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>
76 #include "rx_atomic.h"
77 #include "rx_globals.h"
79 #include "rx_internal.h"
86 #include "rx_packet.h"
88 #include <afs/rxgen_consts.h>
91 #ifdef AFS_PTHREAD_ENV
93 int (*registerProgram) (pid_t, char *) = 0;
94 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
97 int (*registerProgram) (PROCESS, char *) = 0;
98 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
102 /* Local static routines */
103 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
104 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
105 struct rx_call *, struct rx_peer *,
107 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
109 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
110 void *dummy, int dummy2);
111 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
112 void *dummy, int dummy2);
113 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
114 void *unused, int unused2);
115 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
116 void *unused2, int unused3);
117 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
118 struct rx_packet *packet,
119 int istack, int force);
120 static void rxi_AckAll(struct rx_call *call);
121 static struct rx_connection
122 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
123 u_short serviceId, afs_uint32 cid,
124 afs_uint32 epoch, int type, u_int securityIndex);
125 static struct rx_packet
126 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
127 int istack, osi_socket socket,
128 afs_uint32 host, u_short port, int *tnop,
129 struct rx_call **newcallp);
130 static struct rx_packet
131 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
133 static struct rx_packet
134 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
135 struct rx_packet *np, int istack);
136 static struct rx_packet
137 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
138 struct rx_packet *np, int istack);
139 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
140 int *tnop, struct rx_call **newcallp);
141 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
142 static void rxi_ClearReceiveQueue(struct rx_call *call);
143 static void rxi_ResetCall(struct rx_call *call, int newcall);
144 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
145 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
146 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
147 static void rxi_KeepAliveOn(struct rx_call *call);
148 static void rxi_GrowMTUOn(struct rx_call *call);
149 static void rxi_ChallengeOn(struct rx_connection *conn);
151 #ifdef RX_ENABLE_LOCKS
152 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
153 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
155 static int rxi_CheckCall(struct rx_call *call);
158 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
160 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
161 rx_atomic_t rxi_start_in_error;
163 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
165 /* Constant delay time before sending an acknowledge of the last packet
166 * received. This is to avoid sending an extra acknowledge when the
167 * client is about to make another call, anyway, or the server is
170 * The lastAckDelay may not exceeed 400ms without causing peers to
171 * unecessarily timeout.
173 struct clock rx_lastAckDelay = {0, 400000};
175 /* Constant delay time before sending a soft ack when none was requested.
176 * This is to make sure we send soft acks before the sender times out,
177 * Normally we wait and send a hard ack when the receiver consumes the packet
179 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
180 * will require changes to the peer's RTT calculations.
182 struct clock rx_softAckDelay = {0, 100000};
185 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
186 * currently allocated within rx. This number is used to allocate the
187 * memory required to return the statistics when queried.
188 * Protected by the rx_rpc_stats mutex.
191 static unsigned int rxi_rpc_peer_stat_cnt;
194 * rxi_rpc_process_stat_cnt counts the total number of local process stat
195 * structures currently allocated within rx. The number is used to allocate
196 * the memory required to return the statistics when queried.
197 * Protected by the rx_rpc_stats mutex.
200 static unsigned int rxi_rpc_process_stat_cnt;
203 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
204 * errors should be reported to the application when a call channel appears busy
205 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
206 * and there are other call channels in the connection that are not busy.
207 * If 0, we do not return errors upon receiving busy packets; we just keep
208 * trying on the same call channel until we hit a timeout.
210 static afs_int32 rxi_busyChannelError = 0;
212 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
213 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
215 #if !defined(offsetof)
216 #include <stddef.h> /* for definition of offsetof() */
219 #ifdef RX_ENABLE_LOCKS
220 afs_kmutex_t rx_atomic_mutex;
223 /* Forward prototypes */
224 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
227 putConnection (struct rx_connection *conn) {
228 MUTEX_ENTER(&rx_refcnt_mutex);
230 MUTEX_EXIT(&rx_refcnt_mutex);
233 #ifdef AFS_PTHREAD_ENV
236 * Use procedural initialization of mutexes/condition variables
240 extern afs_kmutex_t rx_quota_mutex;
241 extern afs_kmutex_t rx_pthread_mutex;
242 extern afs_kmutex_t rx_packets_mutex;
243 extern afs_kmutex_t rx_refcnt_mutex;
244 extern afs_kmutex_t des_init_mutex;
245 extern afs_kmutex_t des_random_mutex;
246 extern afs_kmutex_t rx_clock_mutex;
247 extern afs_kmutex_t rxi_connCacheMutex;
248 extern afs_kmutex_t event_handler_mutex;
249 extern afs_kmutex_t listener_mutex;
250 extern afs_kmutex_t rx_if_init_mutex;
251 extern afs_kmutex_t rx_if_mutex;
253 extern afs_kcondvar_t rx_event_handler_cond;
254 extern afs_kcondvar_t rx_listener_cond;
256 static afs_kmutex_t epoch_mutex;
257 static afs_kmutex_t rx_init_mutex;
258 static afs_kmutex_t rx_debug_mutex;
259 static afs_kmutex_t rx_rpc_stats;
262 rxi_InitPthread(void)
264 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
265 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
266 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
267 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
268 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
269 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
270 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
271 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
273 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
280 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
281 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
283 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
284 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
286 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
287 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
288 #ifdef RX_ENABLE_LOCKS
291 #endif /* RX_LOCKS_DB */
292 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
293 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
295 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
297 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
299 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
301 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
302 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
303 #endif /* RX_ENABLE_LOCKS */
306 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
307 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
309 * The rx_stats_mutex mutex protects the following global variables:
310 * rxi_lowConnRefCount
311 * rxi_lowPeerRefCount
320 * The rx_quota_mutex mutex protects the following global variables:
328 * The rx_freePktQ_lock protects the following global variables:
333 * The rx_packets_mutex mutex protects the following global variables:
341 * The rx_pthread_mutex mutex protects the following global variables:
342 * rxi_fcfs_thread_num
345 #define INIT_PTHREAD_LOCKS
349 /* Variables for handling the minProcs implementation. availProcs gives the
350 * number of threads available in the pool at this moment (not counting dudes
351 * executing right now). totalMin gives the total number of procs required
352 * for handling all minProcs requests. minDeficit is a dynamic variable
353 * tracking the # of procs required to satisfy all of the remaining minProcs
355 * For fine grain locking to work, the quota check and the reservation of
356 * a server thread has to come while rxi_availProcs and rxi_minDeficit
357 * are locked. To this end, the code has been modified under #ifdef
358 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
359 * same time. A new function, ReturnToServerPool() returns the allocation.
361 * A call can be on several queue's (but only one at a time). When
362 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
363 * that no one else is touching the queue. To this end, we store the address
364 * of the queue lock in the call structure (under the call lock) when we
365 * put the call on a queue, and we clear the call_queue_lock when the
366 * call is removed from a queue (once the call lock has been obtained).
367 * This allows rxi_ResetCall to safely synchronize with others wishing
368 * to manipulate the queue.
371 #if defined(RX_ENABLE_LOCKS)
372 static afs_kmutex_t rx_rpc_stats;
375 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
376 ** pretty good that the next packet coming in is from the same connection
377 ** as the last packet, since we're send multiple packets in a transmit window.
379 struct rx_connection *rxLastConn = 0;
381 #ifdef RX_ENABLE_LOCKS
382 /* The locking hierarchy for rx fine grain locking is composed of these
385 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
386 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
387 * call->lock - locks call data fields.
388 * These are independent of each other:
389 * rx_freeCallQueue_lock
394 * serverQueueEntry->lock
395 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
397 * peer->lock - locks peer data fields.
398 * conn_data_lock - that more than one thread is not updating a conn data
399 * field at the same time.
410 * Do we need a lock to protect the peer field in the conn structure?
411 * conn->peer was previously a constant for all intents and so has no
412 * lock protecting this field. The multihomed client delta introduced
413 * a RX code change : change the peer field in the connection structure
414 * to that remote interface from which the last packet for this
415 * connection was sent out. This may become an issue if further changes
418 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
419 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
421 /* rxdb_fileID is used to identify the lock location, along with line#. */
422 static int rxdb_fileID = RXDB_FILE_RX;
423 #endif /* RX_LOCKS_DB */
424 #else /* RX_ENABLE_LOCKS */
425 #define SET_CALL_QUEUE_LOCK(C, L)
426 #define CLEAR_CALL_QUEUE_LOCK(C)
427 #endif /* RX_ENABLE_LOCKS */
428 struct rx_serverQueueEntry *rx_waitForPacket = 0;
429 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
431 /* ------------Exported Interfaces------------- */
433 /* This function allows rxkad to set the epoch to a suitably random number
434 * which rx_NewConnection will use in the future. The principle purpose is to
435 * get rxnull connections to use the same epoch as the rxkad connections do, at
436 * least once the first rxkad connection is established. This is important now
437 * that the host/port addresses aren't used in FindConnection: the uniqueness
438 * of epoch/cid matters and the start time won't do. */
440 #ifdef AFS_PTHREAD_ENV
442 * This mutex protects the following global variables:
446 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
447 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
451 #endif /* AFS_PTHREAD_ENV */
454 rx_SetEpoch(afs_uint32 epoch)
461 /* Initialize rx. A port number may be mentioned, in which case this
462 * becomes the default port number for any service installed later.
463 * If 0 is provided for the port number, a random port will be chosen
464 * by the kernel. Whether this will ever overlap anything in
465 * /etc/services is anybody's guess... Returns 0 on success, -1 on
470 int rxinit_status = 1;
471 #ifdef AFS_PTHREAD_ENV
473 * This mutex protects the following global variables:
477 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
478 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
481 #define UNLOCK_RX_INIT
485 rx_InitHost(u_int host, u_int port)
492 char *htable, *ptable;
499 if (rxinit_status == 0) {
500 tmp_status = rxinit_status;
502 return tmp_status; /* Already started; return previous error code. */
508 if (afs_winsockInit() < 0)
514 * Initialize anything necessary to provide a non-premptive threading
517 rxi_InitializeThreadSupport();
520 /* Allocate and initialize a socket for client and perhaps server
523 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
524 if (rx_socket == OSI_NULLSOCKET) {
528 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
531 #endif /* RX_LOCKS_DB */
532 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
533 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
534 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
535 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
536 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
537 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
538 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
539 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
540 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
541 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
543 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
545 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
547 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
549 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
550 #if defined(AFS_HPUX110_ENV)
552 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
553 #endif /* AFS_HPUX110_ENV */
554 #endif /* RX_ENABLE_LOCKS && KERNEL */
557 rx_connDeadTime = 12;
558 rx_tranquil = 0; /* reset flag */
559 rxi_ResetStatistics();
560 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
561 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
562 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
563 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
564 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
565 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
567 /* Malloc up a bunch of packets & buffers */
569 queue_Init(&rx_freePacketQueue);
570 rxi_NeedMorePackets = FALSE;
571 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
573 /* enforce a minimum number of allocated packets */
574 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
575 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
577 /* allocate the initial free packet pool */
578 #ifdef RX_ENABLE_TSFPQ
579 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
580 #else /* RX_ENABLE_TSFPQ */
581 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
582 #endif /* RX_ENABLE_TSFPQ */
589 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
590 tv.tv_sec = clock_now.sec;
591 tv.tv_usec = clock_now.usec;
592 srand((unsigned int)tv.tv_usec);
599 #if defined(KERNEL) && !defined(UKERNEL)
600 /* Really, this should never happen in a real kernel */
603 struct sockaddr_in addr;
605 int addrlen = sizeof(addr);
607 socklen_t addrlen = sizeof(addr);
609 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
611 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
614 rx_port = addr.sin_port;
617 rx_stats.minRtt.sec = 9999999;
619 rx_SetEpoch(tv.tv_sec | 0x80000000);
621 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
622 * will provide a randomer value. */
624 MUTEX_ENTER(&rx_quota_mutex);
625 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
626 MUTEX_EXIT(&rx_quota_mutex);
627 /* *Slightly* random start time for the cid. This is just to help
628 * out with the hashing function at the peer */
629 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
630 rx_connHashTable = (struct rx_connection **)htable;
631 rx_peerHashTable = (struct rx_peer **)ptable;
633 rx_hardAckDelay.sec = 0;
634 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
636 rxevent_Init(20, rxi_ReScheduleEvents);
638 /* Initialize various global queues */
639 queue_Init(&rx_idleServerQueue);
640 queue_Init(&rx_incomingCallQueue);
641 queue_Init(&rx_freeCallQueue);
643 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
644 /* Initialize our list of usable IP addresses. */
648 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
649 /* Start listener process (exact function is dependent on the
650 * implementation environment--kernel or user space) */
655 tmp_status = rxinit_status = 0;
663 return rx_InitHost(htonl(INADDR_ANY), port);
669 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
670 * maintaing the round trip timer.
675 * Start a new RTT timer for a given call and packet.
677 * There must be no resendEvent already listed for this call, otherwise this
678 * will leak events - intended for internal use within the RTO code only
681 * the RX call to start the timer for
682 * @param[in] lastPacket
683 * a flag indicating whether the last packet has been sent or not
685 * @pre call must be locked before calling this function
689 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
691 struct clock now, retryTime;
696 clock_Add(&retryTime, &call->rto);
698 /* If we're sending the last packet, and we're the client, then the server
699 * may wait for an additional 400ms before returning the ACK, wait for it
700 * rather than hitting a timeout */
701 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
702 clock_Addmsec(&retryTime, 400);
704 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
705 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
710 * Cancel an RTT timer for a given call.
714 * the RX call to cancel the timer for
716 * @pre call must be locked before calling this function
721 rxi_rto_cancel(struct rx_call *call)
723 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
727 * Tell the RTO timer that we have sent a packet.
729 * If the timer isn't already running, then start it. If the timer is running,
733 * the RX call that the packet has been sent on
734 * @param[in] lastPacket
735 * A flag which is true if this is the last packet for the call
737 * @pre The call must be locked before calling this function
742 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
744 if (call->resendEvent)
747 rxi_rto_startTimer(call, lastPacket, istack);
751 * Tell the RTO timer that we have received an new ACK message
753 * This function should be called whenever a call receives an ACK that
754 * acknowledges new packets. Whatever happens, we stop the current timer.
755 * If there are unacked packets in the queue which have been sent, then
756 * we restart the timer from now. Otherwise, we leave it stopped.
759 * the RX call that the ACK has been received on
763 rxi_rto_packet_acked(struct rx_call *call, int istack)
765 struct rx_packet *p, *nxp;
767 rxi_rto_cancel(call);
769 if (queue_IsEmpty(&call->tq))
772 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
773 if (p->header.seq > call->tfirst + call->twind)
776 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
777 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
785 * Set an initial round trip timeout for a peer connection
787 * @param[in] secs The timeout to set in seconds
791 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
792 peer->rtt = secs * 8000;
796 * Enables or disables the busy call channel error (RX_CALL_BUSY).
798 * @param[in] onoff Non-zero to enable busy call channel errors.
800 * @pre Neither rx_Init nor rx_InitHost have been called yet
803 rx_SetBusyChannelError(afs_int32 onoff)
805 osi_Assert(rxinit_status != 0);
806 rxi_busyChannelError = onoff ? 1 : 0;
810 * Set a delayed ack event on the specified call for the given time
812 * @param[in] call - the call on which to set the event
813 * @param[in] offset - the delay from now after which the event fires
816 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
818 struct clock now, when;
822 clock_Add(&when, offset);
824 if (!call->delayedAckEvent
825 || clock_Gt(&call->delayedAckTime, &when)) {
827 rxevent_Cancel(&call->delayedAckEvent, call,
828 RX_CALL_REFCOUNT_DELAY);
829 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
831 call->delayedAckEvent = rxevent_Post(&when, &now,
834 call->delayedAckTime = when;
838 /* called with unincremented nRequestsRunning to see if it is OK to start
839 * a new thread in this service. Could be "no" for two reasons: over the
840 * max quota, or would prevent others from reaching their min quota.
842 #ifdef RX_ENABLE_LOCKS
843 /* This verion of QuotaOK reserves quota if it's ok while the
844 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
847 QuotaOK(struct rx_service *aservice)
849 /* check if over max quota */
850 if (aservice->nRequestsRunning >= aservice->maxProcs) {
854 /* under min quota, we're OK */
855 /* otherwise, can use only if there are enough to allow everyone
856 * to go to their min quota after this guy starts.
859 MUTEX_ENTER(&rx_quota_mutex);
860 if ((aservice->nRequestsRunning < aservice->minProcs)
861 || (rxi_availProcs > rxi_minDeficit)) {
862 aservice->nRequestsRunning++;
863 /* just started call in minProcs pool, need fewer to maintain
865 if (aservice->nRequestsRunning <= aservice->minProcs)
868 MUTEX_EXIT(&rx_quota_mutex);
871 MUTEX_EXIT(&rx_quota_mutex);
877 ReturnToServerPool(struct rx_service *aservice)
879 aservice->nRequestsRunning--;
880 MUTEX_ENTER(&rx_quota_mutex);
881 if (aservice->nRequestsRunning < aservice->minProcs)
884 MUTEX_EXIT(&rx_quota_mutex);
887 #else /* RX_ENABLE_LOCKS */
889 QuotaOK(struct rx_service *aservice)
892 /* under min quota, we're OK */
893 if (aservice->nRequestsRunning < aservice->minProcs)
896 /* check if over max quota */
897 if (aservice->nRequestsRunning >= aservice->maxProcs)
900 /* otherwise, can use only if there are enough to allow everyone
901 * to go to their min quota after this guy starts.
903 MUTEX_ENTER(&rx_quota_mutex);
904 if (rxi_availProcs > rxi_minDeficit)
906 MUTEX_EXIT(&rx_quota_mutex);
909 #endif /* RX_ENABLE_LOCKS */
912 /* Called by rx_StartServer to start up lwp's to service calls.
913 NExistingProcs gives the number of procs already existing, and which
914 therefore needn't be created. */
916 rxi_StartServerProcs(int nExistingProcs)
918 struct rx_service *service;
923 /* For each service, reserve N processes, where N is the "minimum"
924 * number of processes that MUST be able to execute a request in parallel,
925 * at any time, for that process. Also compute the maximum difference
926 * between any service's maximum number of processes that can run
927 * (i.e. the maximum number that ever will be run, and a guarantee
928 * that this number will run if other services aren't running), and its
929 * minimum number. The result is the extra number of processes that
930 * we need in order to provide the latter guarantee */
931 for (i = 0; i < RX_MAX_SERVICES; i++) {
933 service = rx_services[i];
934 if (service == (struct rx_service *)0)
936 nProcs += service->minProcs;
937 diff = service->maxProcs - service->minProcs;
941 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
942 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
943 for (i = 0; i < nProcs; i++) {
944 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
950 /* This routine is only required on Windows */
952 rx_StartClientThread(void)
954 #ifdef AFS_PTHREAD_ENV
956 pid = pthread_self();
957 #endif /* AFS_PTHREAD_ENV */
959 #endif /* AFS_NT40_ENV */
961 /* This routine must be called if any services are exported. If the
962 * donateMe flag is set, the calling process is donated to the server
965 rx_StartServer(int donateMe)
967 struct rx_service *service;
973 /* Start server processes, if necessary (exact function is dependent
974 * on the implementation environment--kernel or user space). DonateMe
975 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
976 * case, one less new proc will be created rx_StartServerProcs.
978 rxi_StartServerProcs(donateMe);
980 /* count up the # of threads in minProcs, and add set the min deficit to
981 * be that value, too.
983 for (i = 0; i < RX_MAX_SERVICES; i++) {
984 service = rx_services[i];
985 if (service == (struct rx_service *)0)
987 MUTEX_ENTER(&rx_quota_mutex);
988 rxi_totalMin += service->minProcs;
989 /* below works even if a thread is running, since minDeficit would
990 * still have been decremented and later re-incremented.
992 rxi_minDeficit += service->minProcs;
993 MUTEX_EXIT(&rx_quota_mutex);
996 /* Turn on reaping of idle server connections */
997 rxi_ReapConnections(NULL, NULL, NULL, 0);
1002 #ifndef AFS_NT40_ENV
1006 #ifdef AFS_PTHREAD_ENV
1008 pid = afs_pointer_to_int(pthread_self());
1009 #else /* AFS_PTHREAD_ENV */
1011 LWP_CurrentProcess(&pid);
1012 #endif /* AFS_PTHREAD_ENV */
1014 sprintf(name, "srv_%d", ++nProcs);
1015 if (registerProgram)
1016 (*registerProgram) (pid, name);
1018 #endif /* AFS_NT40_ENV */
1019 rx_ServerProc(NULL); /* Never returns */
1021 #ifdef RX_ENABLE_TSFPQ
1022 /* no use leaving packets around in this thread's local queue if
1023 * it isn't getting donated to the server thread pool.
1025 rxi_FlushLocalPacketsTSFPQ();
1026 #endif /* RX_ENABLE_TSFPQ */
1030 /* Create a new client connection to the specified service, using the
1031 * specified security object to implement the security model for this
1033 struct rx_connection *
1034 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1035 struct rx_securityClass *securityObject,
1036 int serviceSecurityIndex)
1040 struct rx_connection *conn;
1045 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1046 "serviceSecurityIndex %d)\n",
1047 ntohl(shost), ntohs(sport), sservice, securityObject,
1048 serviceSecurityIndex));
1050 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1051 * the case of kmem_alloc? */
1052 conn = rxi_AllocConnection();
1053 #ifdef RX_ENABLE_LOCKS
1054 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1055 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1056 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1059 MUTEX_ENTER(&rx_connHashTable_lock);
1060 cid = (rx_nextCid += RX_MAXCALLS);
1061 conn->type = RX_CLIENT_CONNECTION;
1063 conn->epoch = rx_epoch;
1064 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1065 conn->serviceId = sservice;
1066 conn->securityObject = securityObject;
1067 conn->securityData = (void *) 0;
1068 conn->securityIndex = serviceSecurityIndex;
1069 rx_SetConnDeadTime(conn, rx_connDeadTime);
1070 rx_SetConnSecondsUntilNatPing(conn, 0);
1071 conn->ackRate = RX_FAST_ACK_RATE;
1072 conn->nSpecific = 0;
1073 conn->specific = NULL;
1074 conn->challengeEvent = NULL;
1075 conn->delayedAbortEvent = NULL;
1076 conn->abortCount = 0;
1078 for (i = 0; i < RX_MAXCALLS; i++) {
1079 conn->twind[i] = rx_initSendWindow;
1080 conn->rwind[i] = rx_initReceiveWindow;
1081 conn->lastBusy[i] = 0;
1084 RXS_NewConnection(securityObject, conn);
1086 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1088 conn->refCount++; /* no lock required since only this thread knows... */
1089 conn->next = rx_connHashTable[hashindex];
1090 rx_connHashTable[hashindex] = conn;
1091 if (rx_stats_active)
1092 rx_atomic_inc(&rx_stats.nClientConns);
1093 MUTEX_EXIT(&rx_connHashTable_lock);
1099 * Ensure a connection's timeout values are valid.
1101 * @param[in] conn The connection to check
1103 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1104 * unless idleDeadTime and/or hardDeadTime are not set
1108 rxi_CheckConnTimeouts(struct rx_connection *conn)
1110 /* a connection's timeouts must have the relationship
1111 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1112 * total loss of network to a peer may cause an idle timeout instead of a
1113 * dead timeout, simply because the idle timeout gets hit first. Also set
1114 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1115 /* this logic is slightly complicated by the fact that
1116 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1118 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1119 if (conn->idleDeadTime) {
1120 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1122 if (conn->hardDeadTime) {
1123 if (conn->idleDeadTime) {
1124 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1126 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1132 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1134 /* The idea is to set the dead time to a value that allows several
1135 * keepalives to be dropped without timing out the connection. */
1136 conn->secondsUntilDead = seconds;
1137 rxi_CheckConnTimeouts(conn);
1138 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1142 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1144 conn->hardDeadTime = seconds;
1145 rxi_CheckConnTimeouts(conn);
1149 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1151 conn->idleDeadTime = seconds;
1152 conn->idleDeadDetection = (seconds ? 1 : 0);
1153 rxi_CheckConnTimeouts(conn);
1156 int rxi_lowPeerRefCount = 0;
1157 int rxi_lowConnRefCount = 0;
1160 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1161 * NOTE: must not be called with rx_connHashTable_lock held.
