilPing;
6665 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6666 call->keepAliveEvent =
6667 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6672 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6673 if (call->keepAliveEvent) {
6674 rxevent_Cancel(&call->keepAliveEvent);
6675 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6680 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6682 if (!call->growMTUEvent) {
6683 struct clock when, now;
6685 clock_GetTime(&now);
6688 if (call->conn->secondsUntilPing)
6689 secs = (6*call->conn->secondsUntilPing)-1;
6691 if (call->conn->secondsUntilDead)
6692 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6696 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6697 call->growMTUEvent =
6698 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6703 rxi_CancelGrowMTUEvent(struct rx_call *call)
6705 if (call->growMTUEvent) {
6706 rxevent_Cancel(&call->growMTUEvent);
6707 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6712 rxi_KeepAliveOn(struct rx_call *call)
6714 /* Pretend last packet received was received now--i.e. if another
6715 * packet isn't received within the keep alive time, then the call
6716 * will die; Initialize last send time to the current time--even
6717 * if a packet hasn't been sent yet. This will guarantee that a
6718 * keep-alive is sent within the ping time */
6719 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6720 rxi_ScheduleKeepAliveEvent(call);
6724 rx_KeepAliveOff(struct rx_call *call)
6726 MUTEX_ENTER(&call->lock);
6727 rxi_CancelKeepAliveEvent(call);
6728 MUTEX_EXIT(&call->lock);
6732 rx_KeepAliveOn(struct rx_call *call)
6734 MUTEX_ENTER(&call->lock);
6735 rxi_KeepAliveOn(call);
6736 MUTEX_EXIT(&call->lock);
6740 rxi_GrowMTUOn(struct rx_call *call)
6742 struct rx_connection *conn = call->conn;
6743 MUTEX_ENTER(&conn->conn_data_lock);
6744 conn->lastPingSizeSer = conn->lastPingSize = 0;
6745 MUTEX_EXIT(&conn->conn_data_lock);
6746 rxi_ScheduleGrowMTUEvent(call, 1);
6749 /* This routine is called to send connection abort messages
6750 * that have been delayed to throttle looping clients. */
6752 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6755 struct rx_connection *conn = arg1;
6758 struct rx_packet *packet;
6760 MUTEX_ENTER(&conn->conn_data_lock);
6761 rxevent_Put(&conn->delayedAbortEvent);
6762 error = htonl(conn->error);
6764 MUTEX_EXIT(&conn->conn_data_lock);
6765 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6768 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6769 RX_PACKET_TYPE_ABORT, (char *)&error,
6771 rxi_FreePacket(packet);
6775 /* This routine is called to send call abort messages
6776 * that have been delayed to throttle looping clients. */
6778 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6781 struct rx_call *call = arg1;
6784 struct rx_packet *packet;
6786 MUTEX_ENTER(&call->lock);
6787 rxevent_Put(&call->delayedAbortEvent);
6788 error = htonl(call->error);
6790 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6793 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6794 (char *)&error, sizeof(error), 0);
6795 rxi_FreePacket(packet);
6797 MUTEX_EXIT(&call->lock);
6798 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6801 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6802 * seconds) to ask the client to authenticate itself. The routine
6803 * issues a challenge to the client, which is obtained from the
6804 * security object associated with the connection */
6806 rxi_ChallengeEvent(struct rxevent *event,
6807 void *arg0, void *arg1, int tries)
6809 struct rx_connection *conn = arg0;
6812 rxevent_Put(&conn->challengeEvent);
6814 /* If there are no active calls it is not worth re-issuing the
6815 * challenge. If the client issues another call on this connection
6816 * the challenge can be requested at that time.
6818 if (!rxi_HasActiveCalls(conn))
6821 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6822 struct rx_packet *packet;
6823 struct clock when, now;
6826 /* We've failed to authenticate for too long.
6827 * Reset any calls waiting for authentication;
6828 * they are all in RX_STATE_PRECALL.
6832 MUTEX_ENTER(&conn->conn_call_lock);
6833 for (i = 0; i < RX_MAXCALLS; i++) {
6834 struct rx_call *call = conn->call[i];
6836 MUTEX_ENTER(&call->lock);
6837 if (call->state == RX_STATE_PRECALL) {
6838 rxi_CallError(call, RX_CALL_DEAD);
6839 rxi_SendCallAbort(call, NULL, 0, 0);
6841 MUTEX_EXIT(&call->lock);
6844 MUTEX_EXIT(&conn->conn_call_lock);
6848 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6850 /* If there's no packet available, do this later. */
6851 RXS_GetChallenge(conn->securityObject, conn, packet);
6852 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6853 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6854 rxi_FreePacket(packet);
6856 clock_GetTime(&now);
6858 when.sec += RX_CHALLENGE_TIMEOUT;
6859 conn->challengeEvent =
6860 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6865 /* Call this routine to start requesting the client to authenticate
6866 * itself. This will continue until authentication is established,
6867 * the call times out, or an invalid response is returned. The
6868 * security object associated with the connection is asked to create
6869 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6870 * defined earlier. */
6872 rxi_ChallengeOn(struct rx_connection *conn)
6874 if (!conn->challengeEvent) {
6875 RXS_CreateChallenge(conn->securityObject, conn);
6876 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6881 /* rxi_ComputeRoundTripTime is called with peer locked. */
6882 /* peer may be null */
6884 rxi_ComputeRoundTripTime(struct rx_packet *p,
6885 struct rx_ackPacket *ack,
6886 struct rx_call *call,
6887 struct rx_peer *peer,
6890 struct clock thisRtt, *sentp;
6894 /* If the ACK is delayed, then do nothing */
6895 if (ack->reason == RX_ACK_DELAY)
6898 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6899 * their RTT multiple times, so only include the RTT of the last packet
6901 if (p->flags & RX_JUMBO_PACKET)
6904 /* Use the serial number to determine which transmission the ACK is for,
6905 * and set the sent time to match this. If we have no serial number, then
6906 * only use the ACK for RTT calculations if the packet has not been
6910 serial = ntohl(ack->serial);
6912 if (serial == p->header.serial) {
6913 sentp = &p->timeSent;
6914 } else if (serial == p->firstSerial) {
6915 sentp = &p->firstSent;
6916 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6917 sentp = &p->firstSent;
6921 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6922 sentp = &p->firstSent;
6929 if (clock_Lt(&thisRtt, sentp))
6930 return; /* somebody set the clock back, don't count this time. */
6932 clock_Sub(&thisRtt, sentp);
6933 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6934 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6936 if (clock_IsZero(&thisRtt)) {
6938 * The actual round trip time is shorter than the
6939 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6940 * Since we can't tell which at the moment we will assume 1ms.
6942 thisRtt.usec = 1000;
6945 if (rx_stats_active) {
6946 MUTEX_ENTER(&rx_stats_mutex);
6947 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6948 rx_stats.minRtt = thisRtt;
6949 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6950 if (thisRtt.sec > 60) {
6951 MUTEX_EXIT(&rx_stats_mutex);
6952 return; /* somebody set the clock ahead */
6954 rx_stats.maxRtt = thisRtt;
6956 clock_Add(&rx_stats.totalRtt, &thisRtt);
6957 rx_atomic_inc(&rx_stats.nRttSamples);
6958 MUTEX_EXIT(&rx_stats_mutex);
6961 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6963 /* Apply VanJacobson round-trip estimations */
6968 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6969 * srtt is stored as fixed point with 3 bits after the binary
6970 * point (i.e., scaled by 8). The following magic is
6971 * equivalent to the smoothing algorithm in rfc793 with an
6972 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6973 * srtt'*8 = rtt + srtt*7
6974 * srtt'*8 = srtt*8 + rtt - srtt
6975 * srtt' = srtt + rtt/8 - srtt/8
6976 * srtt' = srtt + (rtt - srtt)/8
6979 delta = _8THMSEC(&thisRtt) - call->rtt;
6980 call->rtt += (delta >> 3);
6983 * We accumulate a smoothed rtt variance (actually, a smoothed
6984 * mean difference), then set the retransmit timer to smoothed
6985 * rtt + 4 times the smoothed variance (was 2x in van's original
6986 * paper, but 4x works better for me, and apparently for him as
6988 * rttvar is stored as
6989 * fixed point with 2 bits after the binary point (scaled by
6990 * 4). The following is equivalent to rfc793 smoothing with
6991 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6992 * rttvar'*4 = rttvar*3 + |delta|
6993 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6994 * rttvar' = rttvar + |delta|/4 - rttvar/4
6995 * rttvar' = rttvar + (|delta| - rttvar)/4
6996 * This replaces rfc793's wired-in beta.
6997 * dev*4 = dev*4 + (|actual - expected| - dev)
7003 delta -= (call->rtt_dev << 1);
7004 call->rtt_dev += (delta >> 3);
7006 /* I don't have a stored RTT so I start with this value. Since I'm
7007 * probably just starting a call, and will be pushing more data down
7008 * this, I expect congestion to increase rapidly. So I fudge a
7009 * little, and I set deviance to half the rtt. In practice,
7010 * deviance tends to approach something a little less than
7011 * half the smoothed rtt. */
7012 call->rtt = _8THMSEC(&thisRtt) + 8;
7013 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
7015 /* the smoothed RTT time is RTT + 4*MDEV
7017 * We allow a user specified minimum to be set for this, to allow clamping
7018 * at a minimum value in the same way as TCP. In addition, we have to allow
7019 * for the possibility that this packet is answered by a delayed ACK, so we
7020 * add on a fixed 200ms to account for that timer expiring.
7023 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
7024 rx_minPeerTimeout) + 200;
7025 clock_Zero(&call->rto);
7026 clock_Addmsec(&call->rto, rtt_timeout);
7028 /* Update the peer, so any new calls start with our values */
7029 peer->rtt_dev = call->rtt_dev;
7030 peer->rtt = call->rtt;
7032 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
7033 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
7037 /* Find all server connections that have not been active for a long time, and
7040 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7043 struct clock now, when;
7044 struct rxevent *event;
7045 clock_GetTime(&now);
7047 /* Find server connection structures that haven't been used for
7048 * greater than rx_idleConnectionTime */
7050 struct rx_connection **conn_ptr, **conn_end;
7051 int i, havecalls = 0;
7052 MUTEX_ENTER(&rx_connHashTable_lock);
7053 for (conn_ptr = &rx_connHashTable[0], conn_end =
7054 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7056 struct rx_connection *conn, *next;
7057 struct rx_call *call;
7061 for (conn = *conn_ptr; conn; conn = next) {
7062 /* XXX -- Shouldn't the connection be locked? */
7065 for (i = 0; i < RX_MAXCALLS; i++) {
7066 call = conn->call[i];
7070 code = MUTEX_TRYENTER(&call->lock);
7073 result = rxi_CheckCall(call, 1);
7074 MUTEX_EXIT(&call->lock);
7076 /* If CheckCall freed the call, it might
7077 * have destroyed the connection as well,
7078 * which screws up the linked lists.
7084 if (conn->type == RX_SERVER_CONNECTION) {
7085 /* This only actually destroys the connection if
7086 * there are no outstanding calls */
7087 MUTEX_ENTER(&conn->conn_data_lock);
7088 MUTEX_ENTER(&rx_refcnt_mutex);
7089 if (!havecalls && !conn->refCount
7090 && ((conn->lastSendTime + rx_idleConnectionTime) <
7092 conn->refCount++; /* it will be decr in rx_DestroyConn */
7093 MUTEX_EXIT(&rx_refcnt_mutex);
7094 MUTEX_EXIT(&conn->conn_data_lock);
7095 #ifdef RX_ENABLE_LOCKS
7096 rxi_DestroyConnectionNoLock(conn);
7097 #else /* RX_ENABLE_LOCKS */
7098 rxi_DestroyConnection(conn);
7099 #endif /* RX_ENABLE_LOCKS */
7101 #ifdef RX_ENABLE_LOCKS
7103 MUTEX_EXIT(&rx_refcnt_mutex);
7104 MUTEX_EXIT(&conn->conn_data_lock);
7106 #endif /* RX_ENABLE_LOCKS */
7110 #ifdef RX_ENABLE_LOCKS
7111 while (rx_connCleanup_list) {
7112 struct rx_connection *conn;
7113 conn = rx_connCleanup_list;
7114 rx_connCleanup_list = rx_connCleanup_list->next;
7115 MUTEX_EXIT(&rx_connHashTable_lock);
7116 rxi_CleanupConnection(conn);
7117 MUTEX_ENTER(&rx_connHashTable_lock);
7119 MUTEX_EXIT(&rx_connHashTable_lock);
7120 #endif /* RX_ENABLE_LOCKS */
7123 /* Find any peer structures that haven't been used (haven't had an
7124 * associated connection) for greater than rx_idlePeerTime */
7126 struct rx_peer **peer_ptr, **peer_end;
7130 * Why do we need to hold the rx_peerHashTable_lock across
7131 * the incrementing of peer_ptr since the rx_peerHashTable
7132 * array is not changing? We don't.
7134 * By dropping the lock periodically we can permit other
7135 * activities to be performed while a rxi_ReapConnections
7136 * call is in progress. The goal of reap connections
7137 * is to clean up quickly without causing large amounts
7138 * of contention. Therefore, it is important that global
7139 * mutexes not be held for extended periods of time.
7141 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7142 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7144 struct rx_peer *peer, *next, *prev;
7146 MUTEX_ENTER(&rx_peerHashTable_lock);
7147 for (prev = peer = *peer_ptr; peer; peer = next) {
7149 code = MUTEX_TRYENTER(&peer->peer_lock);
7150 if ((code) && (peer->refCount == 0)
7151 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7152 struct opr_queue *cursor, *store;
7156 * now know that this peer object is one to be
7157 * removed from the hash table. Once it is removed
7158 * it can't be referenced by other threads.
7159 * Lets remove it first and decrement the struct
7160 * nPeerStructs count.
7162 if (peer == *peer_ptr) {
7168 if (rx_stats_active)
7169 rx_atomic_dec(&rx_stats.nPeerStructs);
7172 * Now if we hold references on 'prev' and 'next'
7173 * we can safely drop the rx_peerHashTable_lock
7174 * while we destroy this 'peer' object.
7180 MUTEX_EXIT(&rx_peerHashTable_lock);
7182 MUTEX_EXIT(&peer->peer_lock);
7183 MUTEX_DESTROY(&peer->peer_lock);
7185 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7186 unsigned int num_funcs;
7187 struct rx_interface_stat *rpc_stat
7188 = opr_queue_Entry(cursor, struct rx_interface_stat,
7193 opr_queue_Remove(&rpc_stat->entry);
7194 opr_queue_Remove(&rpc_stat->entryPeers);
7196 num_funcs = rpc_stat->stats[0].func_total;
7198 sizeof(rx_interface_stat_t) +
7199 rpc_stat->stats[0].func_total *
7200 sizeof(rx_function_entry_v1_t);
7202 rxi_Free(rpc_stat, space);
7204 MUTEX_ENTER(&rx_rpc_stats);
7205 rxi_rpc_peer_stat_cnt -= num_funcs;
7206 MUTEX_EXIT(&rx_rpc_stats);
7211 * Regain the rx_peerHashTable_lock and
7212 * decrement the reference count on 'prev'
7215 MUTEX_ENTER(&rx_peerHashTable_lock);
7222 MUTEX_EXIT(&peer->peer_lock);
7227 MUTEX_EXIT(&rx_peerHashTable_lock);
7231 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7232 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7233 * GC, just below. Really, we shouldn't have to keep moving packets from
7234 * one place to another, but instead ought to always know if we can
7235 * afford to hold onto a packet in its particular use. */
7236 MUTEX_ENTER(&rx_freePktQ_lock);
7237 if (rx_waitingForPackets) {
7238 rx_waitingForPackets = 0;
7239 #ifdef RX_ENABLE_LOCKS
7240 CV_BROADCAST(&rx_waitingForPackets_cv);
7242 osi_rxWakeup(&rx_waitingForPackets);
7245 MUTEX_EXIT(&rx_freePktQ_lock);
7248 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7249 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7250 rxevent_Put(&event);
7254 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7255 * rx.h is sort of strange this is better. This is called with a security
7256 * object before it is discarded. Each connection using a security object has
7257 * its own refcount to the object so it won't actually be freed until the last
7258 * connection is destroyed.