1164 rxi_CleanupConnection(struct rx_connection *conn)
1166 /* Notify the service exporter, if requested, that this connection
1167 * is being destroyed */
1168 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1169 (*conn->service->destroyConnProc) (conn);
1171 /* Notify the security module that this connection is being destroyed */
1172 RXS_DestroyConnection(conn->securityObject, conn);
1174 /* If this is the last connection using the rx_peer struct, set its
1175 * idle time to now. rxi_ReapConnections will reap it if it's still
1176 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1178 MUTEX_ENTER(&rx_peerHashTable_lock);
1179 if (conn->peer->refCount < 2) {
1180 conn->peer->idleWhen = clock_Sec();
1181 if (conn->peer->refCount < 1) {
1182 conn->peer->refCount = 1;
1183 if (rx_stats_active) {
1184 MUTEX_ENTER(&rx_stats_mutex);
1185 rxi_lowPeerRefCount++;
1186 MUTEX_EXIT(&rx_stats_mutex);
1190 conn->peer->refCount--;
1191 MUTEX_EXIT(&rx_peerHashTable_lock);
1193 if (rx_stats_active)
1195 if (conn->type == RX_SERVER_CONNECTION)
1196 rx_atomic_dec(&rx_stats.nServerConns);
1198 rx_atomic_dec(&rx_stats.nClientConns);
1201 if (conn->specific) {
1203 for (i = 0; i < conn->nSpecific; i++) {
1204 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1205 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1206 conn->specific[i] = NULL;
1208 free(conn->specific);
1210 conn->specific = NULL;
1211 conn->nSpecific = 0;
1212 #endif /* !KERNEL */
1214 MUTEX_DESTROY(&conn->conn_call_lock);
1215 MUTEX_DESTROY(&conn->conn_data_lock);
1216 CV_DESTROY(&conn->conn_call_cv);
1218 rxi_FreeConnection(conn);
1221 /* Destroy the specified connection */
1223 rxi_DestroyConnection(struct rx_connection *conn)
1225 MUTEX_ENTER(&rx_connHashTable_lock);
1226 rxi_DestroyConnectionNoLock(conn);
1227 /* conn should be at the head of the cleanup list */
1228 if (conn == rx_connCleanup_list) {
1229 rx_connCleanup_list = rx_connCleanup_list->next;
1230 MUTEX_EXIT(&rx_connHashTable_lock);
1231 rxi_CleanupConnection(conn);
1233 #ifdef RX_ENABLE_LOCKS
1235 MUTEX_EXIT(&rx_connHashTable_lock);
1237 #endif /* RX_ENABLE_LOCKS */
1241 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1243 struct rx_connection **conn_ptr;
1245 struct rx_packet *packet;
1252 MUTEX_ENTER(&conn->conn_data_lock);
1253 MUTEX_ENTER(&rx_refcnt_mutex);
1254 if (conn->refCount > 0)
1257 if (rx_stats_active) {
1258 MUTEX_ENTER(&rx_stats_mutex);
1259 rxi_lowConnRefCount++;
1260 MUTEX_EXIT(&rx_stats_mutex);
1264 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1265 /* Busy; wait till the last guy before proceeding */
1266 MUTEX_EXIT(&rx_refcnt_mutex);
1267 MUTEX_EXIT(&conn->conn_data_lock);
1272 /* If the client previously called rx_NewCall, but it is still
1273 * waiting, treat this as a running call, and wait to destroy the
1274 * connection later when the call completes. */
1275 if ((conn->type == RX_CLIENT_CONNECTION)
1276 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1277 conn->flags |= RX_CONN_DESTROY_ME;
1278 MUTEX_EXIT(&conn->conn_data_lock);
1282 MUTEX_EXIT(&rx_refcnt_mutex);
1283 MUTEX_EXIT(&conn->conn_data_lock);
1285 /* Check for extant references to this connection */
1286 MUTEX_ENTER(&conn->conn_call_lock);
1287 for (i = 0; i < RX_MAXCALLS; i++) {
1288 struct rx_call *call = conn->call[i];
1291 if (conn->type == RX_CLIENT_CONNECTION) {
1292 MUTEX_ENTER(&call->lock);
1293 if (call->delayedAckEvent) {
1294 /* Push the final acknowledgment out now--there
1295 * won't be a subsequent call to acknowledge the
1296 * last reply packets */
1297 rxevent_Cancel(&call->delayedAckEvent, call,
1298 RX_CALL_REFCOUNT_DELAY);
1299 if (call->state == RX_STATE_PRECALL
1300 || call->state == RX_STATE_ACTIVE) {
1301 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1306 MUTEX_EXIT(&call->lock);
1310 MUTEX_EXIT(&conn->conn_call_lock);
1312 #ifdef RX_ENABLE_LOCKS
1314 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1315 MUTEX_EXIT(&conn->conn_data_lock);
1317 /* Someone is accessing a packet right now. */
1321 #endif /* RX_ENABLE_LOCKS */
1324 /* Don't destroy the connection if there are any call
1325 * structures still in use */
1326 MUTEX_ENTER(&conn->conn_data_lock);
1327 conn->flags |= RX_CONN_DESTROY_ME;
1328 MUTEX_EXIT(&conn->conn_data_lock);
1333 if (conn->natKeepAliveEvent) {
1334 rxi_NatKeepAliveOff(conn);
1337 if (conn->delayedAbortEvent) {
1338 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1339 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1341 MUTEX_ENTER(&conn->conn_data_lock);
1342 rxi_SendConnectionAbort(conn, packet, 0, 1);
1343 MUTEX_EXIT(&conn->conn_data_lock);
1344 rxi_FreePacket(packet);
1348 /* Remove from connection hash table before proceeding */
1350 &rx_connHashTable[CONN_HASH
1351 (peer->host, peer->port, conn->cid, conn->epoch,
1353 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1354 if (*conn_ptr == conn) {
1355 *conn_ptr = conn->next;
1359 /* if the conn that we are destroying was the last connection, then we
1360 * clear rxLastConn as well */
1361 if (rxLastConn == conn)
1364 /* Make sure the connection is completely reset before deleting it. */
1365 /* get rid of pending events that could zap us later */
1366 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1367 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1368 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1370 /* Add the connection to the list of destroyed connections that
1371 * need to be cleaned up. This is necessary to avoid deadlocks
1372 * in the routines we call to inform others that this connection is
1373 * being destroyed. */
1374 conn->next = rx_connCleanup_list;
1375 rx_connCleanup_list = conn;
1378 /* Externally available version */
1380 rx_DestroyConnection(struct rx_connection *conn)
1385 rxi_DestroyConnection(conn);
1390 rx_GetConnection(struct rx_connection *conn)
1395 MUTEX_ENTER(&rx_refcnt_mutex);
1397 MUTEX_EXIT(&rx_refcnt_mutex);
1401 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1402 /* Wait for the transmit queue to no longer be busy.
1403 * requires the call->lock to be held */
1405 rxi_WaitforTQBusy(struct rx_call *call) {
1406 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1407 call->flags |= RX_CALL_TQ_WAIT;
1409 #ifdef RX_ENABLE_LOCKS
1410 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1411 CV_WAIT(&call->cv_tq, &call->lock);
1412 #else /* RX_ENABLE_LOCKS */
1413 osi_rxSleep(&call->tq);
1414 #endif /* RX_ENABLE_LOCKS */
1416 if (call->tqWaiters == 0) {
1417 call->flags &= ~RX_CALL_TQ_WAIT;
1424 rxi_WakeUpTransmitQueue(struct rx_call *call)
1426 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1427 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1428 call, call->tqWaiters, call->flags));
1429 #ifdef RX_ENABLE_LOCKS
1430 osirx_AssertMine(&call->lock, "rxi_Start start");
1431 CV_BROADCAST(&call->cv_tq);
1432 #else /* RX_ENABLE_LOCKS */
1433 osi_rxWakeup(&call->tq);
1434 #endif /* RX_ENABLE_LOCKS */
1438 /* Start a new rx remote procedure call, on the specified connection.
1439 * If wait is set to 1, wait for a free call channel; otherwise return
1440 * 0. Maxtime gives the maximum number of seconds this call may take,
1441 * after rx_NewCall returns. After this time interval, a call to any
1442 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1443 * For fine grain locking, we hold the conn_call_lock in order to
1444 * to ensure that we don't get signalle after we found a call in an active
1445 * state and before we go to sleep.
1448 rx_NewCall(struct rx_connection *conn)
1450 int i, wait, ignoreBusy = 1;
1451 struct rx_call *call;
1452 struct clock queueTime;
1453 afs_uint32 leastBusy = 0;
1457 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1460 clock_GetTime(&queueTime);
1462 * Check if there are others waiting for a new call.
1463 * If so, let them go first to avoid starving them.
1464 * This is a fairly simple scheme, and might not be
1465 * a complete solution for large numbers of waiters.
1467 * makeCallWaiters keeps track of the number of
1468 * threads waiting to make calls and the
1469 * RX_CONN_MAKECALL_WAITING flag bit is used to
1470 * indicate that there are indeed calls waiting.
1471 * The flag is set when the waiter is incremented.
1472 * It is only cleared when makeCallWaiters is 0.
1473 * This prevents us from accidently destroying the
1474 * connection while it is potentially about to be used.
1476 MUTEX_ENTER(&conn->conn_call_lock);
1477 MUTEX_ENTER(&conn->conn_data_lock);
1478 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1479 conn->flags |= RX_CONN_MAKECALL_WAITING;
1480 conn->makeCallWaiters++;
1481 MUTEX_EXIT(&conn->conn_data_lock);
1483 #ifdef RX_ENABLE_LOCKS
1484 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1488 MUTEX_ENTER(&conn->conn_data_lock);
1489 conn->makeCallWaiters--;
1490 if (conn->makeCallWaiters == 0)
1491 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1494 /* We are now the active thread in rx_NewCall */
1495 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1496 MUTEX_EXIT(&conn->conn_data_lock);
1501 for (i = 0; i < RX_MAXCALLS; i++) {
1502 call = conn->call[i];
1504 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1505 /* we're not ignoring busy call slots; only look at the
1506 * call slot that is the "least" busy */
1510 if (call->state == RX_STATE_DALLY) {
1511 MUTEX_ENTER(&call->lock);
1512 if (call->state == RX_STATE_DALLY) {
1513 if (ignoreBusy && conn->lastBusy[i]) {
1514 /* if we're ignoring busy call slots, skip any ones that
1515 * have lastBusy set */
1516 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1517 leastBusy = conn->lastBusy[i];
1519 MUTEX_EXIT(&call->lock);
1524 * We are setting the state to RX_STATE_RESET to
1525 * ensure that no one else will attempt to use this
1526 * call once we drop the conn->conn_call_lock and
1527 * call->lock. We must drop the conn->conn_call_lock
1528 * before calling rxi_ResetCall because the process
1529 * of clearing the transmit queue can block for an
1530 * extended period of time. If we block while holding
1531 * the conn->conn_call_lock, then all rx_EndCall
1532 * processing will block as well. This has a detrimental
1533 * effect on overall system performance.
1535 call->state = RX_STATE_RESET;
1536 (*call->callNumber)++;
1537 MUTEX_EXIT(&conn->conn_call_lock);
1538 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1539 rxi_ResetCall(call, 0);
1540 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1544 * If we failed to be able to safely obtain the
1545 * conn->conn_call_lock we will have to drop the
1546 * call->lock to avoid a deadlock. When the call->lock
1547 * is released the state of the call can change. If it
1548 * is no longer RX_STATE_RESET then some other thread is
1551 MUTEX_EXIT(&call->lock);
1552 MUTEX_ENTER(&conn->conn_call_lock);
1553 MUTEX_ENTER(&call->lock);
1555 if (call->state == RX_STATE_RESET)
1559 * If we get here it means that after dropping
1560 * the conn->conn_call_lock and call->lock that
1561 * the call is no longer ours. If we can't find
1562 * a free call in the remaining slots we should
1563 * not go immediately to RX_CONN_MAKECALL_WAITING
1564 * because by dropping the conn->conn_call_lock
1565 * we have given up synchronization with rx_EndCall.
1566 * Instead, cycle through one more time to see if
1567 * we can find a call that can call our own.
1569 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1572 MUTEX_EXIT(&call->lock);
1575 if (ignoreBusy && conn->lastBusy[i]) {
1576 /* if we're ignoring busy call slots, skip any ones that
1577 * have lastBusy set */
1578 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1579 leastBusy = conn->lastBusy[i];
1584 /* rxi_NewCall returns with mutex locked */
1585 call = rxi_NewCall(conn, i);
1586 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1590 if (i < RX_MAXCALLS) {
1591 conn->lastBusy[i] = 0;
1592 call->flags &= ~RX_CALL_PEER_BUSY;
1597 if (leastBusy && ignoreBusy) {
1598 /* we didn't find a useable call slot, but we did see at least one
1599 * 'busy' slot; look again and only use a slot with the 'least
1605 MUTEX_ENTER(&conn->conn_data_lock);
1606 conn->flags |= RX_CONN_MAKECALL_WAITING;
1607 conn->makeCallWaiters++;
1608 MUTEX_EXIT(&conn->conn_data_lock);
1610 #ifdef RX_ENABLE_LOCKS
1611 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1615 MUTEX_ENTER(&conn->conn_data_lock);
1616 conn->makeCallWaiters--;
1617 if (conn->makeCallWaiters == 0)
1618 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1619 MUTEX_EXIT(&conn->conn_data_lock);
1621 /* Client is initially in send mode */
1622 call->state = RX_STATE_ACTIVE;
1623 call->error = conn->error;
1625 call->mode = RX_MODE_ERROR;
1627 call->mode = RX_MODE_SENDING;
1629 /* remember start time for call in case we have hard dead time limit */
1630 call->queueTime = queueTime;
1631 clock_GetTime(&call->startTime);
1632 call->bytesSent = 0;
1633 call->bytesRcvd = 0;
1635 /* Turn on busy protocol. */
1636 rxi_KeepAliveOn(call);
1638 /* Attempt MTU discovery */
1639 rxi_GrowMTUOn(call);
1642 * We are no longer the active thread in rx_NewCall
1644 MUTEX_ENTER(&conn->conn_data_lock);
1645 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1646 MUTEX_EXIT(&conn->conn_data_lock);
1649 * Wake up anyone else who might be giving us a chance to
1650 * run (see code above that avoids resource starvation).
1652 #ifdef RX_ENABLE_LOCKS
1653 CV_BROADCAST(&conn->conn_call_cv);
1657 MUTEX_EXIT(&conn->conn_call_lock);
1659 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1660 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1661 osi_Panic("rx_NewCall call about to be used without an empty tq");
1663 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1665 MUTEX_EXIT(&call->lock);
1668 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1673 rxi_HasActiveCalls(struct rx_connection *aconn)
1676 struct rx_call *tcall;
1680 for (i = 0; i < RX_MAXCALLS; i++) {
1681 if ((tcall = aconn->call[i])) {
1682 if ((tcall->state == RX_STATE_ACTIVE)
1683 || (tcall->state == RX_STATE_PRECALL)) {
1694 rxi_GetCallNumberVector(struct rx_connection *aconn,
1695 afs_int32 * aint32s)
1698 struct rx_call *tcall;
1702 MUTEX_ENTER(&aconn->conn_call_lock);
1703 for (i = 0; i < RX_MAXCALLS; i++) {
1704 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1705 aint32s[i] = aconn->callNumber[i] + 1;
1707 aint32s[i] = aconn->callNumber[i];
1709 MUTEX_EXIT(&aconn->conn_call_lock);
1715 rxi_SetCallNumberVector(struct rx_connection *aconn,
1716 afs_int32 * aint32s)
1719 struct rx_call *tcall;
1723 MUTEX_ENTER(&aconn->conn_call_lock);
1724 for (i = 0; i < RX_MAXCALLS; i++) {
1725 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1726 aconn->callNumber[i] = aint32s[i] - 1;
1728 aconn->callNumber[i] = aint32s[i];
1730 MUTEX_EXIT(&aconn->conn_call_lock);
1735 /* Advertise a new service. A service is named locally by a UDP port
1736 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1739 char *serviceName; Name for identification purposes (e.g. the
1740 service name might be used for probing for
1743 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1744 char *serviceName, struct rx_securityClass **securityObjects,
1745 int nSecurityObjects,
1746 afs_int32(*serviceProc) (struct rx_call * acall))
1748 osi_socket socket = OSI_NULLSOCKET;
1749 struct rx_service *tservice;
1755 if (serviceId == 0) {
1757 "rx_NewService: service id for service %s is not non-zero.\n",
1764 "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",
1772 tservice = rxi_AllocService();
1775 #ifdef RX_ENABLE_LOCKS
1776 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1779 for (i = 0; i < RX_MAX_SERVICES; i++) {
1780 struct rx_service *service = rx_services[i];
1782 if (port == service->servicePort && host == service->serviceHost) {
1783 if (service->serviceId == serviceId) {
1784 /* The identical service has already been
1785 * installed; if the caller was intending to
1786 * change the security classes used by this
1787 * service, he/she loses. */
1789 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1790 serviceName, serviceId, service->serviceName);
1792 rxi_FreeService(tservice);
1795 /* Different service, same port: re-use the socket
1796 * which is bound to the same port */
1797 socket = service->socket;
1800 if (socket == OSI_NULLSOCKET) {
1801 /* If we don't already have a socket (from another
1802 * service on same port) get a new one */
1803 socket = rxi_GetHostUDPSocket(host, port);
1804 if (socket == OSI_NULLSOCKET) {
1806 rxi_FreeService(tservice);
1811 service->socket = socket;
1812 service->serviceHost = host;
1813 service->servicePort = port;
1814 service->serviceId = serviceId;
1815 service->serviceName = serviceName;
1816 service->nSecurityObjects = nSecurityObjects;
1817 service->securityObjects = securityObjects;
1818 service->minProcs = 0;
1819 service->maxProcs = 1;
1820 service->idleDeadTime = 60;
1821 service->idleDeadErr = 0;
1822 service->connDeadTime = rx_connDeadTime;
1823 service->executeRequestProc = serviceProc;
1824 service->checkReach = 0;
1825 service->nSpecific = 0;
1826 service->specific = NULL;
1827 rx_services[i] = service; /* not visible until now */
1833 rxi_FreeService(tservice);
1834 (osi_Msg "rx_NewService: cannot support > %d services\n",
1839 /* Set configuration options for all of a service's security objects */
1842 rx_SetSecurityConfiguration(struct rx_service *service,
1843 rx_securityConfigVariables type,
1847 for (i = 0; i<service->nSecurityObjects; i++) {
1848 if (service->securityObjects[i]) {
1849 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1857 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1858 struct rx_securityClass **securityObjects, int nSecurityObjects,
1859 afs_int32(*serviceProc) (struct rx_call * acall))
1861 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1864 /* Generic request processing loop. This routine should be called
1865 * by the implementation dependent rx_ServerProc. If socketp is
1866 * non-null, it will be set to the file descriptor that this thread
1867 * is now listening on. If socketp is null, this routine will never
1870 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1872 struct rx_call *call;
1874 struct rx_service *tservice = NULL;
1881 call = rx_GetCall(threadID, tservice, socketp);
1882 if (socketp && *socketp != OSI_NULLSOCKET) {
1883 /* We are now a listener thread */
1889 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1890 #ifdef RX_ENABLE_LOCKS
1892 #endif /* RX_ENABLE_LOCKS */
1893 afs_termState = AFSOP_STOP_AFS;
1894 afs_osi_Wakeup(&afs_termState);
1895 #ifdef RX_ENABLE_LOCKS
1897 #endif /* RX_ENABLE_LOCKS */
1902 /* if server is restarting( typically smooth shutdown) then do not
1903 * allow any new calls.
1906 if (rx_tranquil && (call != NULL)) {
1910 MUTEX_ENTER(&call->lock);
1912 rxi_CallError(call, RX_RESTARTING);
1913 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1915 MUTEX_EXIT(&call->lock);
1920 tservice = call->conn->service;
1922 if (tservice->beforeProc)
1923 (*tservice->beforeProc) (call);
1925 code = tservice->executeRequestProc(call);
1927 if (tservice->afterProc)
1928 (*tservice->afterProc) (call, code);
1930 rx_EndCall(call, code);
1932 if (tservice->postProc)
1933 (*tservice->postProc) (code);
1935 if (rx_stats_active) {
1936 MUTEX_ENTER(&rx_stats_mutex);
1938 MUTEX_EXIT(&rx_stats_mutex);
1945 rx_WakeupServerProcs(void)
1947 struct rx_serverQueueEntry *np, *tqp;
1951 MUTEX_ENTER(&rx_serverPool_lock);
1953 #ifdef RX_ENABLE_LOCKS
1954 if (rx_waitForPacket)
1955 CV_BROADCAST(&rx_waitForPacket->cv);
1956 #else /* RX_ENABLE_LOCKS */
1957 if (rx_waitForPacket)
1958 osi_rxWakeup(rx_waitForPacket);
1959 #endif /* RX_ENABLE_LOCKS */
1960 MUTEX_ENTER(&freeSQEList_lock);
1961 for (np = rx_FreeSQEList; np; np = tqp) {
1962 tqp = *(struct rx_serverQueueEntry **)np;
1963 #ifdef RX_ENABLE_LOCKS
1964 CV_BROADCAST(&np->cv);
1965 #else /* RX_ENABLE_LOCKS */
1967 #endif /* RX_ENABLE_LOCKS */
1969 MUTEX_EXIT(&freeSQEList_lock);
1970 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1971 #ifdef RX_ENABLE_LOCKS
1972 CV_BROADCAST(&np->cv);
1973 #else /* RX_ENABLE_LOCKS */
1975 #endif /* RX_ENABLE_LOCKS */
1977 MUTEX_EXIT(&rx_serverPool_lock);
1982 * One thing that seems to happen is that all the server threads get
1983 * tied up on some empty or slow call, and then a whole bunch of calls
1984 * arrive at once, using up the packet pool, so now there are more
1985 * empty calls. The most critical resources here are server threads
1986 * and the free packet pool. The "doreclaim" code seems to help in
1987 * general. I think that eventually we arrive in this state: there
1988 * are lots of pending calls which do have all their packets present,
1989 * so they won't be reclaimed, are multi-packet calls, so they won't
1990 * be scheduled until later, and thus are tying up most of the free
1991 * packet pool for a very long time.
1993 * 1. schedule multi-packet calls if all the packets are present.
1994 * Probably CPU-bound operation, useful to return packets to pool.
1995 * Do what if there is a full window, but the last packet isn't here?
1996 * 3. preserve one thread which *only* runs "best" calls, otherwise
1997 * it sleeps and waits for that type of call.
1998 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1999 * the current dataquota business is badly broken. The quota isn't adjusted
2000 * to reflect how many packets are presently queued for a running call.
2001 * So, when we schedule a queued call with a full window of packets queued
2002 * up for it, that *should* free up a window full of packets for other 2d-class
2003 * calls to be able to use from the packet pool. But it doesn't.
2005 * NB. Most of the time, this code doesn't run -- since idle server threads
2006 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2007 * as a new call arrives.
2009 /* Sleep until a call arrives. Returns a pointer to the call, ready
2010 * for an rx_Read. */
2011 #ifdef RX_ENABLE_LOCKS
2013 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2015 struct rx_serverQueueEntry *sq;
2016 struct rx_call *call = (struct rx_call *)0;
2017 struct rx_service *service = NULL;
2019 MUTEX_ENTER(&freeSQEList_lock);
2021 if ((sq = rx_FreeSQEList)) {
2022 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2023 MUTEX_EXIT(&freeSQEList_lock);
2024 } else { /* otherwise allocate a new one and return that */
2025 MUTEX_EXIT(&freeSQEList_lock);
2026 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2027 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2028 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2031 MUTEX_ENTER(&rx_serverPool_lock);
2032 if (cur_service != NULL) {
2033 ReturnToServerPool(cur_service);
2036 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2037 struct rx_call *tcall, *ncall, *choice2 = NULL;
2039 /* Scan for eligible incoming calls. A call is not eligible
2040 * if the maximum number of calls for its service type are
2041 * already executing */
2042 /* One thread will process calls FCFS (to prevent starvation),
2043 * while the other threads may run ahead looking for calls which
2044 * have all their input data available immediately. This helps
2045 * keep threads from blocking, waiting for data from the client. */
2046 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2047 service = tcall->conn->service;
2048 if (!QuotaOK(service)) {
2051 MUTEX_ENTER(&rx_pthread_mutex);
2052 if (tno == rxi_fcfs_thread_num
2053 || queue_IsLast(&rx_incomingCallQueue, tcall)) {
2054 MUTEX_EXIT(&rx_pthread_mutex);
2055 /* If we're the fcfs thread , then we'll just use
2056 * this call. If we haven't been able to find an optimal
2057 * choice, and we're at the end of the list, then use a
2058 * 2d choice if one has been identified. Otherwise... */
2059 call = (choice2 ? choice2 : tcall);
2060 service = call->conn->service;
2062 MUTEX_EXIT(&rx_pthread_mutex);
2063 if (!queue_IsEmpty(&tcall->rq)) {
2064 struct rx_packet *rp;
2065 rp = queue_First(&tcall->rq, rx_packet);
2066 if (rp->header.seq == 1) {
2068 || (rp->header.flags & RX_LAST_PACKET)) {
2070 } else if (rxi_2dchoice && !choice2
2071 && !(tcall->flags & RX_CALL_CLEARED)
2072 && (tcall->rprev > rxi_HardAckRate)) {
2082 ReturnToServerPool(service);
2089 MUTEX_EXIT(&rx_serverPool_lock);
2090 MUTEX_ENTER(&call->lock);
2092 if (call->flags & RX_CALL_WAIT_PROC) {
2093 call->flags &= ~RX_CALL_WAIT_PROC;
2094 rx_atomic_dec(&rx_nWaiting);
2097 if (call->state != RX_STATE_PRECALL || call->error) {
2098 MUTEX_EXIT(&call->lock);
2099 MUTEX_ENTER(&rx_serverPool_lock);
2100 ReturnToServerPool(service);
2105 if (queue_IsEmpty(&call->rq)
2106 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2107 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2109 CLEAR_CALL_QUEUE_LOCK(call);
2112 /* If there are no eligible incoming calls, add this process
2113 * to the idle server queue, to wait for one */
2117 *socketp = OSI_NULLSOCKET;
2119 sq->socketp = socketp;
2120 queue_Append(&rx_idleServerQueue, sq);
2121 #ifndef AFS_AIX41_ENV
2122 rx_waitForPacket = sq;
2124 rx_waitingForPacket = sq;
2125 #endif /* AFS_AIX41_ENV */
2127 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2129 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2130 MUTEX_EXIT(&rx_serverPool_lock);
2131 return (struct rx_call *)0;
2134 } while (!(call = sq->newcall)
2135 && !(socketp && *socketp != OSI_NULLSOCKET));
2136 MUTEX_EXIT(&rx_serverPool_lock);
2138 MUTEX_ENTER(&call->lock);
2144 MUTEX_ENTER(&freeSQEList_lock);
2145 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2146 rx_FreeSQEList = sq;
2147 MUTEX_EXIT(&freeSQEList_lock);
2150 clock_GetTime(&call->startTime);
2151 call->state = RX_STATE_ACTIVE;
2152 call->mode = RX_MODE_RECEIVING;
2153 #ifdef RX_KERNEL_TRACE
2154 if (ICL_SETACTIVE(afs_iclSetp)) {
2155 int glockOwner = ISAFS_GLOCK();
2158 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2159 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2166 rxi_calltrace(RX_CALL_START, call);
2167 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2168 call->conn->service->servicePort, call->conn->service->serviceId,
2171 MUTEX_EXIT(&call->lock);
2172 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2174 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2179 #else /* RX_ENABLE_LOCKS */
2181 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2183 struct rx_serverQueueEntry *sq;
2184 struct rx_call *call = (struct rx_call *)0, *choice2;
2185 struct rx_service *service = NULL;
2189 MUTEX_ENTER(&freeSQEList_lock);
2191 if ((sq = rx_FreeSQEList)) {
2192 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2193 MUTEX_EXIT(&freeSQEList_lock);
2194 } else { /* otherwise allocate a new one and return that */
2195 MUTEX_EXIT(&freeSQEList_lock);
2196 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2197 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2198 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2200 MUTEX_ENTER(&sq->lock);
2202 if (cur_service != NULL) {
2203 cur_service->nRequestsRunning--;
2204 MUTEX_ENTER(&rx_quota_mutex);
2205 if (cur_service->nRequestsRunning < cur_service->minProcs)
2208 MUTEX_EXIT(&rx_quota_mutex);
2210 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2211 struct rx_call *tcall, *ncall;
2212 /* Scan for eligible incoming calls. A call is not eligible
2213 * if the maximum number of calls for its service type are
2214 * already executing */
2215 /* One thread will process calls FCFS (to prevent starvation),
2216 * while the other threads may run ahead looking for calls which
2217 * have all their input data available immediately. This helps
2218 * keep threads from blocking, waiting for data from the client. */
2219 choice2 = (struct rx_call *)0;
2220 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2221 service = tcall->conn->service;
2222 if (QuotaOK(service)) {
2223 MUTEX_ENTER(&rx_pthread_mutex);
2224 if (tno == rxi_fcfs_thread_num
2225 || !tcall->queue_item_header.next) {
2226 MUTEX_EXIT(&rx_pthread_mutex);
2227 /* If we're the fcfs thread, then we'll just use
2228 * this call. If we haven't been able to find an optimal
2229 * choice, and we're at the end of the list, then use a
2230 * 2d choice if one has been identified. Otherwise... */
2231 call = (choice2 ? choice2 : tcall);
2232 service = call->conn->service;
2234 MUTEX_EXIT(&rx_pthread_mutex);
2235 if (!queue_IsEmpty(&tcall->rq)) {
2236 struct rx_packet *rp;
2237 rp = queue_First(&tcall->rq, rx_packet);
2238 if (rp->header.seq == 1
2240 || (rp->header.flags & RX_LAST_PACKET))) {
2242 } else if (rxi_2dchoice && !choice2
2243 && !(tcall->flags & RX_CALL_CLEARED)
2244 && (tcall->rprev > rxi_HardAckRate)) {
2258 /* we can't schedule a call if there's no data!!! */
2259 /* send an ack if there's no data, if we're missing the
2260 * first packet, or we're missing something between first
2261 * and last -- there's a "hole" in the incoming data. */
2262 if (queue_IsEmpty(&call->rq)
2263 || queue_First(&call->rq, rx_packet)->header.seq != 1
2264 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2265 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2267 call->flags &= (~RX_CALL_WAIT_PROC);
2268 service->nRequestsRunning++;
2269 /* just started call in minProcs pool, need fewer to maintain
2271 MUTEX_ENTER(&rx_quota_mutex);
2272 if (service->nRequestsRunning <= service->minProcs)
2275 MUTEX_EXIT(&rx_quota_mutex);
2276 rx_atomic_dec(&rx_nWaiting);
2277 /* MUTEX_EXIT(&call->lock); */
2279 /* If there are no eligible incoming calls, add this process
2280 * to the idle server queue, to wait for one */
2283 *socketp = OSI_NULLSOCKET;
2285 sq->socketp = socketp;
2286 queue_Append(&rx_idleServerQueue, sq);
2290 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2292 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2293 return (struct rx_call *)0;
2296 } while (!(call = sq->newcall)
2297 && !(socketp && *socketp != OSI_NULLSOCKET));
2299 MUTEX_EXIT(&sq->lock);
2301 MUTEX_ENTER(&freeSQEList_lock);
2302 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2303 rx_FreeSQEList = sq;
2304 MUTEX_EXIT(&freeSQEList_lock);
2307 clock_GetTime(&call->startTime);
2308 call->state = RX_STATE_ACTIVE;
2309 call->mode = RX_MODE_RECEIVING;
2310 #ifdef RX_KERNEL_TRACE
2311 if (ICL_SETACTIVE(afs_iclSetp)) {
2312 int glockOwner = ISAFS_GLOCK();
2315 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2316 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2323 rxi_calltrace(RX_CALL_START, call);
2324 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2325 call->conn->service->servicePort, call->conn->service->serviceId,
2328 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2335 #endif /* RX_ENABLE_LOCKS */
2339 /* Establish a procedure to be called when a packet arrives for a
2340 * call. This routine will be called at most once after each call,
2341 * and will also be called if there is an error condition on the or
2342 * the call is complete. Used by multi rx to build a selection
2343 * function which determines which of several calls is likely to be a
2344 * good one to read from.
2345 * NOTE: the way this is currently implemented it is probably only a
2346 * good idea to (1) use it immediately after a newcall (clients only)
2347 * and (2) only use it once. Other uses currently void your warranty
2350 rx_SetArrivalProc(struct rx_call *call,
2351 void (*proc) (struct rx_call * call,
2354 void * handle, int arg)
2356 call->arrivalProc = proc;
2357 call->arrivalProcHandle = handle;
2358 call->arrivalProcArg = arg;
2361 /* Call is finished (possibly prematurely). Return rc to the peer, if
2362 * appropriate, and return the final error code from the conversation
2366 rx_EndCall(struct rx_call *call, afs_int32 rc)
2368 struct rx_connection *conn = call->conn;
2372 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2373 call, rc, call->error, call->abortCode));
2376 MUTEX_ENTER(&call->lock);
2378 if (rc == 0 && call->error == 0) {
2379 call->abortCode = 0;
2380 call->abortCount = 0;
2383 call->arrivalProc = (void (*)())0;
2384 if (rc && call->error == 0) {
2385 rxi_CallError(call, rc);
2386 call->mode = RX_MODE_ERROR;
2387 /* Send an abort message to the peer if this error code has
2388 * only just been set. If it was set previously, assume the
2389 * peer has already been sent the error code or will request it
2391 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2393 if (conn->type == RX_SERVER_CONNECTION) {
2394 /* Make sure reply or at least dummy reply is sent */
2395 if (call->mode == RX_MODE_RECEIVING) {
2396 MUTEX_EXIT(&call->lock);
2397 rxi_WriteProc(call, 0, 0);
2398 MUTEX_ENTER(&call->lock);
2400 if (call->mode == RX_MODE_SENDING) {
2401 MUTEX_EXIT(&call->lock);
2402 rxi_FlushWrite(call);
2403 MUTEX_ENTER(&call->lock);
2405 rxi_calltrace(RX_CALL_END, call);
2406 /* Call goes to hold state until reply packets are acknowledged */
2407 if (call->tfirst + call->nSoftAcked < call->tnext) {
2408 call->state = RX_STATE_HOLD;
2410 call->state = RX_STATE_DALLY;
2411 rxi_ClearTransmitQueue(call, 0);
2412 rxi_rto_cancel(call);
2413 rxevent_Cancel(&call->keepAliveEvent, call,
2414 RX_CALL_REFCOUNT_ALIVE);
2416 } else { /* Client connection */
2418 /* Make sure server receives input packets, in the case where
2419 * no reply arguments are expected */
2420 if ((call->mode == RX_MODE_SENDING)
2421 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2422 MUTEX_EXIT(&call->lock);
2423 (void)rxi_ReadProc(call, &dummy, 1);
2424 MUTEX_ENTER(&call->lock);
2427 /* If we had an outstanding delayed ack, be nice to the server
2428 * and force-send it now.
2430 if (call->delayedAckEvent) {
2431 rxevent_Cancel(&call->delayedAckEvent, call,
2432 RX_CALL_REFCOUNT_DELAY);
2433 rxi_SendDelayedAck(NULL, call, NULL, 0);
2436 /* We need to release the call lock since it's lower than the
2437 * conn_call_lock and we don't want to hold the conn_call_lock
2438 * over the rx_ReadProc call. The conn_call_lock needs to be held
2439 * here for the case where rx_NewCall is perusing the calls on
2440 * the connection structure. We don't want to signal until
2441 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2442 * have checked this call, found it active and by the time it
2443 * goes to sleep, will have missed the signal.
2445 MUTEX_EXIT(&call->lock);
2446 MUTEX_ENTER(&conn->conn_call_lock);
2447 MUTEX_ENTER(&call->lock);
2449 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2450 conn->lastBusy[call->channel] = 0;
2453 MUTEX_ENTER(&conn->conn_data_lock);
2454 conn->flags |= RX_CONN_BUSY;
2455 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2456 MUTEX_EXIT(&conn->conn_data_lock);
2457 #ifdef RX_ENABLE_LOCKS
2458 CV_BROADCAST(&conn->conn_call_cv);
2463 #ifdef RX_ENABLE_LOCKS
2465 MUTEX_EXIT(&conn->conn_data_lock);
2467 #endif /* RX_ENABLE_LOCKS */
2468 call->state = RX_STATE_DALLY;
2470 error = call->error;
2472 /* currentPacket, nLeft, and NFree must be zeroed here, because
2473 * ResetCall cannot: ResetCall may be called at splnet(), in the
2474 * kernel version, and may interrupt the macros rx_Read or
2475 * rx_Write, which run at normal priority for efficiency. */
2476 if (call->currentPacket) {
2477 #ifdef RX_TRACK_PACKETS
2478 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2480 rxi_FreePacket(call->currentPacket);
2481 call->currentPacket = (struct rx_packet *)0;
2484 call->nLeft = call->nFree = call->curlen = 0;
2486 /* Free any packets from the last call to ReadvProc/WritevProc */
2487 #ifdef RXDEBUG_PACKET
2489 #endif /* RXDEBUG_PACKET */
2490 rxi_FreePackets(0, &call->iovq);
2491 MUTEX_EXIT(&call->lock);
2493 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2494 if (conn->type == RX_CLIENT_CONNECTION) {
2495 MUTEX_ENTER(&conn->conn_data_lock);
2496 conn->flags &= ~RX_CONN_BUSY;
2497 MUTEX_EXIT(&conn->conn_data_lock);
2498 MUTEX_EXIT(&conn->conn_call_lock);
2502 * Map errors to the local host's errno.h format.
2504 error = ntoh_syserr_conv(error);
2508 #if !defined(KERNEL)
2510 /* Call this routine when shutting down a server or client (especially
2511 * clients). This will allow Rx to gracefully garbage collect server
2512 * connections, and reduce the number of retries that a server might
2513 * make to a dead client.
2514 * This is not quite right, since some calls may still be ongoing and
2515 * we can't lock them to destroy them. */
2519 struct rx_connection **conn_ptr, **conn_end;
2523 if (rxinit_status == 1) {
2525 return; /* Already shutdown. */
2527 rxi_DeleteCachedConnections();
2528 if (rx_connHashTable) {
2529 MUTEX_ENTER(&rx_connHashTable_lock);
2530 for (conn_ptr = &rx_connHashTable[0], conn_end =
2531 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2533 struct rx_connection *conn, *next;
2534 for (conn = *conn_ptr; conn; conn = next) {
2536 if (conn->type == RX_CLIENT_CONNECTION) {
2537 MUTEX_ENTER(&rx_refcnt_mutex);
2539 MUTEX_EXIT(&rx_refcnt_mutex);
2540 #ifdef RX_ENABLE_LOCKS
2541 rxi_DestroyConnectionNoLock(conn);
2542 #else /* RX_ENABLE_LOCKS */
2543 rxi_DestroyConnection(conn);
2544 #endif /* RX_ENABLE_LOCKS */
2548 #ifdef RX_ENABLE_LOCKS
2549 while (rx_connCleanup_list) {
2550 struct rx_connection *conn;
2551 conn = rx_connCleanup_list;
2552 rx_connCleanup_list = rx_connCleanup_list->next;
2553 MUTEX_EXIT(&rx_connHashTable_lock);
2554 rxi_CleanupConnection(conn);
2555 MUTEX_ENTER(&rx_connHashTable_lock);
2557 MUTEX_EXIT(&rx_connHashTable_lock);
2558 #endif /* RX_ENABLE_LOCKS */
2563 afs_winsockCleanup();
2571 /* if we wakeup packet waiter too often, can get in loop with two
2572 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2574 rxi_PacketsUnWait(void)
2576 if (!rx_waitingForPackets) {
2580 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2581 return; /* still over quota */
2584 rx_waitingForPackets = 0;
2585 #ifdef RX_ENABLE_LOCKS
2586 CV_BROADCAST(&rx_waitingForPackets_cv);
2588 osi_rxWakeup(&rx_waitingForPackets);
2594 /* ------------------Internal interfaces------------------------- */
2596 /* Return this process's service structure for the
2597 * specified socket and service */
2598 static struct rx_service *
2599 rxi_FindService(osi_socket socket, u_short serviceId)
2601 struct rx_service **sp;
2602 for (sp = &rx_services[0]; *sp; sp++) {
2603 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2609 #ifdef RXDEBUG_PACKET
2610 #ifdef KDUMP_RX_LOCK
2611 static struct rx_call_rx_lock *rx_allCallsp = 0;
2613 static struct rx_call *rx_allCallsp = 0;
2615 #endif /* RXDEBUG_PACKET */
2617 /* Allocate a call structure, for the indicated channel of the
2618 * supplied connection. The mode and state of the call must be set by
2619 * the caller. Returns the call with mutex locked. */
2620 static struct rx_call *
2621 rxi_NewCall(struct rx_connection *conn, int channel)
2623 struct rx_call *call;
2624 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2625 struct rx_call *cp; /* Call pointer temp */
2626 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2627 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2629 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2631 /* Grab an existing call structure, or allocate a new one.
2632 * Existing call structures are assumed to have been left reset by
2634 MUTEX_ENTER(&rx_freeCallQueue_lock);
2636 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2638 * EXCEPT that the TQ might not yet be cleared out.
2639 * Skip over those with in-use TQs.
2642 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2643 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2649 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2650 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2651 call = queue_First(&rx_freeCallQueue, rx_call);
2652 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2654 if (rx_stats_active)
2655 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2656 MUTEX_EXIT(&rx_freeCallQueue_lock);
2657 MUTEX_ENTER(&call->lock);
2658 CLEAR_CALL_QUEUE_LOCK(call);
2659 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2660 /* Now, if TQ wasn't cleared earlier, do it now. */
2661 rxi_WaitforTQBusy(call);
2662 if (call->flags & RX_CALL_TQ_CLEARME) {
2663 rxi_ClearTransmitQueue(call, 1);
2664 /*queue_Init(&call->tq);*/
2666 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2667 /* Bind the call to its connection structure */
2669 rxi_ResetCall(call, 1);
2672 call = rxi_Alloc(sizeof(struct rx_call));
2673 #ifdef RXDEBUG_PACKET
2674 call->allNextp = rx_allCallsp;
2675 rx_allCallsp = call;
2677 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2678 #else /* RXDEBUG_PACKET */
2679 rx_atomic_inc(&rx_stats.nCallStructs);
2680 #endif /* RXDEBUG_PACKET */
2682 MUTEX_EXIT(&rx_freeCallQueue_lock);
2683 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2684 MUTEX_ENTER(&call->lock);
2685 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2686 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2687 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2689 /* Initialize once-only items */
2690 queue_Init(&call->tq);
2691 queue_Init(&call->rq);
2692 queue_Init(&call->iovq);
2693 #ifdef RXDEBUG_PACKET
2694 call->rqc = call->tqc = call->iovqc = 0;
2695 #endif /* RXDEBUG_PACKET */
2696 /* Bind the call to its connection structure (prereq for reset) */
2698 rxi_ResetCall(call, 1);
2700 call->channel = channel;
2701 call->callNumber = &conn->callNumber[channel];
2702 call->rwind = conn->rwind[channel];
2703 call->twind = conn->twind[channel];
2704 /* Note that the next expected call number is retained (in
2705 * conn->callNumber[i]), even if we reallocate the call structure
2707 conn->call[channel] = call;
2708 /* if the channel's never been used (== 0), we should start at 1, otherwise
2709 * the call number is valid from the last time this channel was used */
2710 if (*call->callNumber == 0)
2711 *call->callNumber = 1;
2716 /* A call has been inactive long enough that so we can throw away
2717 * state, including the call structure, which is placed on the call
2720 * call->lock amd rx_refcnt_mutex are held upon entry.
2721 * haveCTLock is set when called from rxi_ReapConnections.
2723 * return 1 if the call is freed, 0 if not.
2726 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2728 int channel = call->channel;
2729 struct rx_connection *conn = call->conn;
2730 u_char state = call->state;
2733 * We are setting the state to RX_STATE_RESET to
2734 * ensure that no one else will attempt to use this
2735 * call once we drop the refcnt lock. We must drop
2736 * the refcnt lock before calling rxi_ResetCall
2737 * because it cannot be held across acquiring the
2738 * freepktQ lock. NewCall does the same.
2740 call->state = RX_STATE_RESET;
2741 MUTEX_EXIT(&rx_refcnt_mutex);
2742 rxi_ResetCall(call, 0);
2744 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2746 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2747 (*call->callNumber)++;
2749 if (call->conn->call[channel] == call)
2750 call->conn->call[channel] = 0;
2751 MUTEX_EXIT(&conn->conn_call_lock);
2754 * We couldn't obtain the conn_call_lock so we can't
2755 * disconnect the call from the connection. Set the
2756 * call state to dally so that the call can be reused.
2758 MUTEX_ENTER(&rx_refcnt_mutex);
2759 call->state = RX_STATE_DALLY;
2763 MUTEX_ENTER(&rx_freeCallQueue_lock);
2764 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2765 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2766 /* A call may be free even though its transmit queue is still in use.
2767 * Since we search the call list from head to tail, put busy calls at
2768 * the head of the list, and idle calls at the tail.
2770 if (call->flags & RX_CALL_TQ_BUSY)
2771 queue_Prepend(&rx_freeCallQueue, call);
2773 queue_Append(&rx_freeCallQueue, call);
2774 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2775 queue_Append(&rx_freeCallQueue, call);
2776 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2777 if (rx_stats_active)
2778 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2779 MUTEX_EXIT(&rx_freeCallQueue_lock);
2781 /* Destroy the connection if it was previously slated for
2782 * destruction, i.e. the Rx client code previously called
2783 * rx_DestroyConnection (client connections), or
2784 * rxi_ReapConnections called the same routine (server
2785 * connections). Only do this, however, if there are no
2786 * outstanding calls. Note that for fine grain locking, there appears
2787 * to be a deadlock in that rxi_FreeCall has a call locked and
2788 * DestroyConnectionNoLock locks each call in the conn. But note a
2789 * few lines up where we have removed this call from the conn.
2790 * If someone else destroys a connection, they either have no
2791 * call lock held or are going through this section of code.
2793 MUTEX_ENTER(&conn->conn_data_lock);
2794 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2795 MUTEX_ENTER(&rx_refcnt_mutex);
2797 MUTEX_EXIT(&rx_refcnt_mutex);
2798 MUTEX_EXIT(&conn->conn_data_lock);
2799 #ifdef RX_ENABLE_LOCKS
2801 rxi_DestroyConnectionNoLock(conn);
2803 rxi_DestroyConnection(conn);
2804 #else /* RX_ENABLE_LOCKS */
2805 rxi_DestroyConnection(conn);
2806 #endif /* RX_ENABLE_LOCKS */
2808 MUTEX_EXIT(&conn->conn_data_lock);
2810 MUTEX_ENTER(&rx_refcnt_mutex);
2814 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2815 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2818 rxi_Alloc(size_t size)
2822 if (rx_stats_active) {
2823 rx_atomic_add(&rxi_Allocsize, (int) size);
2824 rx_atomic_inc(&rxi_Alloccnt);
2828 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2829 afs_osi_Alloc_NoSleep(size);
2834 osi_Panic("rxi_Alloc error");
2840 rxi_Free(void *addr, size_t size)
2842 if (rx_stats_active) {
2843 rx_atomic_sub(&rxi_Allocsize, (int) size);
2844 rx_atomic_dec(&rxi_Alloccnt);
2846 osi_Free(addr, size);
2850 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2852 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2853 struct rx_peer *next = NULL;
2857 MUTEX_ENTER(&rx_peerHashTable_lock);
2859 peer_ptr = &rx_peerHashTable[0];
2860 peer_end = &rx_peerHashTable[rx_hashTableSize];
2863 for ( ; peer_ptr < peer_end; peer_ptr++) {
2866 for ( ; peer; peer = next) {
2868 if (host == peer->host)
2873 hashIndex = PEER_HASH(host, port);
2874 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2875 if ((peer->host == host) && (peer->port == port))
2880 MUTEX_ENTER(&rx_peerHashTable_lock);
2885 MUTEX_EXIT(&rx_peerHashTable_lock);
2887 MUTEX_ENTER(&peer->peer_lock);
2888 /* We don't handle dropping below min, so don't */
2889 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2890 peer->ifMTU=MIN(mtu, peer->ifMTU);
2891 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2892 /* if we tweaked this down, need to tune our peer MTU too */
2893 peer->MTU = MIN(peer->MTU, peer->natMTU);
2894 /* if we discovered a sub-1500 mtu, degrade */
2895 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2896 peer->maxDgramPackets = 1;
2897 /* We no longer have valid peer packet information */
2898 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2899 peer->maxPacketSize = 0;
2900 MUTEX_EXIT(&peer->peer_lock);
2902 MUTEX_ENTER(&rx_peerHashTable_lock);
2904 if (host && !port) {
2906 /* pick up where we left off */
2910 MUTEX_EXIT(&rx_peerHashTable_lock);
2913 /* Find the peer process represented by the supplied (host,port)
2914 * combination. If there is no appropriate active peer structure, a
2915 * new one will be allocated and initialized
2916 * The origPeer, if set, is a pointer to a peer structure on which the
2917 * refcount will be be decremented. This is used to replace the peer
2918 * structure hanging off a connection structure */
2920 rxi_FindPeer(afs_uint32 host, u_short port,
2921 struct rx_peer *origPeer, int create)
2925 hashIndex = PEER_HASH(host, port);
2926 MUTEX_ENTER(&rx_peerHashTable_lock);
2927 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2928 if ((pp->host == host) && (pp->port == port))
2933 pp = rxi_AllocPeer(); /* This bzero's *pp */
2934 pp->host = host; /* set here or in InitPeerParams is zero */
2936 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2937 queue_Init(&pp->rpcStats);
2938 pp->next = rx_peerHashTable[hashIndex];
2939 rx_peerHashTable[hashIndex] = pp;
2940 rxi_InitPeerParams(pp);
2941 if (rx_stats_active)
2942 rx_atomic_inc(&rx_stats.nPeerStructs);
2949 origPeer->refCount--;
2950 MUTEX_EXIT(&rx_peerHashTable_lock);
2955 /* Find the connection at (host, port) started at epoch, and with the
2956 * given connection id. Creates the server connection if necessary.
2957 * The type specifies whether a client connection or a server
2958 * connection is desired. In both cases, (host, port) specify the
2959 * peer's (host, pair) pair. Client connections are not made
2960 * automatically by this routine. The parameter socket gives the
2961 * socket descriptor on which the packet was received. This is used,
2962 * in the case of server connections, to check that *new* connections
2963 * come via a valid (port, serviceId). Finally, the securityIndex
2964 * parameter must match the existing index for the connection. If a
2965 * server connection is created, it will be created using the supplied
2966 * index, if the index is valid for this service */
2967 static struct rx_connection *
2968 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2969 u_short port, u_short serviceId, afs_uint32 cid,
2970 afs_uint32 epoch, int type, u_int securityIndex)
2972 int hashindex, flag, i;
2973 struct rx_connection *conn;
2974 hashindex = CONN_HASH(host, port, cid, epoch, type);
2975 MUTEX_ENTER(&rx_connHashTable_lock);
2976 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2977 rx_connHashTable[hashindex],
2980 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2981 && (epoch == conn->epoch)) {
2982 struct rx_peer *pp = conn->peer;
2983 if (securityIndex != conn->securityIndex) {
2984 /* this isn't supposed to happen, but someone could forge a packet
2985 * like this, and there seems to be some CM bug that makes this
2986 * happen from time to time -- in which case, the fileserver
2988 MUTEX_EXIT(&rx_connHashTable_lock);
2989 return (struct rx_connection *)0;
2991 if (pp->host == host && pp->port == port)
2993 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2995 /* So what happens when it's a callback connection? */
2996 if ( /*type == RX_CLIENT_CONNECTION && */
2997 (conn->epoch & 0x80000000))
3001 /* the connection rxLastConn that was used the last time is not the
3002 ** one we are looking for now. Hence, start searching in the hash */
3004 conn = rx_connHashTable[hashindex];
3009 struct rx_service *service;
3010 if (type == RX_CLIENT_CONNECTION) {
3011 MUTEX_EXIT(&rx_connHashTable_lock);
3012 return (struct rx_connection *)0;
3014 service = rxi_FindService(socket, serviceId);
3015 if (!service || (securityIndex >= service->nSecurityObjects)
3016 || (service->securityObjects[securityIndex] == 0)) {
3017 MUTEX_EXIT(&rx_connHashTable_lock);
3018 return (struct rx_connection *)0;
3020 conn = rxi_AllocConnection(); /* This bzero's the connection */
3021 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3022 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3023 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3024 conn->next = rx_connHashTable[hashindex];
3025 rx_connHashTable[hashindex] = conn;
3026 conn->peer = rxi_FindPeer(host, port, 0, 1);
3027 conn->type = RX_SERVER_CONNECTION;
3028 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3029 conn->epoch = epoch;
3030 conn->cid = cid & RX_CIDMASK;
3031 conn->ackRate = RX_FAST_ACK_RATE;
3032 conn->service = service;
3033 conn->serviceId = serviceId;
3034 conn->securityIndex = securityIndex;
3035 conn->securityObject = service->securityObjects[securityIndex];
3036 conn->nSpecific = 0;
3037 conn->specific = NULL;
3038 rx_SetConnDeadTime(conn, service->connDeadTime);
3039 conn->idleDeadTime = service->idleDeadTime;
3040 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3041 for (i = 0; i < RX_MAXCALLS; i++) {
3042 conn->twind[i] = rx_initSendWindow;
3043 conn->rwind[i] = rx_initReceiveWindow;
3045 /* Notify security object of the new connection */
3046 RXS_NewConnection(conn->securityObject, conn);
3047 /* XXXX Connection timeout? */
3048 if (service->newConnProc)
3049 (*service->newConnProc) (conn);
3050 if (rx_stats_active)
3051 rx_atomic_inc(&rx_stats.nServerConns);
3054 MUTEX_ENTER(&rx_refcnt_mutex);
3056 MUTEX_EXIT(&rx_refcnt_mutex);
3058 rxLastConn = conn; /* store this connection as the last conn used */
3059 MUTEX_EXIT(&rx_connHashTable_lock);
3064 * Timeout a call on a busy call channel if appropriate.