7260 * This is the only rxs module call. A hold could also be written but no one
7264 rxs_Release(struct rx_securityClass *aobj)
7266 return RXS_Close(aobj);
7274 #define TRACE_OPTION_RX_DEBUG 16
7282 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7283 0, KEY_QUERY_VALUE, &parmKey);
7284 if (code != ERROR_SUCCESS)
7287 dummyLen = sizeof(TraceOption);
7288 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7289 (BYTE *) &TraceOption, &dummyLen);
7290 if (code == ERROR_SUCCESS) {
7291 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7293 RegCloseKey (parmKey);
7294 #endif /* AFS_NT40_ENV */
7299 rx_DebugOnOff(int on)
7303 rxdebug_active = on;
7309 rx_StatsOnOff(int on)
7311 rx_stats_active = on;
7315 /* Don't call this debugging routine directly; use dpf */
7317 rxi_DebugPrint(char *format, ...)
7326 va_start(ap, format);
7328 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7331 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7333 OutputDebugString(msg);
7339 va_start(ap, format);
7341 clock_GetTime(&now);
7342 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7343 (unsigned int)now.usec);
7344 vfprintf(rx_Log, format, ap);
7352 * This function is used to process the rx_stats structure that is local
7353 * to a process as well as an rx_stats structure received from a remote
7354 * process (via rxdebug). Therefore, it needs to do minimal version
7358 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7359 afs_int32 freePackets, char version)
7363 if (size != sizeof(struct rx_statistics)) {
7365 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7366 size, sizeof(struct rx_statistics));
7369 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7372 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7373 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7374 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7375 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7376 s->specialPktAllocFailures);
7378 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7379 s->receivePktAllocFailures, s->sendPktAllocFailures,
7380 s->specialPktAllocFailures);
7384 " greedy %u, " "bogusReads %u (last from host %x), "
7385 "noPackets %u, " "noBuffers %u, " "selects %u, "
7386 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7387 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7388 s->selects, s->sendSelects);
7390 fprintf(file, " packets read: ");
7391 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7392 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7394 fprintf(file, "\n");
7397 " other read counters: data %u, " "ack %u, " "dup %u "
7398 "spurious %u " "dally %u\n", s->dataPacketsRead,
7399 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7400 s->ignorePacketDally);
7402 fprintf(file, " packets sent: ");
7403 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7404 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7406 fprintf(file, "\n");
7409 " other send counters: ack %u, " "data %u (not resends), "
7410 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7411 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7412 s->dataPacketsPushed, s->ignoreAckedPacket);
7415 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7416 s->netSendFailures, (int)s->fatalErrors);
7418 if (s->nRttSamples) {
7419 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7420 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7422 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7423 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7427 " %d server connections, " "%d client connections, "
7428 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7429 s->nServerConns, s->nClientConns, s->nPeerStructs,
7430 s->nCallStructs, s->nFreeCallStructs);
7432 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7433 fprintf(file, " %d clock updates\n", clock_nUpdates);
7437 /* for backward compatibility */
7439 rx_PrintStats(FILE * file)
7441 MUTEX_ENTER(&rx_stats_mutex);
7442 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7443 sizeof(rx_stats), rx_nFreePackets,
7445 MUTEX_EXIT(&rx_stats_mutex);
7449 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7451 fprintf(file, "Peer %x.%d.\n",
7452 ntohl(peer->host), (int)ntohs(peer->port));
7455 " Rtt %d, " "total sent %d, " "resent %d\n",
7456 peer->rtt, peer->nSent, peer->reSends);
7458 fprintf(file, " Packet size %d\n", peer->ifMTU);
7462 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7464 * This mutex protects the following static variables:
7468 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7469 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7471 #define LOCK_RX_DEBUG
7472 #define UNLOCK_RX_DEBUG
7473 #endif /* AFS_PTHREAD_ENV */
7475 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7477 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7478 u_char type, void *inputData, size_t inputLength,
7479 void *outputData, size_t outputLength)
7481 static afs_int32 counter = 100;
7482 time_t waitTime, waitCount;
7483 struct rx_header theader;
7486 struct timeval tv_now, tv_wake, tv_delta;
7487 struct sockaddr_in taddr, faddr;
7501 tp = &tbuffer[sizeof(struct rx_header)];
7502 taddr.sin_family = AF_INET;
7503 taddr.sin_port = remotePort;
7504 taddr.sin_addr.s_addr = remoteAddr;
7505 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7506 taddr.sin_len = sizeof(struct sockaddr_in);
7509 memset(&theader, 0, sizeof(theader));
7510 theader.epoch = htonl(999);
7512 theader.callNumber = htonl(counter);
7515 theader.type = type;
7516 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7517 theader.serviceId = 0;
7519 memcpy(tbuffer, &theader, sizeof(theader));
7520 memcpy(tp, inputData, inputLength);
7522 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7523 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7525 /* see if there's a packet available */
7526 gettimeofday(&tv_wake, NULL);
7527 tv_wake.tv_sec += waitTime;
7530 FD_SET(socket, &imask);
7531 tv_delta.tv_sec = tv_wake.tv_sec;
7532 tv_delta.tv_usec = tv_wake.tv_usec;
7533 gettimeofday(&tv_now, NULL);
7535 if (tv_delta.tv_usec < tv_now.tv_usec) {
7537 tv_delta.tv_usec += 1000000;
7540 tv_delta.tv_usec -= tv_now.tv_usec;
7542 if (tv_delta.tv_sec < tv_now.tv_sec) {
7546 tv_delta.tv_sec -= tv_now.tv_sec;
7549 code = select(0, &imask, 0, 0, &tv_delta);
7550 #else /* AFS_NT40_ENV */
7551 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7552 #endif /* AFS_NT40_ENV */
7553 if (code == 1 && FD_ISSET(socket, &imask)) {
7554 /* now receive a packet */
7555 faddrLen = sizeof(struct sockaddr_in);
7557 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7558 (struct sockaddr *)&faddr, &faddrLen);
7561 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7562 if (counter == ntohl(theader.callNumber))
7570 /* see if we've timed out */
7578 code -= sizeof(struct rx_header);
7579 if (code > outputLength)
7580 code = outputLength;
7581 memcpy(outputData, tp, code);
7584 #endif /* RXDEBUG */
7587 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7588 afs_uint16 remotePort, struct rx_debugStats * stat,
7589 afs_uint32 * supportedValues)
7591 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7593 struct rx_debugIn in;
7595 *supportedValues = 0;
7596 in.type = htonl(RX_DEBUGI_GETSTATS);
7599 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7600 &in, sizeof(in), stat, sizeof(*stat));
7603 * If the call was successful, fixup the version and indicate
7604 * what contents of the stat structure are valid.
7605 * Also do net to host conversion of fields here.