3066 * @param[in] call The busy call.
3068 * @pre 'call' is marked as busy (namely,
3069 * call->conn->lastBusy[call->channel] != 0)
3071 * @pre call->lock is held
3072 * @pre rxi_busyChannelError is nonzero
3074 * @note call->lock is dropped and reacquired
3077 rxi_CheckBusy(struct rx_call *call)
3079 struct rx_connection *conn = call->conn;
3080 int channel = call->channel;
3081 int freechannel = 0;
3083 afs_uint32 callNumber;
3085 MUTEX_EXIT(&call->lock);
3087 MUTEX_ENTER(&conn->conn_call_lock);
3088 callNumber = *call->callNumber;
3090 /* Are there any other call slots on this conn that we should try? Look for
3091 * slots that are empty and are either non-busy, or were marked as busy
3092 * longer than conn->secondsUntilDead seconds before this call started. */
3094 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3096 /* only look at channels that aren't us */
3100 if (conn->lastBusy[i]) {
3101 /* if this channel looked busy too recently, don't look at it */
3102 if (conn->lastBusy[i] >= call->startTime.sec) {
3105 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3110 if (conn->call[i]) {
3111 struct rx_call *tcall = conn->call[i];
3112 MUTEX_ENTER(&tcall->lock);
3113 if (tcall->state == RX_STATE_DALLY) {
3116 MUTEX_EXIT(&tcall->lock);
3122 MUTEX_ENTER(&call->lock);
3124 /* Since the call->lock and conn->conn_call_lock have been released it is
3125 * possible that (1) the call may no longer be busy and/or (2) the call may
3126 * have been reused by another waiting thread. Therefore, we must confirm
3127 * that the call state has not changed when deciding whether or not to
3128 * force this application thread to retry by forcing a Timeout error. */
3130 if (freechannel && *call->callNumber == callNumber &&
3131 (call->flags & RX_CALL_PEER_BUSY)) {
3132 /* Since 'freechannel' is set, there exists another channel in this
3133 * rx_conn that the application thread might be able to use. We know
3134 * that we have the correct call since callNumber is unchanged, and we
3135 * know that the call is still busy. So, set the call error state to
3136 * rxi_busyChannelError so the application can retry the request,
3137 * presumably on a less-busy call channel. */
3139 rxi_CallError(call, RX_CALL_BUSY);
3141 MUTEX_EXIT(&conn->conn_call_lock);
3144 /* There are two packet tracing routines available for testing and monitoring
3145 * Rx. One is called just after every packet is received and the other is
3146 * called just before every packet is sent. Received packets, have had their
3147 * headers decoded, and packets to be sent have not yet had their headers
3148 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3149 * containing the network address. Both can be modified. The return value, if
3150 * non-zero, indicates that the packet should be dropped. */
3152 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3153 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3155 /* A packet has been received off the interface. Np is the packet, socket is
3156 * the socket number it was received from (useful in determining which service
3157 * this packet corresponds to), and (host, port) reflect the host,port of the
3158 * sender. This call returns the packet to the caller if it is finished with
3159 * it, rather than de-allocating it, just as a small performance hack */
3162 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3163 afs_uint32 host, u_short port, int *tnop,
3164 struct rx_call **newcallp)
3166 struct rx_call *call;
3167 struct rx_connection *conn;
3169 afs_uint32 currentCallNumber;
3174 struct rx_packet *tnp;
3177 /* We don't print out the packet until now because (1) the time may not be
3178 * accurate enough until now in the lwp implementation (rx_Listener only gets
3179 * the time after the packet is read) and (2) from a protocol point of view,
3180 * this is the first time the packet has been seen */
3181 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3182 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3183 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3184 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3185 np->header.epoch, np->header.cid, np->header.callNumber,
3186 np->header.seq, np->header.flags, np));
3189 /* Account for connectionless packets */
3190 if (rx_stats_active &&
3191 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3192 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3193 struct rx_peer *peer;
3195 /* Try to look up the peer structure, but don't create one */
3196 peer = rxi_FindPeer(host, port, 0, 0);
3198 /* Since this may not be associated with a connection, it may have
3199 * no refCount, meaning we could race with ReapConnections
3202 if (peer && (peer->refCount > 0)) {
3203 MUTEX_ENTER(&peer->peer_lock);
3204 peer->bytesReceived += np->length;
3205 MUTEX_EXIT(&peer->peer_lock);
3209 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3210 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3213 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3214 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3217 /* If an input tracer function is defined, call it with the packet and
3218 * network address. Note this function may modify its arguments. */
3219 if (rx_justReceived) {
3220 struct sockaddr_in addr;
3222 addr.sin_family = AF_INET;
3223 addr.sin_port = port;
3224 addr.sin_addr.s_addr = host;
3225 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3226 addr.sin_len = sizeof(addr);
3227 #endif /* AFS_OSF_ENV */
3228 drop = (*rx_justReceived) (np, &addr);
3229 /* drop packet if return value is non-zero */
3232 port = addr.sin_port; /* in case fcn changed addr */
3233 host = addr.sin_addr.s_addr;
3237 /* If packet was not sent by the client, then *we* must be the client */
3238 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3239 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3241 /* Find the connection (or fabricate one, if we're the server & if
3242 * necessary) associated with this packet */
3244 rxi_FindConnection(socket, host, port, np->header.serviceId,
3245 np->header.cid, np->header.epoch, type,
3246 np->header.securityIndex);
3248 /* To avoid having 2 connections just abort at each other,
3249 don't abort an abort. */
3251 if (np->header.type != RX_PACKET_TYPE_ABORT)
3252 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3257 /* If we're doing statistics, then account for the incoming packet */
3258 if (rx_stats_active) {
3259 MUTEX_ENTER(&conn->peer->peer_lock);
3260 conn->peer->bytesReceived += np->length;
3261 MUTEX_EXIT(&conn->peer->peer_lock);
3264 /* If the connection is in an error state, send an abort packet and ignore
3265 * the incoming packet */
3267 /* Don't respond to an abort packet--we don't want loops! */
3268 MUTEX_ENTER(&conn->conn_data_lock);
3269 if (np->header.type != RX_PACKET_TYPE_ABORT)
3270 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3271 putConnection(conn);
3272 MUTEX_EXIT(&conn->conn_data_lock);
3276 /* Check for connection-only requests (i.e. not call specific). */
3277 if (np->header.callNumber == 0) {
3278 switch (np->header.type) {
3279 case RX_PACKET_TYPE_ABORT: {
3280 /* What if the supplied error is zero? */
3281 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3282 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3283 rxi_ConnectionError(conn, errcode);
3284 putConnection(conn);
3287 case RX_PACKET_TYPE_CHALLENGE:
3288 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3289 putConnection(conn);
3291 case RX_PACKET_TYPE_RESPONSE:
3292 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3293 putConnection(conn);
3295 case RX_PACKET_TYPE_PARAMS:
3296 case RX_PACKET_TYPE_PARAMS + 1:
3297 case RX_PACKET_TYPE_PARAMS + 2:
3298 /* ignore these packet types for now */
3299 putConnection(conn);
3303 /* Should not reach here, unless the peer is broken: send an
3305 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3306 MUTEX_ENTER(&conn->conn_data_lock);
3307 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3308 putConnection(conn);
3309 MUTEX_EXIT(&conn->conn_data_lock);
3314 channel = np->header.cid & RX_CHANNELMASK;
3315 MUTEX_ENTER(&conn->conn_call_lock);
3316 call = conn->call[channel];
3319 MUTEX_ENTER(&call->lock);
3320 currentCallNumber = conn->callNumber[channel];
3321 MUTEX_EXIT(&conn->conn_call_lock);
3322 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3323 call = conn->call[channel];
3325 MUTEX_ENTER(&call->lock);
3326 currentCallNumber = conn->callNumber[channel];
3327 MUTEX_EXIT(&conn->conn_call_lock);
3329 call = rxi_NewCall(conn, channel); /* returns locked call */
3330 *call->callNumber = currentCallNumber = np->header.callNumber;
3331 MUTEX_EXIT(&conn->conn_call_lock);
3333 if (np->header.callNumber == 0)
3334 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3335 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3336 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3337 np->header.flags, np, np->length));
3339 call->state = RX_STATE_PRECALL;
3340 clock_GetTime(&call->queueTime);
3341 call->bytesSent = 0;
3342 call->bytesRcvd = 0;
3344 * If the number of queued calls exceeds the overload
3345 * threshold then abort this call.
3347 if ((rx_BusyThreshold > 0) &&
3348 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3349 struct rx_packet *tp;
3351 rxi_CallError(call, rx_BusyError);
3352 tp = rxi_SendCallAbort(call, np, 1, 0);
3353 MUTEX_EXIT(&call->lock);
3354 putConnection(conn);
3355 if (rx_stats_active)
3356 rx_atomic_inc(&rx_stats.nBusies);
3359 rxi_KeepAliveOn(call);
3361 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3362 /* This packet can't be for this call. If the new call address is
3363 * 0 then no call is running on this channel. If there is a call
3364 * then, since this is a client connection we're getting data for
3365 * it must be for the previous call.
3367 MUTEX_EXIT(&conn->conn_call_lock);
3368 if (rx_stats_active)
3369 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3370 putConnection(conn);
3374 /* There is a non-NULL locked call at this point */
3375 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3376 if (np->header.callNumber < currentCallNumber) {
3377 MUTEX_EXIT(&call->lock);
3378 if (rx_stats_active)
3379 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3380 putConnection(conn);
3382 } else if (np->header.callNumber != currentCallNumber) {
3383 /* Wait until the transmit queue is idle before deciding
3384 * whether to reset the current call. Chances are that the
3385 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3388 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3389 if (call->state == RX_STATE_ACTIVE) {
3390 rxi_WaitforTQBusy(call);
3392 * If we entered error state while waiting,
3393 * must call rxi_CallError to permit rxi_ResetCall
3394 * to processed when the tqWaiter count hits zero.
3397 rxi_CallError(call, call->error);
3398 MUTEX_EXIT(&call->lock);
3399 putConnection(conn);
3403 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3404 /* If the new call cannot be taken right now send a busy and set
3405 * the error condition in this call, so that it terminates as
3406 * quickly as possible */
3407 if (call->state == RX_STATE_ACTIVE) {
3408 struct rx_packet *tp;
3410 rxi_CallError(call, RX_CALL_DEAD);
3411 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3413 MUTEX_EXIT(&call->lock);
3414 putConnection(conn);
3417 rxi_ResetCall(call, 0);
3419 * The conn_call_lock is not held but no one else should be
3420 * using this call channel while we are processing this incoming
3421 * packet. This assignment should be safe.
3423 *call->callNumber = np->header.callNumber;
3425 if (np->header.callNumber == 0)
3426 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3427 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3428 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3429 np->header.flags, np, np->length));
3431 call->state = RX_STATE_PRECALL;
3432 clock_GetTime(&call->queueTime);
3433 call->bytesSent = 0;
3434 call->bytesRcvd = 0;
3436 * If the number of queued calls exceeds the overload
3437 * threshold then abort this call.
3439 if ((rx_BusyThreshold > 0) &&
3440 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3441 struct rx_packet *tp;
3443 rxi_CallError(call, rx_BusyError);
3444 tp = rxi_SendCallAbort(call, np, 1, 0);
3445 MUTEX_EXIT(&call->lock);
3446 putConnection(conn);
3447 if (rx_stats_active)
3448 rx_atomic_inc(&rx_stats.nBusies);
3451 rxi_KeepAliveOn(call);
3453 /* Continuing call; do nothing here. */
3455 } else { /* we're the client */
3456 /* Ignore all incoming acknowledgements for calls in DALLY state */
3457 if ((call->state == RX_STATE_DALLY)
3458 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3459 if (rx_stats_active)
3460 rx_atomic_inc(&rx_stats.ignorePacketDally);
3461 MUTEX_EXIT(&call->lock);
3462 putConnection(conn);
3466 /* Ignore anything that's not relevant to the current call. If there
3467 * isn't a current call, then no packet is relevant. */
3468 if (np->header.callNumber != currentCallNumber) {
3469 if (rx_stats_active)
3470 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3471 MUTEX_EXIT(&call->lock);
3472 putConnection(conn);
3475 /* If the service security object index stamped in the packet does not
3476 * match the connection's security index, ignore the packet */
3477 if (np->header.securityIndex != conn->securityIndex) {
3478 MUTEX_EXIT(&call->lock);
3479 putConnection(conn);
3483 /* If we're receiving the response, then all transmit packets are
3484 * implicitly acknowledged. Get rid of them. */
3485 if (np->header.type == RX_PACKET_TYPE_DATA) {
3486 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3487 /* XXX Hack. Because we must release the global rx lock when
3488 * sending packets (osi_NetSend) we drop all acks while we're
3489 * traversing the tq in rxi_Start sending packets out because
3490 * packets may move to the freePacketQueue as result of being here!
3491 * So we drop these packets until we're safely out of the
3492 * traversing. Really ugly!
3493 * For fine grain RX locking, we set the acked field in the
3494 * packets and let rxi_Start remove them from the transmit queue.
3496 if (call->flags & RX_CALL_TQ_BUSY) {
3497 #ifdef RX_ENABLE_LOCKS
3498 rxi_SetAcksInTransmitQueue(call);
3500 putConnection(conn);
3501 return np; /* xmitting; drop packet */
3504 rxi_ClearTransmitQueue(call, 0);
3506 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3507 rxi_ClearTransmitQueue(call, 0);
3508 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3510 if (np->header.type == RX_PACKET_TYPE_ACK) {
3511 /* now check to see if this is an ack packet acknowledging that the
3512 * server actually *lost* some hard-acked data. If this happens we
3513 * ignore this packet, as it may indicate that the server restarted in
3514 * the middle of a call. It is also possible that this is an old ack
3515 * packet. We don't abort the connection in this case, because this
3516 * *might* just be an old ack packet. The right way to detect a server
3517 * restart in the midst of a call is to notice that the server epoch
3519 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3520 * XXX unacknowledged. I think that this is off-by-one, but
3521 * XXX I don't dare change it just yet, since it will
3522 * XXX interact badly with the server-restart detection
3523 * XXX code in receiveackpacket. */
3524 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3525 if (rx_stats_active)
3526 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3527 MUTEX_EXIT(&call->lock);
3528 putConnection(conn);
3532 } /* else not a data packet */
3535 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3536 /* Set remote user defined status from packet */
3537 call->remoteStatus = np->header.userStatus;
3539 /* Now do packet type-specific processing */
3540 switch (np->header.type) {
3541 case RX_PACKET_TYPE_DATA:
3542 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3545 case RX_PACKET_TYPE_ACK:
3546 /* Respond immediately to ack packets requesting acknowledgement
3548 if (np->header.flags & RX_REQUEST_ACK) {
3550 (void)rxi_SendCallAbort(call, 0, 1, 0);
3552 (void)rxi_SendAck(call, 0, np->header.serial,
3553 RX_ACK_PING_RESPONSE, 1);
3555 np = rxi_ReceiveAckPacket(call, np, 1);
3557 case RX_PACKET_TYPE_ABORT: {
3558 /* An abort packet: reset the call, passing the error up to the user. */
3559 /* What if error is zero? */
3560 /* What if the error is -1? the application will treat it as a timeout. */
3561 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3562 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3563 rxi_CallError(call, errdata);
3564 MUTEX_EXIT(&call->lock);
3565 putConnection(conn);
3566 return np; /* xmitting; drop packet */
3568 case RX_PACKET_TYPE_BUSY: {
3569 struct clock busyTime;
3571 clock_GetTime(&busyTime);
3573 MUTEX_EXIT(&call->lock);
3575 MUTEX_ENTER(&conn->conn_call_lock);
3576 MUTEX_ENTER(&call->lock);
3577 conn->lastBusy[call->channel] = busyTime.sec;
3578 call->flags |= RX_CALL_PEER_BUSY;
3579 MUTEX_EXIT(&call->lock);
3580 MUTEX_EXIT(&conn->conn_call_lock);
3582 putConnection(conn);
3586 case RX_PACKET_TYPE_ACKALL:
3587 /* All packets acknowledged, so we can drop all packets previously
3588 * readied for sending */
3589 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3590 /* XXX Hack. We because we can't release the global rx lock when
3591 * sending packets (osi_NetSend) we drop all ack pkts while we're
3592 * traversing the tq in rxi_Start sending packets out because
3593 * packets may move to the freePacketQueue as result of being
3594 * here! So we drop these packets until we're safely out of the
3595 * traversing. Really ugly!
3596 * For fine grain RX locking, we set the acked field in the packets
3597 * and let rxi_Start remove the packets from the transmit queue.
3599 if (call->flags & RX_CALL_TQ_BUSY) {
3600 #ifdef RX_ENABLE_LOCKS
3601 rxi_SetAcksInTransmitQueue(call);
3603 #else /* RX_ENABLE_LOCKS */
3604 MUTEX_EXIT(&call->lock);
3605 putConnection(conn);
3606 return np; /* xmitting; drop packet */
3607 #endif /* RX_ENABLE_LOCKS */
3609 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3610 rxi_ClearTransmitQueue(call, 0);
3613 /* Should not reach here, unless the peer is broken: send an abort
3615 rxi_CallError(call, RX_PROTOCOL_ERROR);
3616 np = rxi_SendCallAbort(call, np, 1, 0);
3619 /* Note when this last legitimate packet was received, for keep-alive
3620 * processing. Note, we delay getting the time until now in the hope that
3621 * the packet will be delivered to the user before any get time is required
3622 * (if not, then the time won't actually be re-evaluated here). */
3623 call->lastReceiveTime = clock_Sec();
3624 /* we've received a legit packet, so the channel is not busy */
3625 call->flags &= ~RX_CALL_PEER_BUSY;
3626 MUTEX_EXIT(&call->lock);
3627 putConnection(conn);
3631 /* return true if this is an "interesting" connection from the point of view
3632 of someone trying to debug the system */
3634 rxi_IsConnInteresting(struct rx_connection *aconn)
3637 struct rx_call *tcall;
3639 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3642 for (i = 0; i < RX_MAXCALLS; i++) {
3643 tcall = aconn->call[i];
3645 if ((tcall->state == RX_STATE_PRECALL)
3646 || (tcall->state == RX_STATE_ACTIVE))
3648 if ((tcall->mode == RX_MODE_SENDING)
3649 || (tcall->mode == RX_MODE_RECEIVING))
3657 /* if this is one of the last few packets AND it wouldn't be used by the
3658 receiving call to immediately satisfy a read request, then drop it on
3659 the floor, since accepting it might prevent a lock-holding thread from
3660 making progress in its reading. If a call has been cleared while in
3661 the precall state then ignore all subsequent packets until the call
3662 is assigned to a thread. */
3665 TooLow(struct rx_packet *ap, struct rx_call *acall)
3669 MUTEX_ENTER(&rx_quota_mutex);
3670 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3671 && (acall->state == RX_STATE_PRECALL))
3672 || ((rx_nFreePackets < rxi_dataQuota + 2)
3673 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3674 && (acall->flags & RX_CALL_READER_WAIT)))) {
3677 MUTEX_EXIT(&rx_quota_mutex);
3683 * Clear the attach wait flag on a connection and proceed.
3685 * Any processing waiting for a connection to be attached should be
3686 * unblocked. We clear the flag and do any other needed tasks.
3689 * the conn to unmark waiting for attach
3691 * @pre conn's conn_data_lock must be locked before calling this function
3695 rxi_ConnClearAttachWait(struct rx_connection *conn)
3697 /* Indicate that rxi_CheckReachEvent is no longer running by
3698 * clearing the flag. Must be atomic under conn_data_lock to
3699 * avoid a new call slipping by: rxi_CheckConnReach holds
3700 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3702 conn->flags &= ~RX_CONN_ATTACHWAIT;
3703 if (conn->flags & RX_CONN_NAT_PING) {
3704 conn->flags &= ~RX_CONN_NAT_PING;
3705 rxi_ScheduleNatKeepAliveEvent(conn);
3710 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3712 struct rx_connection *conn = arg1;
3713 struct rx_call *acall = arg2;
3714 struct rx_call *call = acall;
3715 struct clock when, now;
3718 MUTEX_ENTER(&conn->conn_data_lock);
3721 rxevent_Put(conn->checkReachEvent);
3722 conn->checkReachEvent = NULL;
3725 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3727 putConnection(conn);
3729 MUTEX_EXIT(&conn->conn_data_lock);
3733 MUTEX_ENTER(&conn->conn_call_lock);
3734 MUTEX_ENTER(&conn->conn_data_lock);
3735 for (i = 0; i < RX_MAXCALLS; i++) {
3736 struct rx_call *tc = conn->call[i];
3737 if (tc && tc->state == RX_STATE_PRECALL) {
3743 rxi_ConnClearAttachWait(conn);
3744 MUTEX_EXIT(&conn->conn_data_lock);
3745 MUTEX_EXIT(&conn->conn_call_lock);
3750 MUTEX_ENTER(&call->lock);
3751 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3753 MUTEX_EXIT(&call->lock);
3755 clock_GetTime(&now);
3757 when.sec += RX_CHECKREACH_TIMEOUT;
3758 MUTEX_ENTER(&conn->conn_data_lock);
3759 if (!conn->checkReachEvent) {
3760 MUTEX_ENTER(&rx_refcnt_mutex);
3762 MUTEX_EXIT(&rx_refcnt_mutex);
3763 conn->checkReachEvent = rxevent_Post(&when, &now,
3764 rxi_CheckReachEvent, conn,
3767 MUTEX_EXIT(&conn->conn_data_lock);
3773 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3775 struct rx_service *service = conn->service;
3776 struct rx_peer *peer = conn->peer;
3777 afs_uint32 now, lastReach;
3779 if (service->checkReach == 0)
3783 MUTEX_ENTER(&peer->peer_lock);
3784 lastReach = peer->lastReachTime;
3785 MUTEX_EXIT(&peer->peer_lock);
3786 if (now - lastReach < RX_CHECKREACH_TTL)
3789 MUTEX_ENTER(&conn->conn_data_lock);
3790 if (conn->flags & RX_CONN_ATTACHWAIT) {
3791 MUTEX_EXIT(&conn->conn_data_lock);
3794 conn->flags |= RX_CONN_ATTACHWAIT;
3795 MUTEX_EXIT(&conn->conn_data_lock);
3796 if (!conn->checkReachEvent)
3797 rxi_CheckReachEvent(NULL, conn, call, 0);
3802 /* try to attach call, if authentication is complete */
3804 TryAttach(struct rx_call *acall, osi_socket socket,
3805 int *tnop, struct rx_call **newcallp,
3808 struct rx_connection *conn = acall->conn;
3810 if (conn->type == RX_SERVER_CONNECTION
3811 && acall->state == RX_STATE_PRECALL) {
3812 /* Don't attach until we have any req'd. authentication. */
3813 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3814 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3815 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3816 /* Note: this does not necessarily succeed; there
3817 * may not any proc available
3820 rxi_ChallengeOn(acall->conn);
3825 /* A data packet has been received off the interface. This packet is
3826 * appropriate to the call (the call is in the right state, etc.). This
3827 * routine can return a packet to the caller, for re-use */
3829 static struct rx_packet *
3830 rxi_ReceiveDataPacket(struct rx_call *call,
3831 struct rx_packet *np, int istack,
3832 osi_socket socket, afs_uint32 host, u_short port,
3833 int *tnop, struct rx_call **newcallp)
3835 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3840 afs_uint32 serial=0, flags=0;
3842 struct rx_packet *tnp;
3843 if (rx_stats_active)
3844 rx_atomic_inc(&rx_stats.dataPacketsRead);
3847 /* If there are no packet buffers, drop this new packet, unless we can find
3848 * packet buffers from inactive calls */
3850 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3851 MUTEX_ENTER(&rx_freePktQ_lock);
3852 rxi_NeedMorePackets = TRUE;
3853 MUTEX_EXIT(&rx_freePktQ_lock);
3854 if (rx_stats_active)
3855 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3856 rxi_calltrace(RX_TRACE_DROP, call);
3857 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3858 /* We used to clear the receive queue here, in an attempt to free
3859 * packets. However this is unsafe if the queue has received a
3860 * soft ACK for the final packet */
3861 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3867 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3868 * packet is one of several packets transmitted as a single
3869 * datagram. Do not send any soft or hard acks until all packets
3870 * in a jumbogram have been processed. Send negative acks right away.
3872 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3873 /* tnp is non-null when there are more packets in the
3874 * current jumbo gram */
3881 seq = np->header.seq;
3882 serial = np->header.serial;
3883 flags = np->header.flags;
3885 /* If the call is in an error state, send an abort message */
3887 return rxi_SendCallAbort(call, np, istack, 0);
3889 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3890 * AFS 3.5 jumbogram. */
3891 if (flags & RX_JUMBO_PACKET) {
3892 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3897 if (np->header.spare != 0) {
3898 MUTEX_ENTER(&call->conn->conn_data_lock);
3899 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3900 MUTEX_EXIT(&call->conn->conn_data_lock);
3903 /* The usual case is that this is the expected next packet */
3904 if (seq == call->rnext) {
3906 /* Check to make sure it is not a duplicate of one already queued */
3907 if (queue_IsNotEmpty(&call->rq)
3908 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3909 if (rx_stats_active)
3910 rx_atomic_inc(&rx_stats.dupPacketsRead);
3911 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3912 rxevent_Cancel(&call->delayedAckEvent, call,
3913 RX_CALL_REFCOUNT_DELAY);
3914 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3920 /* It's the next packet. Stick it on the receive queue
3921 * for this call. Set newPackets to make sure we wake
3922 * the reader once all packets have been processed */
3923 #ifdef RX_TRACK_PACKETS
3924 np->flags |= RX_PKTFLAG_RQ;
3926 queue_Prepend(&call->rq, np);
3927 #ifdef RXDEBUG_PACKET
3929 #endif /* RXDEBUG_PACKET */
3931 np = NULL; /* We can't use this anymore */
3934 /* If an ack is requested then set a flag to make sure we
3935 * send an acknowledgement for this packet */
3936 if (flags & RX_REQUEST_ACK) {
3937 ackNeeded = RX_ACK_REQUESTED;
3940 /* Keep track of whether we have received the last packet */
3941 if (flags & RX_LAST_PACKET) {
3942 call->flags |= RX_CALL_HAVE_LAST;
3946 /* Check whether we have all of the packets for this call */
3947 if (call->flags & RX_CALL_HAVE_LAST) {
3948 afs_uint32 tseq; /* temporary sequence number */
3949 struct rx_packet *tp; /* Temporary packet pointer */
3950 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3952 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3953 if (tseq != tp->header.seq)
3955 if (tp->header.flags & RX_LAST_PACKET) {
3956 call->flags |= RX_CALL_RECEIVE_DONE;
3963 /* Provide asynchronous notification for those who want it
3964 * (e.g. multi rx) */
3965 if (call->arrivalProc) {
3966 (*call->arrivalProc) (call, call->arrivalProcHandle,
3967 call->arrivalProcArg);
3968 call->arrivalProc = (void (*)())0;
3971 /* Update last packet received */
3974 /* If there is no server process serving this call, grab
3975 * one, if available. We only need to do this once. If a
3976 * server thread is available, this thread becomes a server
3977 * thread and the server thread becomes a listener thread. */
3979 TryAttach(call, socket, tnop, newcallp, 0);
3982 /* This is not the expected next packet. */
3984 /* Determine whether this is a new or old packet, and if it's
3985 * a new one, whether it fits into the current receive window.