7609 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7610 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7612 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7613 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7615 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7616 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7618 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7619 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7621 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7622 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7624 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7625 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7627 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7628 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7630 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7631 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7633 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7634 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7636 stat->nFreePackets = ntohl(stat->nFreePackets);
7637 stat->packetReclaims = ntohl(stat->packetReclaims);
7638 stat->callsExecuted = ntohl(stat->callsExecuted);
7639 stat->nWaiting = ntohl(stat->nWaiting);
7640 stat->idleThreads = ntohl(stat->idleThreads);
7641 stat->nWaited = ntohl(stat->nWaited);
7642 stat->nPackets = ntohl(stat->nPackets);
7651 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7652 afs_uint16 remotePort, struct rx_statistics * stat,
7653 afs_uint32 * supportedValues)
7655 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7657 struct rx_debugIn in;
7658 afs_int32 *lp = (afs_int32 *) stat;
7662 * supportedValues is currently unused, but added to allow future
7663 * versioning of this function.
7666 *supportedValues = 0;
7667 in.type = htonl(RX_DEBUGI_RXSTATS);
7669 memset(stat, 0, sizeof(*stat));
7671 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7672 &in, sizeof(in), stat, sizeof(*stat));
7677 * Do net to host conversion here
7680 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7691 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7692 afs_uint16 remotePort, size_t version_length,
7695 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7697 return MakeDebugCall(socket, remoteAddr, remotePort,
7698 RX_PACKET_TYPE_VERSION, a, 1, version,
7706 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7707 afs_uint16 remotePort, afs_int32 * nextConnection,
7708 int allConnections, afs_uint32 debugSupportedValues,
7709 struct rx_debugConn * conn,
7710 afs_uint32 * supportedValues)
7712 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7714 struct rx_debugIn in;
7718 * supportedValues is currently unused, but added to allow future
7719 * versioning of this function.
7722 *supportedValues = 0;
7723 if (allConnections) {
7724 in.type = htonl(RX_DEBUGI_GETALLCONN);
7726 in.type = htonl(RX_DEBUGI_GETCONN);
7728 in.index = htonl(*nextConnection);
7729 memset(conn, 0, sizeof(*conn));
7731 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7732 &in, sizeof(in), conn, sizeof(*conn));
7735 *nextConnection += 1;
7738 * Convert old connection format to new structure.
7741 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7742 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7743 #define MOVEvL(a) (conn->a = vL->a)
7745 /* any old or unrecognized version... */
7746 for (i = 0; i < RX_MAXCALLS; i++) {
7747 MOVEvL(callState[i]);
7748 MOVEvL(callMode[i]);
7749 MOVEvL(callFlags[i]);
7750 MOVEvL(callOther[i]);
7752 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7753 MOVEvL(secStats.type);
7754 MOVEvL(secStats.level);
7755 MOVEvL(secStats.flags);
7756 MOVEvL(secStats.expires);
7757 MOVEvL(secStats.packetsReceived);
7758 MOVEvL(secStats.packetsSent);
7759 MOVEvL(secStats.bytesReceived);
7760 MOVEvL(secStats.bytesSent);
7765 * Do net to host conversion here
7767 * I don't convert host or port since we are most likely
7768 * going to want these in NBO.
7770 conn->cid = ntohl(conn->cid);
7771 conn->serial = ntohl(conn->serial);
7772 for (i = 0; i < RX_MAXCALLS; i++) {
7773 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7775 conn->error = ntohl(conn->error);
7776 conn->secStats.flags = ntohl(conn->secStats.flags);
7777 conn->secStats.expires = ntohl(conn->secStats.expires);
7778 conn->secStats.packetsReceived =
7779 ntohl(conn->secStats.packetsReceived);
7780 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7781 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7782 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7783 conn->epoch = ntohl(conn->epoch);
7784 conn->natMTU = ntohl(conn->natMTU);
7793 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7794 afs_uint16 remotePort, afs_int32 * nextPeer,
7795 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7796 afs_uint32 * supportedValues)
7798 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7800 struct rx_debugIn in;
7803 * supportedValues is currently unused, but added to allow future
7804 * versioning of this function.
7807 *supportedValues = 0;
7808 in.type = htonl(RX_DEBUGI_GETPEER);
7809 in.index = htonl(*nextPeer);
7810 memset(peer, 0, sizeof(*peer));
7812 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7813 &in, sizeof(in), peer, sizeof(*peer));
7819 * Do net to host conversion here
7821 * I don't convert host or port since we are most likely
7822 * going to want these in NBO.
7824 peer->ifMTU = ntohs(peer->ifMTU);
7825 peer->idleWhen = ntohl(peer->idleWhen);
7826 peer->refCount = ntohs(peer->refCount);
7827 peer->rtt = ntohl(peer->rtt);
7828 peer->rtt_dev = ntohl(peer->rtt_dev);
7829 peer->timeout.sec = 0;
7830 peer->timeout.usec = 0;
7831 peer->nSent = ntohl(peer->nSent);
7832 peer->reSends = ntohl(peer->reSends);
7833 peer->natMTU = ntohs(peer->natMTU);
7834 peer->maxMTU = ntohs(peer->maxMTU);
7835 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7836 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7837 peer->MTU = ntohs(peer->MTU);
7838 peer->cwind = ntohs(peer->cwind);
7839 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7840 peer->congestSeq = ntohs(peer->congestSeq);
7841 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7842 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7843 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7844 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7853 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7854 struct rx_debugPeer * peerStats)
7857 afs_int32 error = 1; /* default to "did not succeed" */
7858 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7860 MUTEX_ENTER(&rx_peerHashTable_lock);
7861 for(tp = rx_peerHashTable[hashValue];
7862 tp != NULL; tp = tp->next) {
7863 if (tp->host == peerHost)
7869 MUTEX_EXIT(&rx_peerHashTable_lock);
7873 MUTEX_ENTER(&tp->peer_lock);
7874 peerStats->host = tp->host;
7875 peerStats->port = tp->port;
7876 peerStats->ifMTU = tp->ifMTU;
7877 peerStats->idleWhen = tp->idleWhen;
7878 peerStats->refCount = tp->refCount;
7879 peerStats->burstSize = 0;
7880 peerStats->burst = 0;
7881 peerStats->burstWait.sec = 0;
7882 peerStats->burstWait.usec = 0;
7883 peerStats->rtt = tp->rtt;
7884 peerStats->rtt_dev = tp->rtt_dev;