3986 * Also figure out whether the packet was delivered in sequence.
3987 * We use the prev variable to determine whether the new packet
3988 * is the successor of its immediate predecessor in the
3989 * receive queue, and the missing flag to determine whether
3990 * any of this packets predecessors are missing. */
3992 afs_uint32 prev; /* "Previous packet" sequence number */
3993 struct rx_packet *tp; /* Temporary packet pointer */
3994 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3995 int missing; /* Are any predecessors missing? */
3997 /* If the new packet's sequence number has been sent to the
3998 * application already, then this is a duplicate */
3999 if (seq < call->rnext) {
4000 if (rx_stats_active)
4001 rx_atomic_inc(&rx_stats.dupPacketsRead);
4002 rxevent_Cancel(&call->delayedAckEvent, call,
4003 RX_CALL_REFCOUNT_DELAY);
4004 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4010 /* If the sequence number is greater than what can be
4011 * accomodated by the current window, then send a negative
4012 * acknowledge and drop the packet */
4013 if ((call->rnext + call->rwind) <= seq) {
4014 rxevent_Cancel(&call->delayedAckEvent, call,
4015 RX_CALL_REFCOUNT_DELAY);
4016 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4023 /* Look for the packet in the queue of old received packets */
4024 for (prev = call->rnext - 1, missing =
4025 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4026 /*Check for duplicate packet */
4027 if (seq == tp->header.seq) {
4028 if (rx_stats_active)
4029 rx_atomic_inc(&rx_stats.dupPacketsRead);
4030 rxevent_Cancel(&call->delayedAckEvent, call,
4031 RX_CALL_REFCOUNT_DELAY);
4032 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4038 /* If we find a higher sequence packet, break out and
4039 * insert the new packet here. */
4040 if (seq < tp->header.seq)
4042 /* Check for missing packet */
4043 if (tp->header.seq != prev + 1) {
4047 prev = tp->header.seq;
4050 /* Keep track of whether we have received the last packet. */
4051 if (flags & RX_LAST_PACKET) {
4052 call->flags |= RX_CALL_HAVE_LAST;
4055 /* It's within the window: add it to the the receive queue.
4056 * tp is left by the previous loop either pointing at the
4057 * packet before which to insert the new packet, or at the
4058 * queue head if the queue is empty or the packet should be
4060 #ifdef RX_TRACK_PACKETS
4061 np->flags |= RX_PKTFLAG_RQ;
4063 #ifdef RXDEBUG_PACKET
4065 #endif /* RXDEBUG_PACKET */
4066 queue_InsertBefore(tp, np);
4070 /* Check whether we have all of the packets for this call */
4071 if ((call->flags & RX_CALL_HAVE_LAST)
4072 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4073 afs_uint32 tseq; /* temporary sequence number */
4076 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4077 if (tseq != tp->header.seq)
4079 if (tp->header.flags & RX_LAST_PACKET) {
4080 call->flags |= RX_CALL_RECEIVE_DONE;
4087 /* We need to send an ack of the packet is out of sequence,
4088 * or if an ack was requested by the peer. */
4089 if (seq != prev + 1 || missing) {
4090 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4091 } else if (flags & RX_REQUEST_ACK) {
4092 ackNeeded = RX_ACK_REQUESTED;
4095 /* Acknowledge the last packet for each call */
4096 if (flags & RX_LAST_PACKET) {
4107 * If the receiver is waiting for an iovec, fill the iovec
4108 * using the data from the receive queue */
4109 if (call->flags & RX_CALL_IOVEC_WAIT) {
4110 didHardAck = rxi_FillReadVec(call, serial);
4111 /* the call may have been aborted */
4120 /* Wakeup the reader if any */
4121 if ((call->flags & RX_CALL_READER_WAIT)
4122 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4123 || (call->iovNext >= call->iovMax)
4124 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4125 call->flags &= ~RX_CALL_READER_WAIT;
4126 #ifdef RX_ENABLE_LOCKS
4127 CV_BROADCAST(&call->cv_rq);
4129 osi_rxWakeup(&call->rq);
4135 * Send an ack when requested by the peer, or once every
4136 * rxi_SoftAckRate packets until the last packet has been
4137 * received. Always send a soft ack for the last packet in
4138 * the server's reply. */
4140 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4141 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4142 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4143 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4144 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4145 } else if (call->nSoftAcks) {
4146 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4147 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4149 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4150 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4151 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4158 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4160 struct rx_peer *peer = conn->peer;
4162 MUTEX_ENTER(&peer->peer_lock);
4163 peer->lastReachTime = clock_Sec();
4164 MUTEX_EXIT(&peer->peer_lock);
4166 MUTEX_ENTER(&conn->conn_data_lock);
4167 if (conn->flags & RX_CONN_ATTACHWAIT) {
4170 rxi_ConnClearAttachWait(conn);
4171 MUTEX_EXIT(&conn->conn_data_lock);
4173 for (i = 0; i < RX_MAXCALLS; i++) {
4174 struct rx_call *call = conn->call[i];
4177 MUTEX_ENTER(&call->lock);
4178 /* tnop can be null if newcallp is null */
4179 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4181 MUTEX_EXIT(&call->lock);
4185 MUTEX_EXIT(&conn->conn_data_lock);
4188 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4190 rx_ack_reason(int reason)
4193 case RX_ACK_REQUESTED:
4195 case RX_ACK_DUPLICATE:
4197 case RX_ACK_OUT_OF_SEQUENCE:
4199 case RX_ACK_EXCEEDS_WINDOW:
4201 case RX_ACK_NOSPACE:
4205 case RX_ACK_PING_RESPONSE:
4218 /* The real smarts of the whole thing. */
4219 static struct rx_packet *
4220 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4223 struct rx_ackPacket *ap;
4225 struct rx_packet *tp;
4226 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4227 struct rx_connection *conn = call->conn;
4228 struct rx_peer *peer = conn->peer;
4229 struct clock now; /* Current time, for RTT calculations */
4237 int newAckCount = 0;
4238 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4239 int pktsize = 0; /* Set if we need to update the peer mtu */
4240 int conn_data_locked = 0;
4242 if (rx_stats_active)
4243 rx_atomic_inc(&rx_stats.ackPacketsRead);
4244 ap = (struct rx_ackPacket *)rx_DataOf(np);
4245 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4247 return np; /* truncated ack packet */
4249 /* depends on ack packet struct */
4250 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4251 first = ntohl(ap->firstPacket);
4252 prev = ntohl(ap->previousPacket);
4253 serial = ntohl(ap->serial);
4255 /* Ignore ack packets received out of order */
4256 if (first < call->tfirst ||
4257 (first == call->tfirst && prev < call->tprev)) {
4263 if (np->header.flags & RX_SLOW_START_OK) {
4264 call->flags |= RX_CALL_SLOW_START_OK;
4267 if (ap->reason == RX_ACK_PING_RESPONSE)
4268 rxi_UpdatePeerReach(conn, call);
4270 if (conn->lastPacketSizeSeq) {
4271 MUTEX_ENTER(&conn->conn_data_lock);
4272 conn_data_locked = 1;
4273 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4274 pktsize = conn->lastPacketSize;
4275 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4278 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4279 if (!conn_data_locked) {
4280 MUTEX_ENTER(&conn->conn_data_lock);
4281 conn_data_locked = 1;
4283 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4284 /* process mtu ping ack */
4285 pktsize = conn->lastPingSize;
4286 conn->lastPingSizeSer = conn->lastPingSize = 0;
4290 if (conn_data_locked) {
4291 MUTEX_EXIT(&conn->conn_data_lock);
4292 conn_data_locked = 0;
4296 if (rxdebug_active) {
4300 len = _snprintf(msg, sizeof(msg),
4301 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4302 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4303 ntohl(ap->serial), ntohl(ap->previousPacket),
4304 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4305 ap->nAcks, ntohs(ap->bufferSpace) );
4309 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4310 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4314 OutputDebugString(msg);
4316 #else /* AFS_NT40_ENV */
4319 "RACK: reason %x previous %u seq %u serial %u first %u",
4320 ap->reason, ntohl(ap->previousPacket),
4321 (unsigned int)np->header.seq, (unsigned int)serial,
4322 ntohl(ap->firstPacket));
4325 for (offset = 0; offset < nAcks; offset++)
4326 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4331 #endif /* AFS_NT40_ENV */
4334 MUTEX_ENTER(&peer->peer_lock);
4337 * Start somewhere. Can't assume we can send what we can receive,
4338 * but we are clearly receiving.
4340 if (!peer->maxPacketSize)
4341 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4343 if (pktsize > peer->maxPacketSize) {
4344 peer->maxPacketSize = pktsize;
4345 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4346 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4347 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4348 rxi_ScheduleGrowMTUEvent(call, 1);
4353 clock_GetTime(&now);
4355 /* The transmit queue splits into 4 sections.
4357 * The first section is packets which have now been acknowledged
4358 * by a window size change in the ack. These have reached the
4359 * application layer, and may be discarded. These are packets
4360 * with sequence numbers < ap->firstPacket.
4362 * The second section is packets which have sequence numbers in
4363 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4364 * contents of the packet's ack array determines whether these
4365 * packets are acknowledged or not.
4367 * The third section is packets which fall above the range
4368 * addressed in the ack packet. These have not yet been received
4371 * The four section is packets which have not yet been transmitted.
4372 * These packets will have a header.serial of 0.
4375 /* First section - implicitly acknowledged packets that can be
4379 tp = queue_First(&call->tq, rx_packet);
4380 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4381 struct rx_packet *next;
4383 next = queue_Next(tp, rx_packet);
4384 call->tfirst = tp->header.seq + 1;
4386 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4388 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4391 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4392 /* XXX Hack. Because we have to release the global rx lock when sending
4393 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4394 * in rxi_Start sending packets out because packets may move to the
4395 * freePacketQueue as result of being here! So we drop these packets until
4396 * we're safely out of the traversing. Really ugly!
4397 * To make it even uglier, if we're using fine grain locking, we can
4398 * set the ack bits in the packets and have rxi_Start remove the packets
4399 * when it's done transmitting.
4401 if (call->flags & RX_CALL_TQ_BUSY) {
4402 #ifdef RX_ENABLE_LOCKS
4403 tp->flags |= RX_PKTFLAG_ACKED;
4404 call->flags |= RX_CALL_TQ_SOME_ACKED;
4405 #else /* RX_ENABLE_LOCKS */
4407 #endif /* RX_ENABLE_LOCKS */
4409 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4412 #ifdef RX_TRACK_PACKETS
4413 tp->flags &= ~RX_PKTFLAG_TQ;
4415 #ifdef RXDEBUG_PACKET
4417 #endif /* RXDEBUG_PACKET */
4418 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4423 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4425 /* Second section of the queue - packets for which we are receiving
4428 * Go through the explicit acks/nacks and record the results in
4429 * the waiting packets. These are packets that can't be released
4430 * yet, even with a positive acknowledge. This positive
4431 * acknowledge only means the packet has been received by the
4432 * peer, not that it will be retained long enough to be sent to
4433 * the peer's upper level. In addition, reset the transmit timers
4434 * of any missing packets (those packets that must be missing
4435 * because this packet was out of sequence) */
4437 call->nSoftAcked = 0;
4439 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4440 /* Set the acknowledge flag per packet based on the
4441 * information in the ack packet. An acknowlegded packet can
4442 * be downgraded when the server has discarded a packet it
4443 * soacked previously, or when an ack packet is received
4444 * out of sequence. */
4445 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4446 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4448 tp->flags |= RX_PKTFLAG_ACKED;
4449 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4456 } else /* RX_ACK_TYPE_NACK */ {
4457 tp->flags &= ~RX_PKTFLAG_ACKED;
4461 tp = queue_Next(tp, rx_packet);
4464 /* We don't need to take any action with the 3rd or 4th section in the
4465 * queue - they're not addressed by the contents of this ACK packet.
4468 /* If the window has been extended by this acknowledge packet,
4469 * then wakeup a sender waiting in alloc for window space, or try
4470 * sending packets now, if he's been sitting on packets due to
4471 * lack of window space */
4472 if (call->tnext < (call->tfirst + call->twind)) {
4473 #ifdef RX_ENABLE_LOCKS
4474 CV_SIGNAL(&call->cv_twind);
4476 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4477 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4478 osi_rxWakeup(&call->twind);
4481 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4482 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4486 /* if the ack packet has a receivelen field hanging off it,
4487 * update our state */
4488 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4491 /* If the ack packet has a "recommended" size that is less than
4492 * what I am using now, reduce my size to match */
4493 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4494 (int)sizeof(afs_int32), &tSize);
4495 tSize = (afs_uint32) ntohl(tSize);
4496 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4498 /* Get the maximum packet size to send to this peer */
4499 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4501 tSize = (afs_uint32) ntohl(tSize);
4502 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4503 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4505 /* sanity check - peer might have restarted with different params.
4506 * If peer says "send less", dammit, send less... Peer should never
4507 * be unable to accept packets of the size that prior AFS versions would
4508 * send without asking. */
4509 if (peer->maxMTU != tSize) {
4510 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4512 peer->maxMTU = tSize;
4513 peer->MTU = MIN(tSize, peer->MTU);
4514 call->MTU = MIN(call->MTU, tSize);
4517 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4520 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4521 (int)sizeof(afs_int32), &tSize);
4522 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4523 if (tSize < call->twind) { /* smaller than our send */
4524 call->twind = tSize; /* window, we must send less... */
4525 call->ssthresh = MIN(call->twind, call->ssthresh);
4526 call->conn->twind[call->channel] = call->twind;
4529 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4530 * network MTU confused with the loopback MTU. Calculate the
4531 * maximum MTU here for use in the slow start code below.
4533 /* Did peer restart with older RX version? */
4534 if (peer->maxDgramPackets > 1) {
4535 peer->maxDgramPackets = 1;
4537 } else if (np->length >=
4538 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4541 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4542 sizeof(afs_int32), &tSize);
4543 tSize = (afs_uint32) ntohl(tSize);
4545 * As of AFS 3.5 we set the send window to match the receive window.
4547 if (tSize < call->twind) {
4548 call->twind = tSize;
4549 call->conn->twind[call->channel] = call->twind;
4550 call->ssthresh = MIN(call->twind, call->ssthresh);
4551 } else if (tSize > call->twind) {
4552 call->twind = tSize;
4553 call->conn->twind[call->channel] = call->twind;
4557 * As of AFS 3.5, a jumbogram is more than one fixed size
4558 * packet transmitted in a single UDP datagram. If the remote
4559 * MTU is smaller than our local MTU then never send a datagram
4560 * larger than the natural MTU.
4563 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4564 (int)sizeof(afs_int32), &tSize);
4565 maxDgramPackets = (afs_uint32) ntohl(tSize);
4566 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4568 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4569 if (maxDgramPackets > 1) {
4570 peer->maxDgramPackets = maxDgramPackets;
4571 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4573 peer->maxDgramPackets = 1;
4574 call->MTU = peer->natMTU;
4576 } else if (peer->maxDgramPackets > 1) {
4577 /* Restarted with lower version of RX */
4578 peer->maxDgramPackets = 1;
4580 } else if (peer->maxDgramPackets > 1
4581 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4582 /* Restarted with lower version of RX */
4583 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4584 peer->natMTU = OLD_MAX_PACKET_SIZE;
4585 peer->MTU = OLD_MAX_PACKET_SIZE;
4586 peer->maxDgramPackets = 1;
4587 peer->nDgramPackets = 1;
4589 call->MTU = OLD_MAX_PACKET_SIZE;
4594 * Calculate how many datagrams were successfully received after
4595 * the first missing packet and adjust the negative ack counter
4600 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4601 if (call->nNacks < nNacked) {
4602 call->nNacks = nNacked;
4605 call->nAcks += newAckCount;
4609 /* If the packet contained new acknowledgements, rather than just
4610 * being a duplicate of one we have previously seen, then we can restart
4613 if (newAckCount > 0)
4614 rxi_rto_packet_acked(call, istack);
4616 if (call->flags & RX_CALL_FAST_RECOVER) {
4617 if (newAckCount == 0) {
4618 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4620 call->flags &= ~RX_CALL_FAST_RECOVER;
4621 call->cwind = call->nextCwind;
4622 call->nextCwind = 0;
4625 call->nCwindAcks = 0;
4626 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4627 /* Three negative acks in a row trigger congestion recovery */
4628 call->flags |= RX_CALL_FAST_RECOVER;
4629 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4631 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4632 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4633 call->nextCwind = call->ssthresh;
4636 peer->MTU = call->MTU;
4637 peer->cwind = call->nextCwind;
4638 peer->nDgramPackets = call->nDgramPackets;
4640 call->congestSeq = peer->congestSeq;
4642 /* Reset the resend times on the packets that were nacked
4643 * so we will retransmit as soon as the window permits
4646 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4648 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4649 tp->flags &= ~RX_PKTFLAG_SENT;
4651 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4656 /* If cwind is smaller than ssthresh, then increase
4657 * the window one packet for each ack we receive (exponential
4659 * If cwind is greater than or equal to ssthresh then increase
4660 * the congestion window by one packet for each cwind acks we
4661 * receive (linear growth). */
4662 if (call->cwind < call->ssthresh) {
4664 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4665 call->nCwindAcks = 0;
4667 call->nCwindAcks += newAckCount;
4668 if (call->nCwindAcks >= call->cwind) {
4669 call->nCwindAcks = 0;
4670 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4674 * If we have received several acknowledgements in a row then
4675 * it is time to increase the size of our datagrams
4677 if ((int)call->nAcks > rx_nDgramThreshold) {
4678 if (peer->maxDgramPackets > 1) {
4679 if (call->nDgramPackets < peer->maxDgramPackets) {
4680 call->nDgramPackets++;
4682 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4683 } else if (call->MTU < peer->maxMTU) {
4684 /* don't upgrade if we can't handle it */
4685 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4686 call->MTU = peer->ifMTU;
4688 call->MTU += peer->natMTU;
4689 call->MTU = MIN(call->MTU, peer->maxMTU);
4696 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4698 /* Servers need to hold the call until all response packets have
4699 * been acknowledged. Soft acks are good enough since clients
4700 * are not allowed to clear their receive queues. */
4701 if (call->state == RX_STATE_HOLD
4702 && call->tfirst + call->nSoftAcked >= call->tnext) {
4703 call->state = RX_STATE_DALLY;
4704 rxi_ClearTransmitQueue(call, 0);
4705 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4706 } else if (!queue_IsEmpty(&call->tq)) {
4707 rxi_Start(call, istack);
4712 /* Received a response to a challenge packet */
4713 static struct rx_packet *
4714 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4715 struct rx_packet *np, int istack)
4719 /* Ignore the packet if we're the client */
4720 if (conn->type == RX_CLIENT_CONNECTION)
4723 /* If already authenticated, ignore the packet (it's probably a retry) */
4724 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4727 /* Otherwise, have the security object evaluate the response packet */
4728 error = RXS_CheckResponse(conn->securityObject, conn, np);
4730 /* If the response is invalid, reset the connection, sending
4731 * an abort to the peer */
4735 rxi_ConnectionError(conn, error);
4736 MUTEX_ENTER(&conn->conn_data_lock);
4737 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4738 MUTEX_EXIT(&conn->conn_data_lock);
4741 /* If the response is valid, any calls waiting to attach
4742 * servers can now do so */
4745 for (i = 0; i < RX_MAXCALLS; i++) {
4746 struct rx_call *call = conn->call[i];
4748 MUTEX_ENTER(&call->lock);
4749 if (call->state == RX_STATE_PRECALL)
4750 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4751 /* tnop can be null if newcallp is null */
4752 MUTEX_EXIT(&call->lock);
4756 /* Update the peer reachability information, just in case
4757 * some calls went into attach-wait while we were waiting
4758 * for authentication..
4760 rxi_UpdatePeerReach(conn, NULL);
4765 /* A client has received an authentication challenge: the security
4766 * object is asked to cough up a respectable response packet to send
4767 * back to the server. The server is responsible for retrying the
4768 * challenge if it fails to get a response. */
4770 static struct rx_packet *
4771 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4772 struct rx_packet *np, int istack)
4776 /* Ignore the challenge if we're the server */
4777 if (conn->type == RX_SERVER_CONNECTION)
4780 /* Ignore the challenge if the connection is otherwise idle; someone's
4781 * trying to use us as an oracle. */
4782 if (!rxi_HasActiveCalls(conn))
4785 /* Send the security object the challenge packet. It is expected to fill
4786 * in the response. */
4787 error = RXS_GetResponse(conn->securityObject, conn, np);
4789 /* If the security object is unable to return a valid response, reset the
4790 * connection and send an abort to the peer. Otherwise send the response
4791 * packet to the peer connection. */
4793 rxi_ConnectionError(conn, error);
4794 MUTEX_ENTER(&conn->conn_data_lock);
4795 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4796 MUTEX_EXIT(&conn->conn_data_lock);
4798 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4799 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4805 /* Find an available server process to service the current request in
4806 * the given call structure. If one isn't available, queue up this
4807 * call so it eventually gets one */
4809 rxi_AttachServerProc(struct rx_call *call,
4810 osi_socket socket, int *tnop,
4811 struct rx_call **newcallp)
4813 struct rx_serverQueueEntry *sq;
4814 struct rx_service *service = call->conn->service;
4817 /* May already be attached */
4818 if (call->state == RX_STATE_ACTIVE)
4821 MUTEX_ENTER(&rx_serverPool_lock);
4823 haveQuota = QuotaOK(service);
4824 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4825 /* If there are no processes available to service this call,
4826 * put the call on the incoming call queue (unless it's
4827 * already on the queue).
4829 #ifdef RX_ENABLE_LOCKS
4831 ReturnToServerPool(service);
4832 #endif /* RX_ENABLE_LOCKS */
4834 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4835 call->flags |= RX_CALL_WAIT_PROC;
4836 rx_atomic_inc(&rx_nWaiting);
4837 rx_atomic_inc(&rx_nWaited);
4838 rxi_calltrace(RX_CALL_ARRIVAL, call);
4839 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4840 queue_Append(&rx_incomingCallQueue, call);
4843 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4845 /* If hot threads are enabled, and both newcallp and sq->socketp
4846 * are non-null, then this thread will process the call, and the
4847 * idle server thread will start listening on this threads socket.
4850 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4853 *sq->socketp = socket;
4854 clock_GetTime(&call->startTime);
4855 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4859 if (call->flags & RX_CALL_WAIT_PROC) {
4860 /* Conservative: I don't think this should happen */
4861 call->flags &= ~RX_CALL_WAIT_PROC;
4862 rx_atomic_dec(&rx_nWaiting);
4863 if (queue_IsOnQueue(call)) {
4867 call->state = RX_STATE_ACTIVE;
4868 call->mode = RX_MODE_RECEIVING;
4869 #ifdef RX_KERNEL_TRACE
4871 int glockOwner = ISAFS_GLOCK();
4874 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4875 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4881 if (call->flags & RX_CALL_CLEARED) {
4882 /* send an ack now to start the packet flow up again */
4883 call->flags &= ~RX_CALL_CLEARED;
4884 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4886 #ifdef RX_ENABLE_LOCKS
4889 service->nRequestsRunning++;
4890 MUTEX_ENTER(&rx_quota_mutex);
4891 if (service->nRequestsRunning <= service->minProcs)
4894 MUTEX_EXIT(&rx_quota_mutex);
4898 MUTEX_EXIT(&rx_serverPool_lock);
4901 /* Delay the sending of an acknowledge event for a short while, while
4902 * a new call is being prepared (in the case of a client) or a reply
4903 * is being prepared (in the case of a server). Rather than sending
4904 * an ack packet, an ACKALL packet is sent. */
4906 rxi_AckAll(struct rx_call *call)
4908 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4910 call->flags |= RX_CALL_ACKALL_SENT;
4914 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4917 struct rx_call *call = arg1;
4918 #ifdef RX_ENABLE_LOCKS
4920 MUTEX_ENTER(&call->lock);
4921 if (event == call->delayedAckEvent) {
4922 rxevent_Put(call->delayedAckEvent);
4923 call->delayedAckEvent = NULL;
4925 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4927 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4929 MUTEX_EXIT(&call->lock);
4930 #else /* RX_ENABLE_LOCKS */
4932 rxevent_Put(call->delayedAckEvent);
4933 call->delayedAckEvent = NULL;
4935 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4936 #endif /* RX_ENABLE_LOCKS */
4940 #ifdef RX_ENABLE_LOCKS
4941 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4942 * clearing them out.