7885 peerStats->timeout.sec = 0;
7886 peerStats->timeout.usec = 0;
7887 peerStats->nSent = tp->nSent;
7888 peerStats->reSends = tp->reSends;
7889 peerStats->natMTU = tp->natMTU;
7890 peerStats->maxMTU = tp->maxMTU;
7891 peerStats->maxDgramPackets = tp->maxDgramPackets;
7892 peerStats->ifDgramPackets = tp->ifDgramPackets;
7893 peerStats->MTU = tp->MTU;
7894 peerStats->cwind = tp->cwind;
7895 peerStats->nDgramPackets = tp->nDgramPackets;
7896 peerStats->congestSeq = tp->congestSeq;
7897 peerStats->bytesSent.high = tp->bytesSent >> 32;
7898 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7899 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7900 peerStats->bytesReceived.low
7901 = tp->bytesReceived & MAX_AFS_UINT32;
7902 MUTEX_EXIT(&tp->peer_lock);
7904 MUTEX_ENTER(&rx_peerHashTable_lock);
7907 MUTEX_EXIT(&rx_peerHashTable_lock);
7915 struct rx_serverQueueEntry *np;
7918 struct rx_call *call;
7919 struct rx_serverQueueEntry *sq;
7922 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7923 return; /* Already shutdown. */
7927 #ifndef AFS_PTHREAD_ENV
7928 FD_ZERO(&rx_selectMask);
7929 #endif /* AFS_PTHREAD_ENV */
7930 rxi_dataQuota = RX_MAX_QUOTA;
7931 #ifndef AFS_PTHREAD_ENV
7933 #endif /* AFS_PTHREAD_ENV */
7936 #ifndef AFS_PTHREAD_ENV
7937 #ifndef AFS_USE_GETTIMEOFDAY
7939 #endif /* AFS_USE_GETTIMEOFDAY */
7940 #endif /* AFS_PTHREAD_ENV */
7942 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7943 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7944 opr_queue_Remove(&call->entry);
7945 rxi_Free(call, sizeof(struct rx_call));
7948 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7949 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7951 opr_queue_Remove(&sq->entry);
7956 struct rx_peer **peer_ptr, **peer_end;
7957 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7958 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7960 struct rx_peer *peer, *next;
7962 MUTEX_ENTER(&rx_peerHashTable_lock);
7963 for (peer = *peer_ptr; peer; peer = next) {
7964 struct opr_queue *cursor, *store;
7967 MUTEX_ENTER(&rx_rpc_stats);
7968 MUTEX_ENTER(&peer->peer_lock);
7969 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7970 unsigned int num_funcs;
7971 struct rx_interface_stat *rpc_stat
7972 = opr_queue_Entry(cursor, struct rx_interface_stat,
7976 opr_queue_Remove(&rpc_stat->entry);
7977 opr_queue_Remove(&rpc_stat->entryPeers);
7978 num_funcs = rpc_stat->stats[0].func_total;
7980 sizeof(rx_interface_stat_t) +
7981 rpc_stat->stats[0].func_total *
7982 sizeof(rx_function_entry_v1_t);
7984 rxi_Free(rpc_stat, space);
7986 /* rx_rpc_stats must be held */
7987 rxi_rpc_peer_stat_cnt -= num_funcs;
7989 MUTEX_EXIT(&peer->peer_lock);
7990 MUTEX_EXIT(&rx_rpc_stats);
7994 if (rx_stats_active)
7995 rx_atomic_dec(&rx_stats.nPeerStructs);
7997 MUTEX_EXIT(&rx_peerHashTable_lock);
8000 for (i = 0; i < RX_MAX_SERVICES; i++) {
8002 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
8004 for (i = 0; i < rx_hashTableSize; i++) {
8005 struct rx_connection *tc, *ntc;
8006 MUTEX_ENTER(&rx_connHashTable_lock);
8007 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8009 for (j = 0; j < RX_MAXCALLS; j++) {
8011 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8014 rxi_Free(tc, sizeof(*tc));
8016 MUTEX_EXIT(&rx_connHashTable_lock);
8019 MUTEX_ENTER(&freeSQEList_lock);
8021 while ((np = rx_FreeSQEList)) {
8022 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8023 MUTEX_DESTROY(&np->lock);
8024 rxi_Free(np, sizeof(*np));
8027 MUTEX_EXIT(&freeSQEList_lock);
8028 MUTEX_DESTROY(&freeSQEList_lock);
8029 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8030 MUTEX_DESTROY(&rx_connHashTable_lock);
8031 MUTEX_DESTROY(&rx_peerHashTable_lock);
8032 MUTEX_DESTROY(&rx_serverPool_lock);
8034 osi_Free(rx_connHashTable,
8035 rx_hashTableSize * sizeof(struct rx_connection *));
8036 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8038 UNPIN(rx_connHashTable,
8039 rx_hashTableSize * sizeof(struct rx_connection *));
8040 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8042 rxi_FreeAllPackets();
8044 MUTEX_ENTER(&rx_quota_mutex);
8045 rxi_dataQuota = RX_MAX_QUOTA;
8046 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8047 MUTEX_EXIT(&rx_quota_mutex);
8053 * Routines to implement connection specific data.
8057 rx_KeyCreate(rx_destructor_t rtn)
8060 MUTEX_ENTER(&rxi_keyCreate_lock);
8061 key = rxi_keyCreate_counter++;
8062 rxi_keyCreate_destructor = (rx_destructor_t *)
8063 realloc((void *)rxi_keyCreate_destructor,
8064 (key + 1) * sizeof(rx_destructor_t));
8065 rxi_keyCreate_destructor[key] = rtn;
8066 MUTEX_EXIT(&rxi_keyCreate_lock);
8071 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8074 MUTEX_ENTER(&conn->conn_data_lock);
8075 if (!conn->specific) {
8076 conn->specific = malloc((key + 1) * sizeof(void *));
8077 for (i = 0; i < key; i++)
8078 conn->specific[i] = NULL;
8079 conn->nSpecific = key + 1;
8080 conn->specific[key] = ptr;
8081 } else if (key >= conn->nSpecific) {
8082 conn->specific = (void **)
8083 realloc(conn->specific, (key + 1) * sizeof(void *));
8084 for (i = conn->nSpecific; i < key; i++)
8085 conn->specific[i] = NULL;
8086 conn->nSpecific = key + 1;
8087 conn->specific[key] = ptr;
8089 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8090 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8091 conn->specific[key] = ptr;
8093 MUTEX_EXIT(&conn->conn_data_lock);
8097 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8100 MUTEX_ENTER(&svc->svc_data_lock);
8101 if (!svc->specific) {
8102 svc->specific = malloc((key + 1) * sizeof(void *));
8103 for (i = 0; i < key; i++)
8104 svc->specific[i] = NULL;
8105 svc->nSpecific = key + 1;
8106 svc->specific[key] = ptr;
8107 } else if (key >= svc->nSpecific) {
8108 svc->specific = (void **)
8109 realloc(svc->specific, (key + 1) * sizeof(void *));
8110 for (i = svc->nSpecific; i < key; i++)
8111 svc->specific[i] = NULL;
8112 svc->nSpecific = key + 1;
8113 svc->specific[key] = ptr;
8115 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8116 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8117 svc->specific[key] = ptr;
8119 MUTEX_EXIT(&svc->svc_data_lock);
8123 rx_GetSpecific(struct rx_connection *conn, int key)
8126 MUTEX_ENTER(&conn->conn_data_lock);
8127 if (key >= conn->nSpecific)
8130 ptr = conn->specific[key];
8131 MUTEX_EXIT(&conn->conn_data_lock);
8136 rx_GetServiceSpecific(struct rx_service *svc, int key)
8139 MUTEX_ENTER(&svc->svc_data_lock);
8140 if (key >= svc->nSpecific)
8143 ptr = svc->specific[key];
8144 MUTEX_EXIT(&svc->svc_data_lock);
8149 #endif /* !KERNEL */
8152 * processStats is a queue used to store the statistics for the local
8153 * process. Its contents are similar to the contents of the rpcStats
8154 * queue on a rx_peer structure, but the actual data stored within
8155 * this queue contains totals across the lifetime of the process (assuming
8156 * the stats have not been reset) - unlike the per peer structures
8157 * which can come and go based upon the peer lifetime.
8160 static struct opr_queue processStats = { &processStats, &processStats };
8163 * peerStats is a queue used to store the statistics for all peer structs.
8164 * Its contents are the union of all the peer rpcStats queues.
8167 static struct opr_queue peerStats = { &peerStats, &peerStats };
8170 * rxi_monitor_processStats is used to turn process wide stat collection
8174 static int rxi_monitor_processStats = 0;
8177 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8180 static int rxi_monitor_peerStats = 0;
8184 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8186 rpc_stat->invocations = 0;
8187 rpc_stat->bytes_sent = 0;
8188 rpc_stat->bytes_rcvd = 0;
8189 rpc_stat->queue_time_sum.sec = 0;
8190 rpc_stat->queue_time_sum.usec = 0;
8191 rpc_stat->queue_time_sum_sqr.sec = 0;
8192 rpc_stat->queue_time_sum_sqr.usec = 0;
8193 rpc_stat->queue_time_min.sec = 9999999;
8194 rpc_stat->queue_time_min.usec = 9999999;
8195 rpc_stat->queue_time_max.sec = 0;
8196 rpc_stat->queue_time_max.usec = 0;
8197 rpc_stat->execution_time_sum.sec = 0;
8198 rpc_stat->execution_time_sum.usec = 0;
8199 rpc_stat->execution_time_sum_sqr.sec = 0;
8200 rpc_stat->execution_time_sum_sqr.usec = 0;
8201 rpc_stat->execution_time_min.sec = 9999999;
8202 rpc_stat->execution_time_min.usec = 9999999;
8203 rpc_stat->execution_time_max.sec = 0;
8204 rpc_stat->execution_time_max.usec = 0;
8208 * Given all of the information for a particular rpc
8209 * call, find or create (if requested) the stat structure for the rpc.
8212 * the queue of stats that will be updated with the new value
8214 * @param rxInterface
8215 * a unique number that identifies the rpc interface
8218 * the total number of functions in this interface. this is only
8219 * required if create is true
8222 * if true, this invocation was made to a server
8225 * the ip address of the remote host. this is only required if create
8226 * and addToPeerList are true
8229 * the port of the remote host. this is only required if create
8230 * and addToPeerList are true
8232 * @param addToPeerList
8233 * if != 0, add newly created stat to the global peer list
8236 * if a new stats structure is allocated, the counter will
8237 * be updated with the new number of allocated stat structures.
8238 * only required if create is true
8241 * if no stats structure exists, allocate one
8245 static rx_interface_stat_p
8246 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8247 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8248 afs_uint32 remotePort, int addToPeerList,
8249 unsigned int *counter, int create)
8251 rx_interface_stat_p rpc_stat = NULL;
8252 struct opr_queue *cursor;
8255 * See if there's already a structure for this interface
8258 for (opr_queue_Scan(stats, cursor)) {
8259 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8261 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8262 && (rpc_stat->stats[0].remote_is_server == isServer))
8266 /* if they didn't ask us to create, we're done */
8268 if (opr_queue_IsEnd(stats, cursor))
8274 /* can't proceed without these */
8275 if (!totalFunc || !counter)
8279 * Didn't find a match so allocate a new structure and add it to the
8283 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8284 || (rpc_stat->stats[0].interfaceId != rxInterface)
8285 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8290 sizeof(rx_interface_stat_t) +
8291 totalFunc * sizeof(rx_function_entry_v1_t);
8293 rpc_stat = rxi_Alloc(space);
8294 if (rpc_stat == NULL)
8297 *counter += totalFunc;
8298 for (i = 0; i < totalFunc; i++) {
8299 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8300 rpc_stat->stats[i].remote_peer = remoteHost;
8301 rpc_stat->stats[i].remote_port = remotePort;
8302 rpc_stat->stats[i].remote_is_server = isServer;
8303 rpc_stat->stats[i].interfaceId = rxInterface;
8304 rpc_stat->stats[i].func_total = totalFunc;
8305 rpc_stat->stats[i].func_index = i;
8307 opr_queue_Prepend(stats, &rpc_stat->entry);
8308 if (addToPeerList) {
8309 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8316 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8318 rx_interface_stat_p rpc_stat;
8321 if (rxInterface == -1)
8324 MUTEX_ENTER(&rx_rpc_stats);
8325 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8328 totalFunc = rpc_stat->stats[0].func_total;
8329 for (i = 0; i < totalFunc; i++)
8330 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8332 MUTEX_EXIT(&rx_rpc_stats);
8337 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8339 rx_interface_stat_p rpc_stat;
8341 struct rx_peer * peer;
8343 if (rxInterface == -1)
8346 peer = rxi_FindPeer(peerHost, peerPort, 0);
8350 MUTEX_ENTER(&rx_rpc_stats);
8351 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8354 totalFunc = rpc_stat->stats[0].func_total;
8355 for (i = 0; i < totalFunc; i++)
8356 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8358 MUTEX_EXIT(&rx_rpc_stats);
8363 rx_CopyProcessRPCStats(afs_uint64 op)
8365 rx_interface_stat_p rpc_stat;
8366 rx_function_entry_v1_p rpcop_stat =
8367 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8368 int currentFunc = (op & MAX_AFS_UINT32);
8369 afs_int32 rxInterface = (op >> 32);
8371 if (!rxi_monitor_processStats)
8374 if (rxInterface == -1)
8377 if (rpcop_stat == NULL)
8380 MUTEX_ENTER(&rx_rpc_stats);
8381 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8384 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8385 sizeof(rx_function_entry_v1_t));
8386 MUTEX_EXIT(&rx_rpc_stats);
8388 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8395 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8397 rx_interface_stat_p rpc_stat;
8398 rx_function_entry_v1_p rpcop_stat =
8399 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8400 int currentFunc = (op & MAX_AFS_UINT32);
8401 afs_int32 rxInterface = (op >> 32);
8402 struct rx_peer *peer;
8404 if (!rxi_monitor_peerStats)
8407 if (rxInterface == -1)
8410 if (rpcop_stat == NULL)
8413 peer = rxi_FindPeer(peerHost, peerPort, 0);
8417 MUTEX_ENTER(&rx_rpc_stats);
8418 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8421 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8422 sizeof(rx_function_entry_v1_t));
8423 MUTEX_EXIT(&rx_rpc_stats);
8425 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8432 rx_ReleaseRPCStats(void *stats)
8435 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8439 * Given all of the information for a particular rpc
8440 * call, create (if needed) and update the stat totals for the rpc.
8443 * the queue of stats that will be updated with the new value
8445 * @param rxInterface
8446 * a unique number that identifies the rpc interface
8448 * @param currentFunc
8449 * the index of the function being invoked
8452 * the total number of functions in this interface
8455 * the amount of time this function waited for a thread
8458 * the amount of time this function invocation took to execute
8461 * the number bytes sent by this invocation
8464 * the number bytes received by this invocation
8467 * if true, this invocation was made to a server
8470 * the ip address of the remote host
8473 * the port of the remote host
8475 * @param addToPeerList
8476 * if != 0, add newly created stat to the global peer list
8479 * if a new stats structure is allocated, the counter will
8480 * be updated with the new number of allocated stat structures
8485 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8486 afs_uint32 currentFunc, afs_uint32 totalFunc,
8487 struct clock *queueTime, struct clock *execTime,
8488 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8489 afs_uint32 remoteHost, afs_uint32 remotePort,
8490 int addToPeerList, unsigned int *counter)
8493 rx_interface_stat_p rpc_stat;
8495 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8496 remoteHost, remotePort, addToPeerList, counter,
8504 * Increment the stats for this function
8507 rpc_stat->stats[currentFunc].invocations++;
8508 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8509 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8510 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8511 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8512 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8513 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8515 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8516 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8518 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8519 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8521 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8522 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8524 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8525 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8533 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8534 afs_uint32 currentFunc, afs_uint32 totalFunc,
8535 struct clock *queueTime, struct clock *execTime,
8536 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8540 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8543 MUTEX_ENTER(&rx_rpc_stats);
8545 if (rxi_monitor_peerStats) {
8546 MUTEX_ENTER(&peer->peer_lock);
8547 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8548 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8549 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8550 MUTEX_EXIT(&peer->peer_lock);
8553 if (rxi_monitor_processStats) {
8554 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8555 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8556 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8559 MUTEX_EXIT(&rx_rpc_stats);
8563 * Increment the times and count for a particular rpc function.
8565 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8566 * call rx_RecordCallStatistics instead, so the public version of this
8567 * function is left purely for legacy callers.