4945 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4947 struct rx_packet *p, *tp;
4950 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4951 p->flags |= RX_PKTFLAG_ACKED;
4955 call->flags |= RX_CALL_TQ_CLEARME;
4956 call->flags |= RX_CALL_TQ_SOME_ACKED;
4959 rxi_rto_cancel(call);
4961 call->tfirst = call->tnext;
4962 call->nSoftAcked = 0;
4964 if (call->flags & RX_CALL_FAST_RECOVER) {
4965 call->flags &= ~RX_CALL_FAST_RECOVER;
4966 call->cwind = call->nextCwind;
4967 call->nextCwind = 0;
4970 CV_SIGNAL(&call->cv_twind);
4972 #endif /* RX_ENABLE_LOCKS */
4974 /* Clear out the transmit queue for the current call (all packets have
4975 * been received by peer) */
4977 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4979 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4980 struct rx_packet *p, *tp;
4982 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4984 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4985 p->flags |= RX_PKTFLAG_ACKED;
4989 call->flags |= RX_CALL_TQ_CLEARME;
4990 call->flags |= RX_CALL_TQ_SOME_ACKED;
4993 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4994 #ifdef RXDEBUG_PACKET
4996 #endif /* RXDEBUG_PACKET */
4997 rxi_FreePackets(0, &call->tq);
4998 rxi_WakeUpTransmitQueue(call);
4999 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5000 call->flags &= ~RX_CALL_TQ_CLEARME;
5002 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5004 rxi_rto_cancel(call);
5005 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5006 call->nSoftAcked = 0;
5008 if (call->flags & RX_CALL_FAST_RECOVER) {
5009 call->flags &= ~RX_CALL_FAST_RECOVER;
5010 call->cwind = call->nextCwind;
5012 #ifdef RX_ENABLE_LOCKS
5013 CV_SIGNAL(&call->cv_twind);
5015 osi_rxWakeup(&call->twind);
5020 rxi_ClearReceiveQueue(struct rx_call *call)
5022 if (queue_IsNotEmpty(&call->rq)) {
5025 count = rxi_FreePackets(0, &call->rq);
5026 rx_packetReclaims += count;
5027 #ifdef RXDEBUG_PACKET
5029 if ( call->rqc != 0 )
5030 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5032 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5034 if (call->state == RX_STATE_PRECALL) {
5035 call->flags |= RX_CALL_CLEARED;
5039 /* Send an abort packet for the specified call */
5040 static struct rx_packet *
5041 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5042 int istack, int force)
5044 afs_int32 error, cerror;
5045 struct clock when, now;
5050 switch (call->error) {
5053 cerror = RX_CALL_TIMEOUT;
5056 cerror = call->error;
5059 /* Clients should never delay abort messages */
5060 if (rx_IsClientConn(call->conn))
5063 if (call->abortCode != cerror) {
5064 call->abortCode = cerror;
5065 call->abortCount = 0;
5068 if (force || rxi_callAbortThreshhold == 0
5069 || call->abortCount < rxi_callAbortThreshhold) {
5070 if (call->delayedAbortEvent) {
5071 rxevent_Cancel(&call->delayedAbortEvent, call,
5072 RX_CALL_REFCOUNT_ABORT);
5074 error = htonl(cerror);
5077 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5078 (char *)&error, sizeof(error), istack);
5079 } else if (!call->delayedAbortEvent) {
5080 clock_GetTime(&now);
5082 clock_Addmsec(&when, rxi_callAbortDelay);
5083 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5084 call->delayedAbortEvent =
5085 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5090 /* Send an abort packet for the specified connection. Packet is an
5091 * optional pointer to a packet that can be used to send the abort.
5092 * Once the number of abort messages reaches the threshhold, an
5093 * event is scheduled to send the abort. Setting the force flag
5094 * overrides sending delayed abort messages.
5096 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5097 * to send the abort packet.
5100 rxi_SendConnectionAbort(struct rx_connection *conn,
5101 struct rx_packet *packet, int istack, int force)
5104 struct clock when, now;
5109 /* Clients should never delay abort messages */
5110 if (rx_IsClientConn(conn))
5113 if (force || rxi_connAbortThreshhold == 0
5114 || conn->abortCount < rxi_connAbortThreshhold) {
5116 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5117 error = htonl(conn->error);
5119 MUTEX_EXIT(&conn->conn_data_lock);
5121 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5122 RX_PACKET_TYPE_ABORT, (char *)&error,
5123 sizeof(error), istack);
5124 MUTEX_ENTER(&conn->conn_data_lock);
5125 } else if (!conn->delayedAbortEvent) {
5126 clock_GetTime(&now);
5128 clock_Addmsec(&when, rxi_connAbortDelay);
5129 conn->delayedAbortEvent =
5130 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5135 /* Associate an error all of the calls owned by a connection. Called
5136 * with error non-zero. This is only for really fatal things, like
5137 * bad authentication responses. The connection itself is set in
5138 * error at this point, so that future packets received will be
5141 rxi_ConnectionError(struct rx_connection *conn,
5147 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5149 MUTEX_ENTER(&conn->conn_data_lock);
5150 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5151 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5152 if (conn->checkReachEvent) {
5153 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5154 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5155 putConnection(conn);
5157 MUTEX_EXIT(&conn->conn_data_lock);
5158 for (i = 0; i < RX_MAXCALLS; i++) {
5159 struct rx_call *call = conn->call[i];
5161 MUTEX_ENTER(&call->lock);
5162 rxi_CallError(call, error);
5163 MUTEX_EXIT(&call->lock);
5166 conn->error = error;
5167 if (rx_stats_active)
5168 rx_atomic_inc(&rx_stats.fatalErrors);
5173 * Interrupt an in-progress call with the specified error and wakeup waiters.
5175 * @param[in] call The call to interrupt
5176 * @param[in] error The error code to send to the peer
5179 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5181 MUTEX_ENTER(&call->lock);
5182 rxi_CallError(call, error);
5183 rxi_SendCallAbort(call, NULL, 0, 1);
5184 MUTEX_EXIT(&call->lock);
5188 rxi_CallError(struct rx_call *call, afs_int32 error)
5191 osirx_AssertMine(&call->lock, "rxi_CallError");
5193 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5195 error = call->error;
5197 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5198 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5199 rxi_ResetCall(call, 0);
5202 rxi_ResetCall(call, 0);
5204 call->error = error;
5207 /* Reset various fields in a call structure, and wakeup waiting
5208 * processes. Some fields aren't changed: state & mode are not
5209 * touched (these must be set by the caller), and bufptr, nLeft, and
5210 * nFree are not reset, since these fields are manipulated by
5211 * unprotected macros, and may only be reset by non-interrupting code.
5215 rxi_ResetCall(struct rx_call *call, int newcall)
5218 struct rx_peer *peer;
5219 struct rx_packet *packet;
5221 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5223 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5225 /* Notify anyone who is waiting for asynchronous packet arrival */
5226 if (call->arrivalProc) {
5227 (*call->arrivalProc) (call, call->arrivalProcHandle,
5228 call->arrivalProcArg);
5229 call->arrivalProc = (void (*)())0;
5233 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5235 if (call->delayedAbortEvent) {
5236 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5237 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5239 rxi_SendCallAbort(call, packet, 0, 1);
5240 rxi_FreePacket(packet);
5245 * Update the peer with the congestion information in this call
5246 * so other calls on this connection can pick up where this call
5247 * left off. If the congestion sequence numbers don't match then
5248 * another call experienced a retransmission.
5250 peer = call->conn->peer;
5251 MUTEX_ENTER(&peer->peer_lock);
5253 if (call->congestSeq == peer->congestSeq) {
5254 peer->cwind = MAX(peer->cwind, call->cwind);
5255 peer->MTU = MAX(peer->MTU, call->MTU);
5256 peer->nDgramPackets =
5257 MAX(peer->nDgramPackets, call->nDgramPackets);
5260 call->abortCode = 0;
5261 call->abortCount = 0;
5263 if (peer->maxDgramPackets > 1) {
5264 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5266 call->MTU = peer->MTU;
5268 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5269 call->ssthresh = rx_maxSendWindow;
5270 call->nDgramPackets = peer->nDgramPackets;
5271 call->congestSeq = peer->congestSeq;
5272 call->rtt = peer->rtt;
5273 call->rtt_dev = peer->rtt_dev;
5274 clock_Zero(&call->rto);
5275 clock_Addmsec(&call->rto,
5276 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5277 MUTEX_EXIT(&peer->peer_lock);
5279 flags = call->flags;
5280 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5281 rxi_WaitforTQBusy(call);
5282 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5284 rxi_ClearTransmitQueue(call, 1);
5285 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5286 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5290 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5291 /* The call channel is still busy; resetting the call doesn't change
5292 * that. However, if 'newcall' is set, we are processing a call
5293 * structure that has either been recycled from the free list, or has
5294 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5295 * 'newcall' is set, since it describes a completely different call
5296 * channel which we do not care about. */
5297 call->flags |= RX_CALL_PEER_BUSY;
5300 rxi_ClearReceiveQueue(call);
5301 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5305 call->twind = call->conn->twind[call->channel];
5306 call->rwind = call->conn->rwind[call->channel];
5307 call->nSoftAcked = 0;
5308 call->nextCwind = 0;
5311 call->nCwindAcks = 0;
5312 call->nSoftAcks = 0;
5313 call->nHardAcks = 0;
5315 call->tfirst = call->rnext = call->tnext = 1;
5318 call->lastAcked = 0;
5319 call->localStatus = call->remoteStatus = 0;
5321 if (flags & RX_CALL_READER_WAIT) {
5322 #ifdef RX_ENABLE_LOCKS
5323 CV_BROADCAST(&call->cv_rq);
5325 osi_rxWakeup(&call->rq);
5328 if (flags & RX_CALL_WAIT_PACKETS) {
5329 MUTEX_ENTER(&rx_freePktQ_lock);
5330 rxi_PacketsUnWait(); /* XXX */
5331 MUTEX_EXIT(&rx_freePktQ_lock);
5333 #ifdef RX_ENABLE_LOCKS
5334 CV_SIGNAL(&call->cv_twind);
5336 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5337 osi_rxWakeup(&call->twind);
5340 if (flags & RX_CALL_WAIT_PROC) {
5341 rx_atomic_dec(&rx_nWaiting);
5343 #ifdef RX_ENABLE_LOCKS
5344 /* The following ensures that we don't mess with any queue while some
5345 * other thread might also be doing so. The call_queue_lock field is
5346 * is only modified under the call lock. If the call is in the process
5347 * of being removed from a queue, the call is not locked until the
5348 * the queue lock is dropped and only then is the call_queue_lock field
5349 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5350 * Note that any other routine which removes a call from a queue has to
5351 * obtain the queue lock before examing the queue and removing the call.
5353 if (call->call_queue_lock) {
5354 MUTEX_ENTER(call->call_queue_lock);
5355 if (queue_IsOnQueue(call)) {
5358 MUTEX_EXIT(call->call_queue_lock);
5359 CLEAR_CALL_QUEUE_LOCK(call);
5361 #else /* RX_ENABLE_LOCKS */
5362 if (queue_IsOnQueue(call)) {
5365 #endif /* RX_ENABLE_LOCKS */
5367 rxi_KeepAliveOff(call);
5368 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5371 /* Send an acknowledge for the indicated packet (seq,serial) of the
5372 * indicated call, for the indicated reason (reason). This
5373 * acknowledge will specifically acknowledge receiving the packet, and
5374 * will also specify which other packets for this call have been
5375 * received. This routine returns the packet that was used to the
5376 * caller. The caller is responsible for freeing it or re-using it.
5377 * This acknowledgement also returns the highest sequence number
5378 * actually read out by the higher level to the sender; the sender
5379 * promises to keep around packets that have not been read by the
5380 * higher level yet (unless, of course, the sender decides to abort
5381 * the call altogether). Any of p, seq, serial, pflags, or reason may
5382 * be set to zero without ill effect. That is, if they are zero, they
5383 * will not convey any information.
5384 * NOW there is a trailer field, after the ack where it will safely be
5385 * ignored by mundanes, which indicates the maximum size packet this
5386 * host can swallow. */
5388 struct rx_packet *optionalPacket; use to send ack (or null)
5389 int seq; Sequence number of the packet we are acking
5390 int serial; Serial number of the packet
5391 int pflags; Flags field from packet header
5392 int reason; Reason an acknowledge was prompted
5396 rxi_SendAck(struct rx_call *call,
5397 struct rx_packet *optionalPacket, int serial, int reason,
5400 struct rx_ackPacket *ap;
5401 struct rx_packet *rqp;
5402 struct rx_packet *nxp; /* For queue_Scan */
5403 struct rx_packet *p;
5406 afs_uint32 padbytes = 0;
5407 #ifdef RX_ENABLE_TSFPQ
5408 struct rx_ts_info_t * rx_ts_info;
5412 * Open the receive window once a thread starts reading packets
5414 if (call->rnext > 1) {
5415 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5418 /* Don't attempt to grow MTU if this is a critical ping */
5419 if (reason == RX_ACK_MTU) {
5420 /* keep track of per-call attempts, if we're over max, do in small
5421 * otherwise in larger? set a size to increment by, decrease
5424 if (call->conn->peer->maxPacketSize &&
5425 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5427 padbytes = call->conn->peer->maxPacketSize+16;
5429 padbytes = call->conn->peer->maxMTU + 128;
5431 /* do always try a minimum size ping */
5432 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5434 /* subtract the ack payload */
5435 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5436 reason = RX_ACK_PING;
5439 call->nHardAcks = 0;
5440 call->nSoftAcks = 0;
5441 if (call->rnext > call->lastAcked)
5442 call->lastAcked = call->rnext;
5446 rx_computelen(p, p->length); /* reset length, you never know */
5447 } /* where that's been... */
5448 #ifdef RX_ENABLE_TSFPQ
5450 RX_TS_INFO_GET(rx_ts_info);
5451 if ((p = rx_ts_info->local_special_packet)) {
5452 rx_computelen(p, p->length);
5453 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5454 rx_ts_info->local_special_packet = p;
5455 } else { /* We won't send the ack, but don't panic. */
5456 return optionalPacket;
5460 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5461 /* We won't send the ack, but don't panic. */
5462 return optionalPacket;
5467 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5470 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5471 #ifndef RX_ENABLE_TSFPQ
5472 if (!optionalPacket)
5475 return optionalPacket;
5477 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5478 if (rx_Contiguous(p) < templ) {
5479 #ifndef RX_ENABLE_TSFPQ
5480 if (!optionalPacket)
5483 return optionalPacket;
5488 /* MTUXXX failing to send an ack is very serious. We should */
5489 /* try as hard as possible to send even a partial ack; it's */
5490 /* better than nothing. */
5491 ap = (struct rx_ackPacket *)rx_DataOf(p);
5492 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5493 ap->reason = reason;
5495 /* The skew computation used to be bogus, I think it's better now. */
5496 /* We should start paying attention to skew. XXX */
5497 ap->serial = htonl(serial);
5498 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5501 * First packet not yet forwarded to reader. When ACKALL has been
5502 * sent the peer has been told that all received packets will be
5503 * delivered to the reader. The value 'rnext' is used internally
5504 * to refer to the next packet in the receive queue that must be
5505 * delivered to the reader. From the perspective of the peer it
5506 * already has so report the last sequence number plus one if there
5507 * are packets in the receive queue awaiting processing.
5509 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5510 !queue_IsEmpty(&call->rq)) {
5511 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5513 ap->firstPacket = htonl(call->rnext);
5515 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5517 /* No fear of running out of ack packet here because there can only be at most
5518 * one window full of unacknowledged packets. The window size must be constrained
5519 * to be less than the maximum ack size, of course. Also, an ack should always
5520 * fit into a single packet -- it should not ever be fragmented. */
5521 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5522 if (!rqp || !call->rq.next
5523 || (rqp->header.seq > (call->rnext + call->rwind))) {
5524 #ifndef RX_ENABLE_TSFPQ
5525 if (!optionalPacket)
5528 rxi_CallError(call, RX_CALL_DEAD);
5529 return optionalPacket;
5532 while (rqp->header.seq > call->rnext + offset)
5533 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5534 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5536 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5537 #ifndef RX_ENABLE_TSFPQ
5538 if (!optionalPacket)
5541 rxi_CallError(call, RX_CALL_DEAD);
5542 return optionalPacket;
5548 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5550 /* these are new for AFS 3.3 */
5551 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5552 templ = htonl(templ);
5553 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5554 templ = htonl(call->conn->peer->ifMTU);
5555 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5556 sizeof(afs_int32), &templ);
5558 /* new for AFS 3.4 */
5559 templ = htonl(call->rwind);
5560 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5561 sizeof(afs_int32), &templ);
5563 /* new for AFS 3.5 */
5564 templ = htonl(call->conn->peer->ifDgramPackets);
5565 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5566 sizeof(afs_int32), &templ);
5568 p->header.serviceId = call->conn->serviceId;
5569 p->header.cid = (call->conn->cid | call->channel);
5570 p->header.callNumber = *call->callNumber;
5572 p->header.securityIndex = call->conn->securityIndex;
5573 p->header.epoch = call->conn->epoch;
5574 p->header.type = RX_PACKET_TYPE_ACK;
5575 p->header.flags = RX_SLOW_START_OK;
5576 if (reason == RX_ACK_PING) {
5577 p->header.flags |= RX_REQUEST_ACK;
5579 p->length = padbytes +
5580 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5583 /* not fast but we can potentially use this if truncated
5584 * fragments are delivered to figure out the mtu.
5586 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5587 sizeof(afs_int32), sizeof(afs_int32),
5591 if (call->conn->type == RX_CLIENT_CONNECTION)
5592 p->header.flags |= RX_CLIENT_INITIATED;
5596 if (rxdebug_active) {
5600 len = _snprintf(msg, sizeof(msg),
5601 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5602 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5603 ntohl(ap->serial), ntohl(ap->previousPacket),
5604 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5605 ap->nAcks, ntohs(ap->bufferSpace) );
5609 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5610 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5614 OutputDebugString(msg);
5616 #else /* AFS_NT40_ENV */
5618 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5619 ap->reason, ntohl(ap->previousPacket),
5620 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5622 for (offset = 0; offset < ap->nAcks; offset++)
5623 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5628 #endif /* AFS_NT40_ENV */
5631 int i, nbytes = p->length;
5633 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5634 if (nbytes <= p->wirevec[i].iov_len) {
5637 savelen = p->wirevec[i].iov_len;
5639 p->wirevec[i].iov_len = nbytes;
5641 rxi_Send(call, p, istack);
5642 p->wirevec[i].iov_len = savelen;
5646 nbytes -= p->wirevec[i].iov_len;
5649 if (rx_stats_active)
5650 rx_atomic_inc(&rx_stats.ackPacketsSent);
5651 #ifndef RX_ENABLE_TSFPQ
5652 if (!optionalPacket)
5655 return optionalPacket; /* Return packet for re-use by caller */
5659 struct rx_packet **list;
5664 /* Send all of the packets in the list in single datagram */
5666 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5667 int istack, int moreFlag)
5673 struct rx_connection *conn = call->conn;
5674 struct rx_peer *peer = conn->peer;
5676 MUTEX_ENTER(&peer->peer_lock);
5677 peer->nSent += xmit->len;
5678 if (xmit->resending)
5679 peer->reSends += xmit->len;
5680 MUTEX_EXIT(&peer->peer_lock);
5682 if (rx_stats_active) {
5683 if (xmit->resending)
5684 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5686 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5689 clock_GetTime(&now);
5691 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5695 /* Set the packet flags and schedule the resend events */
5696 /* Only request an ack for the last packet in the list */
5697 for (i = 0; i < xmit->len; i++) {
5698 struct rx_packet *packet = xmit->list[i];
5700 /* Record the time sent */
5701 packet->timeSent = now;
5702 packet->flags |= RX_PKTFLAG_SENT;
5704 /* Ask for an ack on retransmitted packets, on every other packet
5705 * if the peer doesn't support slow start. Ask for an ack on every
5706 * packet until the congestion window reaches the ack rate. */
5707 if (packet->header.serial) {
5710 packet->firstSent = now;
5711 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5712 || (!(call->flags & RX_CALL_SLOW_START_OK)
5713 && (packet->header.seq & 1)))) {
5718 /* Tag this packet as not being the last in this group,
5719 * for the receiver's benefit */
5720 if (i < xmit->len - 1 || moreFlag) {
5721 packet->header.flags |= RX_MORE_PACKETS;
5726 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5729 /* Since we're about to send a data packet to the peer, it's
5730 * safe to nuke any scheduled end-of-packets ack */
5731 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5733 MUTEX_EXIT(&call->lock);
5734 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5735 if (xmit->len > 1) {
5736 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5738 rxi_SendPacket(call, conn, xmit->list[0], istack);
5740 MUTEX_ENTER(&call->lock);
5741 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5743 /* Tell the RTO calculation engine that we have sent a packet, and
5744 * if it was the last one */
5745 rxi_rto_packet_sent(call, lastPacket, istack);
5747 /* Update last send time for this call (for keep-alive
5748 * processing), and for the connection (so that we can discover
5749 * idle connections) */
5750 conn->lastSendTime = call->lastSendTime = clock_Sec();
5751 /* Let a set of retransmits trigger an idle timeout */
5752 if (!xmit->resending)
5753 call->lastSendData = call->lastSendTime;
5756 /* When sending packets we need to follow these rules:
5757 * 1. Never send more than maxDgramPackets in a jumbogram.
5758 * 2. Never send a packet with more than two iovecs in a jumbogram.
5759 * 3. Never send a retransmitted packet in a jumbogram.
5760 * 4. Never send more than cwind/4 packets in a jumbogram
5761 * We always keep the last list we should have sent so we
5762 * can set the RX_MORE_PACKETS flags correctly.
5766 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5771 struct xmitlist working;
5772 struct xmitlist last;
5774 struct rx_peer *peer = call->conn->peer;
5775 int morePackets = 0;
5777 memset(&last, 0, sizeof(struct xmitlist));
5778 working.list = &list[0];
5780 working.resending = 0;
5782 recovery = call->flags & RX_CALL_FAST_RECOVER;
5784 for (i = 0; i < len; i++) {
5785 /* Does the current packet force us to flush the current list? */
5787 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5788 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5790 /* This sends the 'last' list and then rolls the current working
5791 * set into the 'last' one, and resets the working set */
5794 rxi_SendList(call, &last, istack, 1);
5795 /* If the call enters an error state stop sending, or if
5796 * we entered congestion recovery mode, stop sending */
5798 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5803 working.resending = 0;
5804 working.list = &list[i];
5806 /* Add the current packet to the list if it hasn't been acked.
5807 * Otherwise adjust the list pointer to skip the current packet. */
5808 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5811 if (list[i]->header.serial)
5812 working.resending = 1;
5814 /* Do we need to flush the list? */
5815 if (working.len >= (int)peer->maxDgramPackets
5816 || working.len >= (int)call->nDgramPackets
5817 || working.len >= (int)call->cwind
5818 || list[i]->header.serial
5819 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5821 rxi_SendList(call, &last, istack, 1);
5822 /* If the call enters an error state stop sending, or if
5823 * we entered congestion recovery mode, stop sending */
5825 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5830 working.resending = 0;
5831 working.list = &list[i + 1];
5834 if (working.len != 0) {
5835 osi_Panic("rxi_SendList error");
5837 working.list = &list[i + 1];
5841 /* Send the whole list when the call is in receive mode, when
5842 * the call is in eof mode, when we are in fast recovery mode,
5843 * and when we have the last packet */
5844 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5845 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5846 || (call->flags & RX_CALL_FAST_RECOVER)) {
5847 /* Check for the case where the current list contains
5848 * an acked packet. Since we always send retransmissions
5849 * in a separate packet, we only need to check the first
5850 * packet in the list */
5851 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5855 rxi_SendList(call, &last, istack, morePackets);
5856 /* If the call enters an error state stop sending, or if
5857 * we entered congestion recovery mode, stop sending */
5859 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5863 rxi_SendList(call, &working, istack, 0);
5865 } else if (last.len > 0) {
5866 rxi_SendList(call, &last, istack, 0);
5867 /* Packets which are in 'working' are not sent by this call */
5872 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5874 struct rx_call *call = arg0;
5875 struct rx_peer *peer;
5876 struct rx_packet *p, *nxp;
5877 struct clock maxTimeout = { 60, 0 };
5879 MUTEX_ENTER(&call->lock);
5881 peer = call->conn->peer;
5883 /* Make sure that the event pointer is removed from the call
5884 * structure, since there is no longer a per-call retransmission
5886 if (event == call->resendEvent) {
5887 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5888 rxevent_Put(call->resendEvent);
5889 call->resendEvent = NULL;
5892 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5893 rxi_CheckBusy(call);
5896 if (queue_IsEmpty(&call->tq)) {
5897 /* Nothing to do. This means that we've been raced, and that an
5898 * ACK has come in between when we were triggered, and when we
5899 * actually got to run. */
5903 /* We're in loss recovery */
5904 call->flags |= RX_CALL_FAST_RECOVER;
5906 /* Mark all of the pending packets in the queue as being lost */
5907 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5908 if (!(p->flags & RX_PKTFLAG_ACKED))
5909 p->flags &= ~RX_PKTFLAG_SENT;
5912 /* We're resending, so we double the timeout of the call. This will be
5913 * dropped back down by the first successful ACK that we receive.
5915 * We apply a maximum value here of 60 seconds
5917 clock_Add(&call->rto, &call->rto);
5918 if (clock_Gt(&call->rto, &maxTimeout))
5919 call->rto = maxTimeout;
5921 /* Packet loss is most likely due to congestion, so drop our window size
5922 * and start again from the beginning */
5923 if (peer->maxDgramPackets >1) {
5924 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5925 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5927 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5928 call->nDgramPackets = 1;
5930 call->nextCwind = 1;
5933 MUTEX_ENTER(&peer->peer_lock);
5934 peer->MTU = call->MTU;
5935 peer->cwind = call->cwind;
5936 peer->nDgramPackets = 1;
5938 call->congestSeq = peer->congestSeq;
5939 MUTEX_EXIT(&peer->peer_lock);
5941 rxi_Start(call, istack);
5944 MUTEX_EXIT(&call->lock);
5947 /* This routine is called when new packets are readied for
5948 * transmission and when retransmission may be necessary, or when the
5949 * transmission window or burst count are favourable. This should be
5950 * better optimized for new packets, the usual case, now that we've
5951 * got rid of queues of send packets. XXXXXXXXXXX */
5953 rxi_Start(struct rx_call *call, int istack)
5956 struct rx_packet *p;
5957 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5962 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5963 if (rx_stats_active)
5964 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5969 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5971 /* Send (or resend) any packets that need it, subject to
5972 * window restrictions and congestion burst control
5973 * restrictions. Ask for an ack on the last packet sent in
5974 * this burst. For now, we're relying upon the window being
5975 * considerably bigger than the largest number of packets that
5976 * are typically sent at once by one initial call to
5977 * rxi_Start. This is probably bogus (perhaps we should ask
5978 * for an ack when we're half way through the current
5979 * window?). Also, for non file transfer applications, this
5980 * may end up asking for an ack for every packet. Bogus. XXXX
5983 * But check whether we're here recursively, and let the other guy
5986 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5987 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5988 call->flags |= RX_CALL_TQ_BUSY;
5990 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5992 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5993 call->flags &= ~RX_CALL_NEED_START;
5994 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5996 maxXmitPackets = MIN(call->twind, call->cwind);
5997 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5998 #ifdef RX_TRACK_PACKETS
5999 if ((p->flags & RX_PKTFLAG_FREE)
6000 || (!queue_IsEnd(&call->tq, nxp)
6001 && (nxp->flags & RX_PKTFLAG_FREE))
6002 || (p == (struct rx_packet *)&rx_freePacketQueue)
6003 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6004 osi_Panic("rxi_Start: xmit queue clobbered");
6007 if (p->flags & RX_PKTFLAG_ACKED) {
6008 /* Since we may block, don't trust this */
6009 if (rx_stats_active)
6010 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6011 continue; /* Ignore this packet if it has been acknowledged */
6014 /* Turn off all flags except these ones, which are the same
6015 * on each transmission */
6016 p->header.flags &= RX_PRESET_FLAGS;
6018 if (p->header.seq >=
6019 call->tfirst + MIN((int)call->twind,
6020 (int)(call->nSoftAcked +
6022 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6023 /* Note: if we're waiting for more window space, we can
6024 * still send retransmits; hence we don't return here, but
6025 * break out to schedule a retransmit event */
6026 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6027 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6032 /* Transmit the packet if it needs to be sent. */
6033 if (!(p->flags & RX_PKTFLAG_SENT)) {
6034 if (nXmitPackets == maxXmitPackets) {
6035 rxi_SendXmitList(call, call->xmitList,
6036 nXmitPackets, istack);
6039 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6040 *(call->callNumber), p));
6041 call->xmitList[nXmitPackets++] = p;
6045 /* xmitList now hold pointers to all of the packets that are
6046 * ready to send. Now we loop to send the packets */
6047 if (nXmitPackets > 0) {
6048 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6052 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6054 /* We went into the error state while sending packets. Now is
6055 * the time to reset the call. This will also inform the using
6056 * process that the call is in an error state.
6058 if (rx_stats_active)
6059 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6060 call->flags &= ~RX_CALL_TQ_BUSY;
6061 rxi_WakeUpTransmitQueue(call);
6062 rxi_CallError(call, call->error);
6065 #ifdef RX_ENABLE_LOCKS
6066 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6068 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6069 /* Some packets have received acks. If they all have, we can clear
6070 * the transmit queue.
6073 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6074 if (p->header.seq < call->tfirst
6075 && (p->flags & RX_PKTFLAG_ACKED)) {
6077 #ifdef RX_TRACK_PACKETS
6078 p->flags &= ~RX_PKTFLAG_TQ;
6080 #ifdef RXDEBUG_PACKET
6088 call->flags |= RX_CALL_TQ_CLEARME;
6090 #endif /* RX_ENABLE_LOCKS */
6091 if (call->flags & RX_CALL_TQ_CLEARME)
6092 rxi_ClearTransmitQueue(call, 1);
6093 } while (call->flags & RX_CALL_NEED_START);
6095 * TQ references no longer protected by this flag; they must remain
6096 * protected by the global lock.
6098 call->flags &= ~RX_CALL_TQ_BUSY;
6099 rxi_WakeUpTransmitQueue(call);
6101 call->flags |= RX_CALL_NEED_START;
6103 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6105 rxi_rto_cancel(call);
6109 /* Also adjusts the keep alive parameters for the call, to reflect
6110 * that we have just sent a packet (so keep alives aren't sent
6113 rxi_Send(struct rx_call *call, struct rx_packet *p,
6116 struct rx_connection *conn = call->conn;
6118 /* Stamp each packet with the user supplied status */
6119 p->header.userStatus = call->localStatus;
6121 /* Allow the security object controlling this call's security to
6122 * make any last-minute changes to the packet */
6123 RXS_SendPacket(conn->securityObject, call, p);
6125 /* Since we're about to send SOME sort of packet to the peer, it's
6126 * safe to nuke any scheduled end-of-packets ack */
6127 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6129 /* Actually send the packet, filling in more connection-specific fields */
6130 MUTEX_EXIT(&call->lock);
6131 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6132 rxi_SendPacket(call, conn, p, istack);
6133 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6134 MUTEX_ENTER(&call->lock);
6136 /* Update last send time for this call (for keep-alive
6137 * processing), and for the connection (so that we can discover
6138 * idle connections) */
6139 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6140 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6141 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6143 conn->lastSendTime = call->lastSendTime = clock_Sec();
6144 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6145 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6146 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6147 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6148 RX_ACK_PING_RESPONSE)))
6149 call->lastSendData = call->lastSendTime;
6153 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6154 * that things are fine. Also called periodically to guarantee that nothing
6155 * falls through the cracks (e.g. (error + dally) connections have keepalive
6156 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6158 * haveCTLock Set if calling from rxi_ReapConnections
6160 #ifdef RX_ENABLE_LOCKS
6162 static rxi_CheckCall(struct rx_call *call, int haveCTLock)
6163 #else /* RX_ENABLE_LOCKS */
6165 static rxi_CheckCall(struct rx_call *call)
6166 #endif /* RX_ENABLE_LOCKS */
6168 struct rx_connection *conn = call->conn;
6170 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6171 afs_uint32 fudgeFactor;
6174 int idle_timeout = 0;
6175 afs_int32 clock_diff = 0;
6179 /* Large swings in the clock can have a significant impact on
6180 * the performance of RX call processing. Forward clock shifts
6181 * will result in premature event triggering or timeouts.
6182 * Backward shifts can result in calls not completing until
6183 * the clock catches up with the original start clock value.
6185 * If a backward clock shift of more than five minutes is noticed,
6186 * just fail the call.
6188 if (now < call->lastSendTime)
6189 clock_diff = call->lastSendTime - now;
6190 if (now < call->startWait)
6191 clock_diff = MAX(clock_diff, call->startWait - now);
6192 if (now < call->lastReceiveTime)
6193 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6194 if (clock_diff > 5 * 60)
6196 if (call->state == RX_STATE_ACTIVE)
6197 rxi_CallError(call, RX_CALL_TIMEOUT);
6201 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6202 if (call->flags & RX_CALL_TQ_BUSY) {
6203 /* Call is active and will be reset by rxi_Start if it's
6204 * in an error state.
6209 /* RTT + 8*MDEV, rounded up to the next second. */
6210 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6211 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6213 deadTime = conn->secondsUntilDead + fudgeFactor;
6214 /* These are computed to the second (+- 1 second). But that's
6215 * good enough for these values, which should be a significant
6216 * number of seconds. */
6217 if (now > (call->lastReceiveTime + deadTime)) {
6218 if (call->state == RX_STATE_ACTIVE) {
6220 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6222 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6223 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6224 ip_stack_t *ipst = ns->netstack_ip;
6226 ire = ire_cache_lookup(conn->peer->host
6227 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6229 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6231 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6238 if (ire && ire->ire_max_frag > 0)
6239 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6241 #if defined(GLOBAL_NETSTACKID)
6245 #endif /* ADAPT_PMTU */
6246 cerror = RX_CALL_DEAD;
6249 #ifdef RX_ENABLE_LOCKS
6250 /* Cancel pending events */
6251 rxevent_Cancel(&call->delayedAckEvent, call,
6252 RX_CALL_REFCOUNT_DELAY);
6253 rxi_rto_cancel(call);
6254 rxevent_Cancel(&call->keepAliveEvent, call,
6255 RX_CALL_REFCOUNT_ALIVE);
6256 rxevent_Cancel(&call->growMTUEvent, call,
6257 RX_CALL_REFCOUNT_MTU);
6258 MUTEX_ENTER(&rx_refcnt_mutex);
6259 /* if rxi_FreeCall returns 1 it has freed the call */
6260 if (call->refCount == 0 &&
6261 rxi_FreeCall(call, haveCTLock))
6263 MUTEX_EXIT(&rx_refcnt_mutex);
6266 MUTEX_EXIT(&rx_refcnt_mutex);
6268 #else /* RX_ENABLE_LOCKS */
6269 rxi_FreeCall(call, 0);
6271 #endif /* RX_ENABLE_LOCKS */
6273 /* Non-active calls are destroyed if they are not responding
6274 * to pings; active calls are simply flagged in error, so the
6275 * attached process can die reasonably gracefully. */
6278 if (conn->idleDeadDetection) {
6279 if (conn->idleDeadTime) {
6280 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6284 /* see if we have a non-activity timeout */
6285 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6286 (call->flags & RX_CALL_READER_WAIT)) {
6287 if (call->state == RX_STATE_ACTIVE) {
6288 cerror = RX_CALL_TIMEOUT;
6293 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6294 if (call->state == RX_STATE_ACTIVE) {
6295 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6303 if (conn->hardDeadTime) {
6304 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6307 /* see if we have a hard timeout */
6309 && (now > (hardDeadTime + call->startTime.sec))) {
6310 if (call->state == RX_STATE_ACTIVE)
6311 rxi_CallError(call, RX_CALL_TIMEOUT);
6316 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6317 call->lastReceiveTime) {
6318 int oldMTU = conn->peer->ifMTU;
6320 /* if we thought we could send more, perhaps things got worse */
6321 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6322 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6323 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6324 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6326 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6328 /* minimum capped in SetPeerMtu */
6329 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6332 conn->lastPacketSize = 0;
6334 /* needed so ResetCall doesn't clobber us. */
6335 call->MTU = conn->peer->ifMTU;
6337 /* if we never succeeded, let the error pass out as-is */
6338 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6339 cerror = conn->msgsizeRetryErr;
6342 rxi_CallError(call, cerror);
6347 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6348 void *dummy, int dummy2)
6350 struct rx_connection *conn = arg1;
6351 struct rx_header theader;
6352 char tbuffer[1 + sizeof(struct rx_header)];
6353 struct sockaddr_in taddr;
6356 struct iovec tmpiov[2];
6359 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6362 tp = &tbuffer[sizeof(struct rx_header)];
6363 taddr.sin_family = AF_INET;
6364 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6365 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6366 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6367 taddr.sin_len = sizeof(struct sockaddr_in);
6369 memset(&theader, 0, sizeof(theader));
6370 theader.epoch = htonl(999);
6372 theader.callNumber = 0;
6375 theader.type = RX_PACKET_TYPE_VERSION;
6376 theader.flags = RX_LAST_PACKET;
6377 theader.serviceId = 0;
6379 memcpy(tbuffer, &theader, sizeof(theader));
6380 memcpy(tp, &a, sizeof(a));
6381 tmpiov[0].iov_base = tbuffer;
6382 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6384 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6386 MUTEX_ENTER(&conn->conn_data_lock);
6387 MUTEX_ENTER(&rx_refcnt_mutex);
6388 /* Only reschedule ourselves if the connection would not be destroyed */
6389 if (conn->refCount <= 1) {
6390 rxevent_Put(conn->natKeepAliveEvent);
6391 conn->natKeepAliveEvent = NULL;
6392 MUTEX_EXIT(&rx_refcnt_mutex);
6393 MUTEX_EXIT(&conn->conn_data_lock);
6394 rx_DestroyConnection(conn); /* drop the reference for this */
6396 conn->refCount--; /* drop the reference for this */
6397 MUTEX_EXIT(&rx_refcnt_mutex);
6398 rxevent_Put(conn->natKeepAliveEvent);
6399 conn->natKeepAliveEvent = NULL;
6400 rxi_ScheduleNatKeepAliveEvent(conn);
6401 MUTEX_EXIT(&conn->conn_data_lock);
6406 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6408 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6409 struct clock when, now;
6410 clock_GetTime(&now);
6412 when.sec += conn->secondsUntilNatPing;
6413 MUTEX_ENTER(&rx_refcnt_mutex);
6414 conn->refCount++; /* hold a reference for this */
6415 MUTEX_EXIT(&rx_refcnt_mutex);
6416 conn->natKeepAliveEvent =
6417 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6422 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6424 MUTEX_ENTER(&conn->conn_data_lock);
6425 conn->secondsUntilNatPing = seconds;
6427 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6428 rxi_ScheduleNatKeepAliveEvent(conn);
6430 conn->flags |= RX_CONN_NAT_PING;
6432 MUTEX_EXIT(&conn->conn_data_lock);
6435 /* When a call is in progress, this routine is called occasionally to
6436 * make sure that some traffic has arrived (or been sent to) the peer.
6437 * If nothing has arrived in a reasonable amount of time, the call is
6438 * declared dead; if nothing has been sent for a while, we send a
6439 * keep-alive packet (if we're actually trying to keep the call alive)
6442 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6445 struct rx_call *call = arg1;
6446 struct rx_connection *conn;
6449 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6450 MUTEX_ENTER(&call->lock);
6452 if (event == call->keepAliveEvent) {
6453 rxevent_Put(call->keepAliveEvent);
6454 call->keepAliveEvent = NULL;
6459 #ifdef RX_ENABLE_LOCKS
6460 if (rxi_CheckCall(call, 0)) {
6461 MUTEX_EXIT(&call->lock);
6464 #else /* RX_ENABLE_LOCKS */
6465 if (rxi_CheckCall(call))
6467 #endif /* RX_ENABLE_LOCKS */
6469 /* Don't try to keep alive dallying calls */
6470 if (call->state == RX_STATE_DALLY) {
6471 MUTEX_EXIT(&call->lock);
6476 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6477 /* Don't try to send keepalives if there is unacknowledged data */
6478 /* the rexmit code should be good enough, this little hack
6479 * doesn't quite work XXX */
6480 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6482 rxi_ScheduleKeepAliveEvent(call);
6483 MUTEX_EXIT(&call->lock);
6486 /* Does what's on the nameplate. */
6488 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6490 struct rx_call *call = arg1;
6491 struct rx_connection *conn;
6493 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6494 MUTEX_ENTER(&call->lock);
6496 if (event == call->growMTUEvent) {
6497 rxevent_Put(call->growMTUEvent);
6498 call->growMTUEvent = NULL;
6501 #ifdef RX_ENABLE_LOCKS
6502 if (rxi_CheckCall(call, 0)) {
6503 MUTEX_EXIT(&call->lock);
6506 #else /* RX_ENABLE_LOCKS */
6507 if (rxi_CheckCall(call))
6509 #endif /* RX_ENABLE_LOCKS */
6511 /* Don't bother with dallying calls */
6512 if (call->state == RX_STATE_DALLY) {
6513 MUTEX_EXIT(&call->lock);
6520 * keep being scheduled, just don't do anything if we're at peak,
6521 * or we're not set up to be properly handled (idle timeout required)
6523 if ((conn->peer->maxPacketSize != 0) &&
6524 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6525 conn->idleDeadDetection)
6526 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6527 rxi_ScheduleGrowMTUEvent(call, 0);
6528 MUTEX_EXIT(&call->lock);
6532 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6534 if (!call->keepAliveEvent) {
6535 struct clock when, now;
6536 clock_GetTime(&now);
6538 when.sec += call->conn->secondsUntilPing;
6539 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6540 call->keepAliveEvent =
6541 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6546 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6548 if (!call->growMTUEvent) {
6549 struct clock when, now;
6551 clock_GetTime(&now);
6554 if (call->conn->secondsUntilPing)
6555 secs = (6*call->conn->secondsUntilPing)-1;
6557 if (call->conn->secondsUntilDead)
6558 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6562 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6563 call->growMTUEvent =
6564 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6568 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6570 rxi_KeepAliveOn(struct rx_call *call)
6572 /* Pretend last packet received was received now--i.e. if another
6573 * packet isn't received within the keep alive time, then the call
6574 * will die; Initialize last send time to the current time--even
6575 * if a packet hasn't been sent yet. This will guarantee that a
6576 * keep-alive is sent within the ping time */
6577 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6578 rxi_ScheduleKeepAliveEvent(call);
6582 * Solely in order that callers not need to include rx_call.h
6585 rx_KeepAliveOff(struct rx_call *call)
6587 rxi_KeepAliveOff(call);
6590 rx_KeepAliveOn(struct rx_call *call)
6592 rxi_KeepAliveOn(call);
6596 rxi_GrowMTUOn(struct rx_call *call)
6598 struct rx_connection *conn = call->conn;
6599 MUTEX_ENTER(&conn->conn_data_lock);
6600 conn->lastPingSizeSer = conn->lastPingSize = 0;
6601 MUTEX_EXIT(&conn->conn_data_lock);
6602 rxi_ScheduleGrowMTUEvent(call, 1);
6605 /* This routine is called to send connection abort messages
6606 * that have been delayed to throttle looping clients. */
6608 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6611 struct rx_connection *conn = arg1;
6614 struct rx_packet *packet;
6616 MUTEX_ENTER(&conn->conn_data_lock);
6617 rxevent_Put(conn->delayedAbortEvent);
6618 conn->delayedAbortEvent = NULL;
6619 error = htonl(conn->error);
6621 MUTEX_EXIT(&conn->conn_data_lock);
6622 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6625 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6626 RX_PACKET_TYPE_ABORT, (char *)&error,
6628 rxi_FreePacket(packet);
6632 /* This routine is called to send call abort messages
6633 * that have been delayed to throttle looping clients. */
6635 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6638 struct rx_call *call = arg1;
6641 struct rx_packet *packet;
6643 MUTEX_ENTER(&call->lock);
6644 rxevent_Put(call->delayedAbortEvent);
6645 call->delayedAbortEvent = NULL;
6646 error = htonl(call->error);
6648 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6651 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6652 (char *)&error, sizeof(error), 0);
6653 rxi_FreePacket(packet);
6655 MUTEX_EXIT(&call->lock);
6656 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6659 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6660 * seconds) to ask the client to authenticate itself. The routine
6661 * issues a challenge to the client, which is obtained from the
6662 * security object associated with the connection */
6664 rxi_ChallengeEvent(struct rxevent *event,
6665 void *arg0, void *arg1, int tries)
6667 struct rx_connection *conn = arg0;
6670 rxevent_Put(conn->challengeEvent);
6671 conn->challengeEvent = NULL;
6674 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6675 struct rx_packet *packet;
6676 struct clock when, now;
6679 /* We've failed to authenticate for too long.
6680 * Reset any calls waiting for authentication;
6681 * they are all in RX_STATE_PRECALL.
6685 MUTEX_ENTER(&conn->conn_call_lock);
6686 for (i = 0; i < RX_MAXCALLS; i++) {
6687 struct rx_call *call = conn->call[i];
6689 MUTEX_ENTER(&call->lock);
6690 if (call->state == RX_STATE_PRECALL) {
6691 rxi_CallError(call, RX_CALL_DEAD);
6692 rxi_SendCallAbort(call, NULL, 0, 0);
6694 MUTEX_EXIT(&call->lock);
6697 MUTEX_EXIT(&conn->conn_call_lock);
6701 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6703 /* If there's no packet available, do this later. */
6704 RXS_GetChallenge(conn->securityObject, conn, packet);
6705 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6706 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6707 rxi_FreePacket(packet);
6709 clock_GetTime(&now);
6711 when.sec += RX_CHALLENGE_TIMEOUT;
6712 conn->challengeEvent =
6713 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6718 /* Call this routine to start requesting the client to authenticate
6719 * itself. This will continue until authentication is established,
6720 * the call times out, or an invalid response is returned. The
6721 * security object associated with the connection is asked to create
6722 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6723 * defined earlier. */
6725 rxi_ChallengeOn(struct rx_connection *conn)
6727 if (!conn->challengeEvent) {
6728 RXS_CreateChallenge(conn->securityObject, conn);
6729 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6734 /* rxi_ComputeRoundTripTime is called with peer locked. */
6735 /* peer may be null */
6737 rxi_ComputeRoundTripTime(struct rx_packet *p,
6738 struct rx_ackPacket *ack,
6739 struct rx_call *call,
6740 struct rx_peer *peer,
6743 struct clock thisRtt, *sentp;
6747 /* If the ACK is delayed, then do nothing */
6748 if (ack->reason == RX_ACK_DELAY)
6751 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6752 * their RTT multiple times, so only include the RTT of the last packet
6754 if (p->flags & RX_JUMBO_PACKET)
6757 /* Use the serial number to determine which transmission the ACK is for,
6758 * and set the sent time to match this. If we have no serial number, then
6759 * only use the ACK for RTT calculations if the packet has not been
6763 serial = ntohl(ack->serial);
6765 if (serial == p->header.serial) {
6766 sentp = &p->timeSent;
6767 } else if (serial == p->firstSerial) {
6768 sentp = &p->firstSent;
6769 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6770 sentp = &p->firstSent;
6774 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6775 sentp = &p->firstSent;
6782 if (clock_Lt(&thisRtt, sentp))
6783 return; /* somebody set the clock back, don't count this time. */
6785 clock_Sub(&thisRtt, sentp);
6786 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6787 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6789 if (clock_IsZero(&thisRtt)) {
6791 * The actual round trip time is shorter than the
6792 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6793 * Since we can't tell which at the moment we will assume 1ms.
6795 thisRtt.usec = 1000;
6798 if (rx_stats_active) {
6799 MUTEX_ENTER(&rx_stats_mutex);
6800 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6801 rx_stats.minRtt = thisRtt;
6802 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6803 if (thisRtt.sec > 60) {
6804 MUTEX_EXIT(&rx_stats_mutex);
6805 return; /* somebody set the clock ahead */
6807 rx_stats.maxRtt = thisRtt;
6809 clock_Add(&rx_stats.totalRtt, &thisRtt);
6810 rx_atomic_inc(&rx_stats.nRttSamples);
6811 MUTEX_EXIT(&rx_stats_mutex);
6814 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6816 /* Apply VanJacobson round-trip estimations */
6821 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6822 * srtt is stored as fixed point with 3 bits after the binary
6823 * point (i.e., scaled by 8). The following magic is
6824 * equivalent to the smoothing algorithm in rfc793 with an
6825 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6826 * srtt'*8 = rtt + srtt*7
6827 * srtt'*8 = srtt*8 + rtt - srtt
6828 * srtt' = srtt + rtt/8 - srtt/8
6829 * srtt' = srtt + (rtt - srtt)/8
6832 delta = _8THMSEC(&thisRtt) - call->rtt;
6833 call->rtt += (delta >> 3);
6836 * We accumulate a smoothed rtt variance (actually, a smoothed
6837 * mean difference), then set the retransmit timer to smoothed
6838 * rtt + 4 times the smoothed variance (was 2x in van's original
6839 * paper, but 4x works better for me, and apparently for him as
6841 * rttvar is stored as
6842 * fixed point with 2 bits after the binary point (scaled by
6843 * 4). The following is equivalent to rfc793 smoothing with
6844 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6845 * rttvar'*4 = rttvar*3 + |delta|
6846 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6847 * rttvar' = rttvar + |delta|/4 - rttvar/4
6848 * rttvar' = rttvar + (|delta| - rttvar)/4
6849 * This replaces rfc793's wired-in beta.
6850 * dev*4 = dev*4 + (|actual - expected| - dev)
6856 delta -= (call->rtt_dev << 1);
6857 call->rtt_dev += (delta >> 3);
6859 /* I don't have a stored RTT so I start with this value. Since I'm
6860 * probably just starting a call, and will be pushing more data down
6861 * this, I expect congestion to increase rapidly. So I fudge a
6862 * little, and I set deviance to half the rtt. In practice,
6863 * deviance tends to approach something a little less than
6864 * half the smoothed rtt. */
6865 call->rtt = _8THMSEC(&thisRtt) + 8;
6866 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6868 /* the smoothed RTT time is RTT + 4*MDEV
6870 * We allow a user specified minimum to be set for this, to allow clamping
6871 * at a minimum value in the same way as TCP. In addition, we have to allow
6872 * for the possibility that this packet is answered by a delayed ACK, so we
6873 * add on a fixed 200ms to account for that timer expiring.
6876 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6877 rx_minPeerTimeout) + 200;
6878 clock_Zero(&call->rto);
6879 clock_Addmsec(&call->rto, rtt_timeout);
6881 /* Update the peer, so any new calls start with our values */
6882 peer->rtt_dev = call->rtt_dev;
6883 peer->rtt = call->rtt;
6885 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6886 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6890 /* Find all server connections that have not been active for a long time, and
6893 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6896 struct clock now, when;
6897 clock_GetTime(&now);
6899 /* Find server connection structures that haven't been used for
6900 * greater than rx_idleConnectionTime */
6902 struct rx_connection **conn_ptr, **conn_end;
6903 int i, havecalls = 0;
6904 MUTEX_ENTER(&rx_connHashTable_lock);
6905 for (conn_ptr = &rx_connHashTable[0], conn_end =
6906 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6908 struct rx_connection *conn, *next;
6909 struct rx_call *call;
6913 for (conn = *conn_ptr; conn; conn = next) {
6914 /* XXX -- Shouldn't the connection be locked? */
6917 for (i = 0; i < RX_MAXCALLS; i++) {
6918 call = conn->call[i];
6922 code = MUTEX_TRYENTER(&call->lock);
6925 #ifdef RX_ENABLE_LOCKS
6926 result = rxi_CheckCall(call, 1);
6927 #else /* RX_ENABLE_LOCKS */
6928 result = rxi_CheckCall(call);
6929 #endif /* RX_ENABLE_LOCKS */
6930 MUTEX_EXIT(&call->lock);
6932 /* If CheckCall freed the call, it might
6933 * have destroyed the connection as well,
6934 * which screws up the linked lists.