8570 * The peer who invoked the rpc
8572 * @param rxInterface
8573 * A unique number that identifies the rpc interface
8575 * @param currentFunc
8576 * The index of the function being invoked
8579 * The total number of functions in this interface
8582 * The amount of time this function waited for a thread
8585 * The amount of time this function invocation took to execute
8588 * The number bytes sent by this invocation
8591 * The number bytes received by this invocation
8594 * If true, this invocation was made to a server
8598 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8599 afs_uint32 currentFunc, afs_uint32 totalFunc,
8600 struct clock *queueTime, struct clock *execTime,
8601 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8607 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8608 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8610 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8611 queueTime, execTime, sent64, rcvd64,
8618 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8622 * IN callerVersion - the rpc stat version of the caller.
8624 * IN count - the number of entries to marshall.
8626 * IN stats - pointer to stats to be marshalled.
8628 * OUT ptr - Where to store the marshalled data.
8635 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8636 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8642 * We only support the first version
8644 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8645 *(ptr++) = stats->remote_peer;
8646 *(ptr++) = stats->remote_port;
8647 *(ptr++) = stats->remote_is_server;
8648 *(ptr++) = stats->interfaceId;
8649 *(ptr++) = stats->func_total;
8650 *(ptr++) = stats->func_index;
8651 *(ptr++) = stats->invocations >> 32;
8652 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8653 *(ptr++) = stats->bytes_sent >> 32;
8654 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8655 *(ptr++) = stats->bytes_rcvd >> 32;
8656 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8657 *(ptr++) = stats->queue_time_sum.sec;
8658 *(ptr++) = stats->queue_time_sum.usec;
8659 *(ptr++) = stats->queue_time_sum_sqr.sec;
8660 *(ptr++) = stats->queue_time_sum_sqr.usec;
8661 *(ptr++) = stats->queue_time_min.sec;
8662 *(ptr++) = stats->queue_time_min.usec;
8663 *(ptr++) = stats->queue_time_max.sec;
8664 *(ptr++) = stats->queue_time_max.usec;
8665 *(ptr++) = stats->execution_time_sum.sec;
8666 *(ptr++) = stats->execution_time_sum.usec;
8667 *(ptr++) = stats->execution_time_sum_sqr.sec;
8668 *(ptr++) = stats->execution_time_sum_sqr.usec;
8669 *(ptr++) = stats->execution_time_min.sec;
8670 *(ptr++) = stats->execution_time_min.usec;
8671 *(ptr++) = stats->execution_time_max.sec;
8672 *(ptr++) = stats->execution_time_max.usec;
8678 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8683 * IN callerVersion - the rpc stat version of the caller
8685 * OUT myVersion - the rpc stat version of this function
8687 * OUT clock_sec - local time seconds
8689 * OUT clock_usec - local time microseconds
8691 * OUT allocSize - the number of bytes allocated to contain stats
8693 * OUT statCount - the number stats retrieved from this process.
8695 * OUT stats - the actual stats retrieved from this process.
8699 * Returns void. If successful, stats will != NULL.
8703 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8704 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8705 size_t * allocSize, afs_uint32 * statCount,
8706 afs_uint32 ** stats)
8716 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8719 * Check to see if stats are enabled
8722 MUTEX_ENTER(&rx_rpc_stats);
8723 if (!rxi_monitor_processStats) {
8724 MUTEX_EXIT(&rx_rpc_stats);
8728 clock_GetTime(&now);
8729 *clock_sec = now.sec;
8730 *clock_usec = now.usec;
8733 * Allocate the space based upon the caller version
8735 * If the client is at an older version than we are,
8736 * we return the statistic data in the older data format, but
8737 * we still return our version number so the client knows we
8738 * are maintaining more data than it can retrieve.
8741 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8742 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8743 *statCount = rxi_rpc_process_stat_cnt;
8746 * This can't happen yet, but in the future version changes
8747 * can be handled by adding additional code here
8751 if (space > (size_t) 0) {
8753 ptr = *stats = rxi_Alloc(space);
8756 struct opr_queue *cursor;
8758 for (opr_queue_Scan(&processStats, cursor)) {
8759 struct rx_interface_stat *rpc_stat =
8760 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8762 * Copy the data based upon the caller version
8764 rx_MarshallProcessRPCStats(callerVersion,
8765 rpc_stat->stats[0].func_total,
8766 rpc_stat->stats, &ptr);
8772 MUTEX_EXIT(&rx_rpc_stats);
8777 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8781 * IN callerVersion - the rpc stat version of the caller
8783 * OUT myVersion - the rpc stat version of this function
8785 * OUT clock_sec - local time seconds
8787 * OUT clock_usec - local time microseconds
8789 * OUT allocSize - the number of bytes allocated to contain stats
8791 * OUT statCount - the number of stats retrieved from the individual
8794 * OUT stats - the actual stats retrieved from the individual peer structures.
8798 * Returns void. If successful, stats will != NULL.
8802 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8803 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8804 size_t * allocSize, afs_uint32 * statCount,
8805 afs_uint32 ** stats)
8815 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8818 * Check to see if stats are enabled
8821 MUTEX_ENTER(&rx_rpc_stats);
8822 if (!rxi_monitor_peerStats) {
8823 MUTEX_EXIT(&rx_rpc_stats);
8827 clock_GetTime(&now);
8828 *clock_sec = now.sec;
8829 *clock_usec = now.usec;
8832 * Allocate the space based upon the caller version
8834 * If the client is at an older version than we are,
8835 * we return the statistic data in the older data format, but
8836 * we still return our version number so the client knows we
8837 * are maintaining more data than it can retrieve.
8840 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8841 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8842 *statCount = rxi_rpc_peer_stat_cnt;
8845 * This can't happen yet, but in the future version changes
8846 * can be handled by adding additional code here
8850 if (space > (size_t) 0) {
8852 ptr = *stats = rxi_Alloc(space);
8855 struct opr_queue *cursor;
8857 for (opr_queue_Scan(&peerStats, cursor)) {
8858 struct rx_interface_stat *rpc_stat
8859 = opr_queue_Entry(cursor, struct rx_interface_stat,
8863 * Copy the data based upon the caller version
8865 rx_MarshallProcessRPCStats(callerVersion,
8866 rpc_stat->stats[0].func_total,
8867 rpc_stat->stats, &ptr);
8873 MUTEX_EXIT(&rx_rpc_stats);
8878 * rx_FreeRPCStats - free memory allocated by
8879 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8883 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8884 * rx_RetrievePeerRPCStats
8886 * IN allocSize - the number of bytes in stats.
8894 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8896 rxi_Free(stats, allocSize);
8900 * rx_queryProcessRPCStats - see if process rpc stat collection is
8901 * currently enabled.
8907 * Returns 0 if stats are not enabled != 0 otherwise
8911 rx_queryProcessRPCStats(void)
8914 MUTEX_ENTER(&rx_rpc_stats);
8915 rc = rxi_monitor_processStats;
8916 MUTEX_EXIT(&rx_rpc_stats);
8921 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8927 * Returns 0 if stats are not enabled != 0 otherwise
8931 rx_queryPeerRPCStats(void)
8934 MUTEX_ENTER(&rx_rpc_stats);
8935 rc = rxi_monitor_peerStats;
8936 MUTEX_EXIT(&rx_rpc_stats);
8941 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8951 rx_enableProcessRPCStats(void)
8953 MUTEX_ENTER(&rx_rpc_stats);
8954 rx_enable_stats = 1;
8955 rxi_monitor_processStats = 1;
8956 MUTEX_EXIT(&rx_rpc_stats);
8960 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8970 rx_enablePeerRPCStats(void)
8972 MUTEX_ENTER(&rx_rpc_stats);
8973 rx_enable_stats = 1;
8974 rxi_monitor_peerStats = 1;
8975 MUTEX_EXIT(&rx_rpc_stats);
8979 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8989 rx_disableProcessRPCStats(void)
8991 struct opr_queue *cursor, *store;
8994 MUTEX_ENTER(&rx_rpc_stats);
8997 * Turn off process statistics and if peer stats is also off, turn
9001 rxi_monitor_processStats = 0;
9002 if (rxi_monitor_peerStats == 0) {
9003 rx_enable_stats = 0;
9006 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
9007 unsigned int num_funcs = 0;
9008 struct rx_interface_stat *rpc_stat
9009 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9011 opr_queue_Remove(&rpc_stat->entry);
9013 num_funcs = rpc_stat->stats[0].func_total;
9015 sizeof(rx_interface_stat_t) +
9016 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
9018 rxi_Free(rpc_stat, space);
9019 rxi_rpc_process_stat_cnt -= num_funcs;
9021 MUTEX_EXIT(&rx_rpc_stats);
9025 * rx_disablePeerRPCStats - stop rpc stat collection for peers
9035 rx_disablePeerRPCStats(void)
9037 struct rx_peer **peer_ptr, **peer_end;
9041 * Turn off peer statistics and if process stats is also off, turn
9045 rxi_monitor_peerStats = 0;
9046 if (rxi_monitor_processStats == 0) {
9047 rx_enable_stats = 0;
9050 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9051 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9053 struct rx_peer *peer, *next, *prev;
9055 MUTEX_ENTER(&rx_peerHashTable_lock);
9056 MUTEX_ENTER(&rx_rpc_stats);
9057 for (prev = peer = *peer_ptr; peer; peer = next) {
9059 code = MUTEX_TRYENTER(&peer->peer_lock);
9062 struct opr_queue *cursor, *store;
9064 if (prev == *peer_ptr) {
9075 MUTEX_EXIT(&rx_peerHashTable_lock);
9077 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9078 unsigned int num_funcs = 0;
9079 struct rx_interface_stat *rpc_stat
9080 = opr_queue_Entry(cursor, struct rx_interface_stat,
9083 opr_queue_Remove(&rpc_stat->entry);
9084 opr_queue_Remove(&rpc_stat->entryPeers);
9085 num_funcs = rpc_stat->stats[0].func_total;
9087 sizeof(rx_interface_stat_t) +
9088 rpc_stat->stats[0].func_total *
9089 sizeof(rx_function_entry_v1_t);
9091 rxi_Free(rpc_stat, space);
9092 rxi_rpc_peer_stat_cnt -= num_funcs;
9094 MUTEX_EXIT(&peer->peer_lock);
9096 MUTEX_ENTER(&rx_peerHashTable_lock);
9106 MUTEX_EXIT(&rx_rpc_stats);
9107 MUTEX_EXIT(&rx_peerHashTable_lock);
9112 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9117 * IN clearFlag - flag indicating which stats to clear
9125 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9127 struct opr_queue *cursor;
9129 MUTEX_ENTER(&rx_rpc_stats);
9131 for (opr_queue_Scan(&processStats, cursor)) {
9132 unsigned int num_funcs = 0, i;
9133 struct rx_interface_stat *rpc_stat
9134 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9136 num_funcs = rpc_stat->stats[0].func_total;
9137 for (i = 0; i < num_funcs; i++) {
9138 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9139 rpc_stat->stats[i].invocations = 0;
9141 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9142 rpc_stat->stats[i].bytes_sent = 0;
9144 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9145 rpc_stat->stats[i].bytes_rcvd = 0;
9147 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9148 rpc_stat->stats[i].queue_time_sum.sec = 0;
9149 rpc_stat->stats[i].queue_time_sum.usec = 0;
9151 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9152 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9153 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9155 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9156 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9157 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9159 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9160 rpc_stat->stats[i].queue_time_max.sec = 0;
9161 rpc_stat->stats[i].queue_time_max.usec = 0;
9163 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9164 rpc_stat->stats[i].execution_time_sum.sec = 0;
9165 rpc_stat->stats[i].execution_time_sum.usec = 0;
9167 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9168 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9169 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9171 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9172 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9173 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9175 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9176 rpc_stat->stats[i].execution_time_max.sec = 0;
9177 rpc_stat->stats[i].execution_time_max.usec = 0;
9182 MUTEX_EXIT(&rx_rpc_stats);
9186 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9191 * IN clearFlag - flag indicating which stats to clear
9199 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9201 struct opr_queue *cursor;
9203 MUTEX_ENTER(&rx_rpc_stats);
9205 for (opr_queue_Scan(&peerStats, cursor)) {
9206 unsigned int num_funcs, i;
9207 struct rx_interface_stat *rpc_stat
9208 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9210 num_funcs = rpc_stat->stats[0].func_total;
9211 for (i = 0; i < num_funcs; i++) {
9212 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9213 rpc_stat->stats[i].invocations = 0;
9215 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9216 rpc_stat->stats[i].bytes_sent = 0;
9218 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9219 rpc_stat->stats[i].bytes_rcvd = 0;
9221 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9222 rpc_stat->stats[i].queue_time_sum.sec = 0;
9223 rpc_stat->stats[i].queue_time_sum.usec = 0;
9225 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9226 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9227 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9229 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9230 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9231 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9233 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9234 rpc_stat->stats[i].queue_time_max.sec = 0;
9235 rpc_stat->stats[i].queue_time_max.usec = 0;
9237 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9238 rpc_stat->stats[i].execution_time_sum.sec = 0;
9239 rpc_stat->stats[i].execution_time_sum.usec = 0;
9241 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9242 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9243 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9245 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9246 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9247 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9249 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9250 rpc_stat->stats[i].execution_time_max.sec = 0;
9251 rpc_stat->stats[i].execution_time_max.usec = 0;
9256 MUTEX_EXIT(&rx_rpc_stats);
9260 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9261 * is authorized to enable/disable/clear RX statistics.
9263 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9266 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9268 rxi_rxstat_userok = proc;
9272 rx_RxStatUserOk(struct rx_call *call)
9274 if (!rxi_rxstat_userok)
9276 return rxi_rxstat_userok(call);
9281 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9282 * function in the MSVC runtime DLL (msvcrt.dll).
9284 * Note: the system serializes calls to this function.
9287 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9288 DWORD reason, /* reason function is being called */
9289 LPVOID reserved) /* reserved for future use */
9292 case DLL_PROCESS_ATTACH:
9293 /* library is being attached to a process */
9297 case DLL_PROCESS_DETACH:
9304 #endif /* AFS_NT40_ENV */
9307 int rx_DumpCalls(FILE *outputFile, char *cookie)
9309 #ifdef RXDEBUG_PACKET
9310 #ifdef KDUMP_RX_LOCK
9311 struct rx_call_rx_lock *c;
9318 #define RXDPRINTF sprintf
9319 #define RXDPRINTOUT output
9321 #define RXDPRINTF fprintf
9322 #define RXDPRINTOUT outputFile
9325 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9327 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9330 for (c = rx_allCallsp; c; c = c->allNextp) {
9331 u_short rqc, tqc, iovqc;
9333 MUTEX_ENTER(&c->lock);
9334 rqc = opr_queue_Count(&c->rq);
9335 tqc = opr_queue_Count(&c->tq);
9336 iovqc = opr_queue_Count(&c->app.iovq);
9338 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, "
9339 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9340 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9341 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9342 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9343 #ifdef RX_ENABLE_LOCKS
9346 #ifdef RX_REFCOUNT_CHECK
9347 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9348 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9351 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,
9352 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9353 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9354 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9355 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9356 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9357 #ifdef RX_ENABLE_LOCKS
9358 , (afs_uint32)c->refCount
9360 #ifdef RX_REFCOUNT_CHECK
9361 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9364 MUTEX_EXIT(&c->lock);
9367 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9370 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9372 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9374 #endif /* RXDEBUG_PACKET */