6940 if (conn->type == RX_SERVER_CONNECTION) {
6941 /* This only actually destroys the connection if
6942 * there are no outstanding calls */
6943 MUTEX_ENTER(&conn->conn_data_lock);
6944 MUTEX_ENTER(&rx_refcnt_mutex);
6945 if (!havecalls && !conn->refCount
6946 && ((conn->lastSendTime + rx_idleConnectionTime) <
6948 conn->refCount++; /* it will be decr in rx_DestroyConn */
6949 MUTEX_EXIT(&rx_refcnt_mutex);
6950 MUTEX_EXIT(&conn->conn_data_lock);
6951 #ifdef RX_ENABLE_LOCKS
6952 rxi_DestroyConnectionNoLock(conn);
6953 #else /* RX_ENABLE_LOCKS */
6954 rxi_DestroyConnection(conn);
6955 #endif /* RX_ENABLE_LOCKS */
6957 #ifdef RX_ENABLE_LOCKS
6959 MUTEX_EXIT(&rx_refcnt_mutex);
6960 MUTEX_EXIT(&conn->conn_data_lock);
6962 #endif /* RX_ENABLE_LOCKS */
6966 #ifdef RX_ENABLE_LOCKS
6967 while (rx_connCleanup_list) {
6968 struct rx_connection *conn;
6969 conn = rx_connCleanup_list;
6970 rx_connCleanup_list = rx_connCleanup_list->next;
6971 MUTEX_EXIT(&rx_connHashTable_lock);
6972 rxi_CleanupConnection(conn);
6973 MUTEX_ENTER(&rx_connHashTable_lock);
6975 MUTEX_EXIT(&rx_connHashTable_lock);
6976 #endif /* RX_ENABLE_LOCKS */
6979 /* Find any peer structures that haven't been used (haven't had an
6980 * associated connection) for greater than rx_idlePeerTime */
6982 struct rx_peer **peer_ptr, **peer_end;
6986 * Why do we need to hold the rx_peerHashTable_lock across
6987 * the incrementing of peer_ptr since the rx_peerHashTable
6988 * array is not changing? We don't.
6990 * By dropping the lock periodically we can permit other
6991 * activities to be performed while a rxi_ReapConnections
6992 * call is in progress. The goal of reap connections
6993 * is to clean up quickly without causing large amounts
6994 * of contention. Therefore, it is important that global
6995 * mutexes not be held for extended periods of time.
6997 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6998 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7000 struct rx_peer *peer, *next, *prev;
7002 MUTEX_ENTER(&rx_peerHashTable_lock);
7003 for (prev = peer = *peer_ptr; peer; peer = next) {
7005 code = MUTEX_TRYENTER(&peer->peer_lock);
7006 if ((code) && (peer->refCount == 0)
7007 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7008 rx_interface_stat_p rpc_stat, nrpc_stat;
7012 * now know that this peer object is one to be
7013 * removed from the hash table. Once it is removed
7014 * it can't be referenced by other threads.
7015 * Lets remove it first and decrement the struct
7016 * nPeerStructs count.
7018 if (peer == *peer_ptr) {
7024 if (rx_stats_active)
7025 rx_atomic_dec(&rx_stats.nPeerStructs);
7028 * Now if we hold references on 'prev' and 'next'
7029 * we can safely drop the rx_peerHashTable_lock
7030 * while we destroy this 'peer' object.
7036 MUTEX_EXIT(&rx_peerHashTable_lock);
7038 MUTEX_EXIT(&peer->peer_lock);
7039 MUTEX_DESTROY(&peer->peer_lock);
7041 (&peer->rpcStats, rpc_stat, nrpc_stat,
7042 rx_interface_stat)) {
7043 unsigned int num_funcs;
7046 queue_Remove(&rpc_stat->queue_header);
7047 queue_Remove(&rpc_stat->all_peers);
7048 num_funcs = rpc_stat->stats[0].func_total;
7050 sizeof(rx_interface_stat_t) +
7051 rpc_stat->stats[0].func_total *
7052 sizeof(rx_function_entry_v1_t);
7054 rxi_Free(rpc_stat, space);
7056 MUTEX_ENTER(&rx_rpc_stats);
7057 rxi_rpc_peer_stat_cnt -= num_funcs;
7058 MUTEX_EXIT(&rx_rpc_stats);
7063 * Regain the rx_peerHashTable_lock and
7064 * decrement the reference count on 'prev'
7067 MUTEX_ENTER(&rx_peerHashTable_lock);
7074 MUTEX_EXIT(&peer->peer_lock);
7079 MUTEX_EXIT(&rx_peerHashTable_lock);
7083 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7084 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7085 * GC, just below. Really, we shouldn't have to keep moving packets from
7086 * one place to another, but instead ought to always know if we can
7087 * afford to hold onto a packet in its particular use. */
7088 MUTEX_ENTER(&rx_freePktQ_lock);
7089 if (rx_waitingForPackets) {
7090 rx_waitingForPackets = 0;
7091 #ifdef RX_ENABLE_LOCKS
7092 CV_BROADCAST(&rx_waitingForPackets_cv);
7094 osi_rxWakeup(&rx_waitingForPackets);
7097 MUTEX_EXIT(&rx_freePktQ_lock);
7100 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7101 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7105 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7106 * rx.h is sort of strange this is better. This is called with a security
7107 * object before it is discarded. Each connection using a security object has
7108 * its own refcount to the object so it won't actually be freed until the last
7109 * connection is destroyed.
7111 * This is the only rxs module call. A hold could also be written but no one
7115 rxs_Release(struct rx_securityClass *aobj)
7117 return RXS_Close(aobj);
7125 #define TRACE_OPTION_RX_DEBUG 16
7133 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7134 0, KEY_QUERY_VALUE, &parmKey);
7135 if (code != ERROR_SUCCESS)
7138 dummyLen = sizeof(TraceOption);
7139 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7140 (BYTE *) &TraceOption, &dummyLen);
7141 if (code == ERROR_SUCCESS) {
7142 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7144 RegCloseKey (parmKey);
7145 #endif /* AFS_NT40_ENV */
7150 rx_DebugOnOff(int on)
7154 rxdebug_active = on;
7160 rx_StatsOnOff(int on)
7162 rx_stats_active = on;
7166 /* Don't call this debugging routine directly; use dpf */
7168 rxi_DebugPrint(char *format, ...)
7177 va_start(ap, format);
7179 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7182 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7184 OutputDebugString(msg);
7190 va_start(ap, format);
7192 clock_GetTime(&now);
7193 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7194 (unsigned int)now.usec);
7195 vfprintf(rx_Log, format, ap);
7203 * This function is used to process the rx_stats structure that is local
7204 * to a process as well as an rx_stats structure received from a remote
7205 * process (via rxdebug). Therefore, it needs to do minimal version
7209 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7210 afs_int32 freePackets, char version)
7214 if (size != sizeof(struct rx_statistics)) {
7216 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7217 size, sizeof(struct rx_statistics));
7220 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7223 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7224 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7225 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7226 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7227 s->specialPktAllocFailures);
7229 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7230 s->receivePktAllocFailures, s->sendPktAllocFailures,
7231 s->specialPktAllocFailures);
7235 " greedy %u, " "bogusReads %u (last from host %x), "
7236 "noPackets %u, " "noBuffers %u, " "selects %u, "
7237 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7238 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7239 s->selects, s->sendSelects);
7241 fprintf(file, " packets read: ");
7242 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7243 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7245 fprintf(file, "\n");
7248 " other read counters: data %u, " "ack %u, " "dup %u "
7249 "spurious %u " "dally %u\n", s->dataPacketsRead,
7250 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7251 s->ignorePacketDally);
7253 fprintf(file, " packets sent: ");
7254 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7255 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7257 fprintf(file, "\n");
7260 " other send counters: ack %u, " "data %u (not resends), "
7261 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7262 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7263 s->dataPacketsPushed, s->ignoreAckedPacket);
7266 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7267 s->netSendFailures, (int)s->fatalErrors);
7269 if (s->nRttSamples) {
7270 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7271 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7273 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7274 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7278 " %d server connections, " "%d client connections, "
7279 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7280 s->nServerConns, s->nClientConns, s->nPeerStructs,
7281 s->nCallStructs, s->nFreeCallStructs);
7283 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7284 fprintf(file, " %d clock updates\n", clock_nUpdates);
7288 /* for backward compatibility */
7290 rx_PrintStats(FILE * file)
7292 MUTEX_ENTER(&rx_stats_mutex);
7293 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7294 sizeof(rx_stats), rx_nFreePackets,
7296 MUTEX_EXIT(&rx_stats_mutex);
7300 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7302 fprintf(file, "Peer %x.%d.\n",
7303 ntohl(peer->host), (int)ntohs(peer->port));
7306 " Rtt %d, " "total sent %d, " "resent %d\n",
7307 peer->rtt, peer->nSent, peer->reSends);
7309 fprintf(file, " Packet size %d\n", peer->ifMTU);
7313 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7315 * This mutex protects the following static variables:
7319 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7320 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7322 #define LOCK_RX_DEBUG
7323 #define UNLOCK_RX_DEBUG
7324 #endif /* AFS_PTHREAD_ENV */
7326 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7328 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7329 u_char type, void *inputData, size_t inputLength,
7330 void *outputData, size_t outputLength)
7332 static afs_int32 counter = 100;
7333 time_t waitTime, waitCount;
7334 struct rx_header theader;
7337 struct timeval tv_now, tv_wake, tv_delta;
7338 struct sockaddr_in taddr, faddr;
7352 tp = &tbuffer[sizeof(struct rx_header)];
7353 taddr.sin_family = AF_INET;
7354 taddr.sin_port = remotePort;
7355 taddr.sin_addr.s_addr = remoteAddr;
7356 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7357 taddr.sin_len = sizeof(struct sockaddr_in);
7360 memset(&theader, 0, sizeof(theader));
7361 theader.epoch = htonl(999);
7363 theader.callNumber = htonl(counter);
7366 theader.type = type;
7367 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7368 theader.serviceId = 0;
7370 memcpy(tbuffer, &theader, sizeof(theader));
7371 memcpy(tp, inputData, inputLength);
7373 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7374 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7376 /* see if there's a packet available */
7377 gettimeofday(&tv_wake, NULL);
7378 tv_wake.tv_sec += waitTime;
7381 FD_SET(socket, &imask);
7382 tv_delta.tv_sec = tv_wake.tv_sec;
7383 tv_delta.tv_usec = tv_wake.tv_usec;
7384 gettimeofday(&tv_now, NULL);
7386 if (tv_delta.tv_usec < tv_now.tv_usec) {
7388 tv_delta.tv_usec += 1000000;
7391 tv_delta.tv_usec -= tv_now.tv_usec;
7393 if (tv_delta.tv_sec < tv_now.tv_sec) {
7397 tv_delta.tv_sec -= tv_now.tv_sec;
7400 code = select(0, &imask, 0, 0, &tv_delta);
7401 #else /* AFS_NT40_ENV */
7402 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7403 #endif /* AFS_NT40_ENV */
7404 if (code == 1 && FD_ISSET(socket, &imask)) {
7405 /* now receive a packet */
7406 faddrLen = sizeof(struct sockaddr_in);
7408 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7409 (struct sockaddr *)&faddr, &faddrLen);
7412 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7413 if (counter == ntohl(theader.callNumber))
7421 /* see if we've timed out */
7429 code -= sizeof(struct rx_header);
7430 if (code > outputLength)
7431 code = outputLength;
7432 memcpy(outputData, tp, code);
7435 #endif /* RXDEBUG */
7438 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7439 afs_uint16 remotePort, struct rx_debugStats * stat,
7440 afs_uint32 * supportedValues)
7442 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7444 struct rx_debugIn in;
7446 *supportedValues = 0;
7447 in.type = htonl(RX_DEBUGI_GETSTATS);
7450 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7451 &in, sizeof(in), stat, sizeof(*stat));
7454 * If the call was successful, fixup the version and indicate
7455 * what contents of the stat structure are valid.
7456 * Also do net to host conversion of fields here.
7460 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7461 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7463 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7464 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7466 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7467 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7469 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7470 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7472 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7473 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7475 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7476 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7478 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7479 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7481 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7482 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7484 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7485 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7487 stat->nFreePackets = ntohl(stat->nFreePackets);
7488 stat->packetReclaims = ntohl(stat->packetReclaims);
7489 stat->callsExecuted = ntohl(stat->callsExecuted);
7490 stat->nWaiting = ntohl(stat->nWaiting);
7491 stat->idleThreads = ntohl(stat->idleThreads);
7492 stat->nWaited = ntohl(stat->nWaited);
7493 stat->nPackets = ntohl(stat->nPackets);
7502 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7503 afs_uint16 remotePort, struct rx_statistics * stat,
7504 afs_uint32 * supportedValues)
7506 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7508 struct rx_debugIn in;
7509 afs_int32 *lp = (afs_int32 *) stat;
7513 * supportedValues is currently unused, but added to allow future
7514 * versioning of this function.
7517 *supportedValues = 0;
7518 in.type = htonl(RX_DEBUGI_RXSTATS);
7520 memset(stat, 0, sizeof(*stat));
7522 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7523 &in, sizeof(in), stat, sizeof(*stat));
7528 * Do net to host conversion here
7531 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7542 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7543 afs_uint16 remotePort, size_t version_length,
7546 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7548 return MakeDebugCall(socket, remoteAddr, remotePort,
7549 RX_PACKET_TYPE_VERSION, a, 1, version,
7557 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7558 afs_uint16 remotePort, afs_int32 * nextConnection,
7559 int allConnections, afs_uint32 debugSupportedValues,
7560 struct rx_debugConn * conn,
7561 afs_uint32 * supportedValues)
7563 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7565 struct rx_debugIn in;
7569 * supportedValues is currently unused, but added to allow future
7570 * versioning of this function.
7573 *supportedValues = 0;
7574 if (allConnections) {
7575 in.type = htonl(RX_DEBUGI_GETALLCONN);
7577 in.type = htonl(RX_DEBUGI_GETCONN);
7579 in.index = htonl(*nextConnection);
7580 memset(conn, 0, sizeof(*conn));
7582 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7583 &in, sizeof(in), conn, sizeof(*conn));
7586 *nextConnection += 1;
7589 * Convert old connection format to new structure.
7592 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7593 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7594 #define MOVEvL(a) (conn->a = vL->a)
7596 /* any old or unrecognized version... */
7597 for (i = 0; i < RX_MAXCALLS; i++) {
7598 MOVEvL(callState[i]);
7599 MOVEvL(callMode[i]);
7600 MOVEvL(callFlags[i]);
7601 MOVEvL(callOther[i]);
7603 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7604 MOVEvL(secStats.type);
7605 MOVEvL(secStats.level);
7606 MOVEvL(secStats.flags);
7607 MOVEvL(secStats.expires);
7608 MOVEvL(secStats.packetsReceived);
7609 MOVEvL(secStats.packetsSent);
7610 MOVEvL(secStats.bytesReceived);
7611 MOVEvL(secStats.bytesSent);
7616 * Do net to host conversion here
7618 * I don't convert host or port since we are most likely
7619 * going to want these in NBO.
7621 conn->cid = ntohl(conn->cid);
7622 conn->serial = ntohl(conn->serial);
7623 for (i = 0; i < RX_MAXCALLS; i++) {
7624 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7626 conn->error = ntohl(conn->error);
7627 conn->secStats.flags = ntohl(conn->secStats.flags);
7628 conn->secStats.expires = ntohl(conn->secStats.expires);
7629 conn->secStats.packetsReceived =
7630 ntohl(conn->secStats.packetsReceived);
7631 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7632 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7633 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7634 conn->epoch = ntohl(conn->epoch);
7635 conn->natMTU = ntohl(conn->natMTU);
7644 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7645 afs_uint16 remotePort, afs_int32 * nextPeer,
7646 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7647 afs_uint32 * supportedValues)
7649 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7651 struct rx_debugIn in;
7654 * supportedValues is currently unused, but added to allow future
7655 * versioning of this function.
7658 *supportedValues = 0;
7659 in.type = htonl(RX_DEBUGI_GETPEER);
7660 in.index = htonl(*nextPeer);
7661 memset(peer, 0, sizeof(*peer));
7663 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7664 &in, sizeof(in), peer, sizeof(*peer));
7670 * Do net to host conversion here
7672 * I don't convert host or port since we are most likely
7673 * going to want these in NBO.
7675 peer->ifMTU = ntohs(peer->ifMTU);
7676 peer->idleWhen = ntohl(peer->idleWhen);
7677 peer->refCount = ntohs(peer->refCount);
7678 peer->rtt = ntohl(peer->rtt);
7679 peer->rtt_dev = ntohl(peer->rtt_dev);
7680 peer->timeout.sec = 0;
7681 peer->timeout.usec = 0;
7682 peer->nSent = ntohl(peer->nSent);
7683 peer->reSends = ntohl(peer->reSends);
7684 peer->natMTU = ntohs(peer->natMTU);
7685 peer->maxMTU = ntohs(peer->maxMTU);
7686 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7687 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7688 peer->MTU = ntohs(peer->MTU);
7689 peer->cwind = ntohs(peer->cwind);
7690 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7691 peer->congestSeq = ntohs(peer->congestSeq);
7692 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7693 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7694 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7695 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7704 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7705 struct rx_debugPeer * peerStats)
7708 afs_int32 error = 1; /* default to "did not succeed" */
7709 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7711 MUTEX_ENTER(&rx_peerHashTable_lock);
7712 for(tp = rx_peerHashTable[hashValue];
7713 tp != NULL; tp = tp->next) {
7714 if (tp->host == peerHost)
7720 MUTEX_EXIT(&rx_peerHashTable_lock);
7724 MUTEX_ENTER(&tp->peer_lock);
7725 peerStats->host = tp->host;
7726 peerStats->port = tp->port;
7727 peerStats->ifMTU = tp->ifMTU;
7728 peerStats->idleWhen = tp->idleWhen;
7729 peerStats->refCount = tp->refCount;
7730 peerStats->burstSize = 0;
7731 peerStats->burst = 0;
7732 peerStats->burstWait.sec = 0;
7733 peerStats->burstWait.usec = 0;
7734 peerStats->rtt = tp->rtt;
7735 peerStats->rtt_dev = tp->rtt_dev;
7736 peerStats->timeout.sec = 0;
7737 peerStats->timeout.usec = 0;
7738 peerStats->nSent = tp->nSent;
7739 peerStats->reSends = tp->reSends;
7740 peerStats->natMTU = tp->natMTU;
7741 peerStats->maxMTU = tp->maxMTU;
7742 peerStats->maxDgramPackets = tp->maxDgramPackets;
7743 peerStats->ifDgramPackets = tp->ifDgramPackets;
7744 peerStats->MTU = tp->MTU;
7745 peerStats->cwind = tp->cwind;
7746 peerStats->nDgramPackets = tp->nDgramPackets;
7747 peerStats->congestSeq = tp->congestSeq;
7748 peerStats->bytesSent.high = tp->bytesSent >> 32;
7749 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7750 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7751 peerStats->bytesReceived.low
7752 = tp->bytesReceived & MAX_AFS_UINT32;
7753 MUTEX_EXIT(&tp->peer_lock);
7755 MUTEX_ENTER(&rx_peerHashTable_lock);
7758 MUTEX_EXIT(&rx_peerHashTable_lock);
7766 struct rx_serverQueueEntry *np;
7769 struct rx_call *call;
7770 struct rx_serverQueueEntry *sq;
7774 if (rxinit_status == 1) {
7776 return; /* Already shutdown. */
7780 #ifndef AFS_PTHREAD_ENV
7781 FD_ZERO(&rx_selectMask);
7782 #endif /* AFS_PTHREAD_ENV */
7783 rxi_dataQuota = RX_MAX_QUOTA;
7784 #ifndef AFS_PTHREAD_ENV
7786 #endif /* AFS_PTHREAD_ENV */
7789 #ifndef AFS_PTHREAD_ENV
7790 #ifndef AFS_USE_GETTIMEOFDAY
7792 #endif /* AFS_USE_GETTIMEOFDAY */
7793 #endif /* AFS_PTHREAD_ENV */
7795 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7796 call = queue_First(&rx_freeCallQueue, rx_call);
7798 rxi_Free(call, sizeof(struct rx_call));
7801 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7802 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7808 struct rx_peer **peer_ptr, **peer_end;
7809 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7810 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7812 struct rx_peer *peer, *next;
7814 MUTEX_ENTER(&rx_peerHashTable_lock);
7815 for (peer = *peer_ptr; peer; peer = next) {
7816 rx_interface_stat_p rpc_stat, nrpc_stat;
7819 MUTEX_ENTER(&rx_rpc_stats);
7820 MUTEX_ENTER(&peer->peer_lock);
7822 (&peer->rpcStats, rpc_stat, nrpc_stat,
7823 rx_interface_stat)) {
7824 unsigned int num_funcs;
7827 queue_Remove(&rpc_stat->queue_header);
7828 queue_Remove(&rpc_stat->all_peers);
7829 num_funcs = rpc_stat->stats[0].func_total;
7831 sizeof(rx_interface_stat_t) +
7832 rpc_stat->stats[0].func_total *
7833 sizeof(rx_function_entry_v1_t);
7835 rxi_Free(rpc_stat, space);
7837 /* rx_rpc_stats must be held */
7838 rxi_rpc_peer_stat_cnt -= num_funcs;
7840 MUTEX_EXIT(&peer->peer_lock);
7841 MUTEX_EXIT(&rx_rpc_stats);
7845 if (rx_stats_active)
7846 rx_atomic_dec(&rx_stats.nPeerStructs);
7848 MUTEX_EXIT(&rx_peerHashTable_lock);
7851 for (i = 0; i < RX_MAX_SERVICES; i++) {
7853 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7855 for (i = 0; i < rx_hashTableSize; i++) {
7856 struct rx_connection *tc, *ntc;
7857 MUTEX_ENTER(&rx_connHashTable_lock);
7858 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7860 for (j = 0; j < RX_MAXCALLS; j++) {
7862 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7865 rxi_Free(tc, sizeof(*tc));
7867 MUTEX_EXIT(&rx_connHashTable_lock);
7870 MUTEX_ENTER(&freeSQEList_lock);
7872 while ((np = rx_FreeSQEList)) {
7873 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7874 MUTEX_DESTROY(&np->lock);
7875 rxi_Free(np, sizeof(*np));
7878 MUTEX_EXIT(&freeSQEList_lock);
7879 MUTEX_DESTROY(&freeSQEList_lock);
7880 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7881 MUTEX_DESTROY(&rx_connHashTable_lock);
7882 MUTEX_DESTROY(&rx_peerHashTable_lock);
7883 MUTEX_DESTROY(&rx_serverPool_lock);
7885 osi_Free(rx_connHashTable,
7886 rx_hashTableSize * sizeof(struct rx_connection *));
7887 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7889 UNPIN(rx_connHashTable,
7890 rx_hashTableSize * sizeof(struct rx_connection *));
7891 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7893 rxi_FreeAllPackets();
7895 MUTEX_ENTER(&rx_quota_mutex);
7896 rxi_dataQuota = RX_MAX_QUOTA;
7897 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7898 MUTEX_EXIT(&rx_quota_mutex);
7903 #ifdef RX_ENABLE_LOCKS
7905 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7907 if (!MUTEX_ISMINE(lockaddr))
7908 osi_Panic("Lock not held: %s", msg);
7910 #endif /* RX_ENABLE_LOCKS */
7915 * Routines to implement connection specific data.
7919 rx_KeyCreate(rx_destructor_t rtn)
7922 MUTEX_ENTER(&rxi_keyCreate_lock);
7923 key = rxi_keyCreate_counter++;
7924 rxi_keyCreate_destructor = (rx_destructor_t *)
7925 realloc((void *)rxi_keyCreate_destructor,
7926 (key + 1) * sizeof(rx_destructor_t));
7927 rxi_keyCreate_destructor[key] = rtn;
7928 MUTEX_EXIT(&rxi_keyCreate_lock);
7933 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7936 MUTEX_ENTER(&conn->conn_data_lock);
7937 if (!conn->specific) {
7938 conn->specific = malloc((key + 1) * sizeof(void *));
7939 for (i = 0; i < key; i++)
7940 conn->specific[i] = NULL;
7941 conn->nSpecific = key + 1;
7942 conn->specific[key] = ptr;
7943 } else if (key >= conn->nSpecific) {
7944 conn->specific = (void **)
7945 realloc(conn->specific, (key + 1) * sizeof(void *));
7946 for (i = conn->nSpecific; i < key; i++)
7947 conn->specific[i] = NULL;
7948 conn->nSpecific = key + 1;
7949 conn->specific[key] = ptr;
7951 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7952 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7953 conn->specific[key] = ptr;
7955 MUTEX_EXIT(&conn->conn_data_lock);
7959 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7962 MUTEX_ENTER(&svc->svc_data_lock);
7963 if (!svc->specific) {
7964 svc->specific = malloc((key + 1) * sizeof(void *));
7965 for (i = 0; i < key; i++)
7966 svc->specific[i] = NULL;
7967 svc->nSpecific = key + 1;
7968 svc->specific[key] = ptr;
7969 } else if (key >= svc->nSpecific) {
7970 svc->specific = (void **)
7971 realloc(svc->specific, (key + 1) * sizeof(void *));
7972 for (i = svc->nSpecific; i < key; i++)
7973 svc->specific[i] = NULL;
7974 svc->nSpecific = key + 1;
7975 svc->specific[key] = ptr;
7977 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7978 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7979 svc->specific[key] = ptr;
7981 MUTEX_EXIT(&svc->svc_data_lock);
7985 rx_GetSpecific(struct rx_connection *conn, int key)
7988 MUTEX_ENTER(&conn->conn_data_lock);
7989 if (key >= conn->nSpecific)
7992 ptr = conn->specific[key];
7993 MUTEX_EXIT(&conn->conn_data_lock);
7998 rx_GetServiceSpecific(struct rx_service *svc, int key)
8001 MUTEX_ENTER(&svc->svc_data_lock);
8002 if (key >= svc->nSpecific)
8005 ptr = svc->specific[key];
8006 MUTEX_EXIT(&svc->svc_data_lock);
8011 #endif /* !KERNEL */
8014 * processStats is a queue used to store the statistics for the local
8015 * process. Its contents are similar to the contents of the rpcStats
8016 * queue on a rx_peer structure, but the actual data stored within
8017 * this queue contains totals across the lifetime of the process (assuming
8018 * the stats have not been reset) - unlike the per peer structures
8019 * which can come and go based upon the peer lifetime.
8022 static struct rx_queue processStats = { &processStats, &processStats };
8025 * peerStats is a queue used to store the statistics for all peer structs.
8026 * Its contents are the union of all the peer rpcStats queues.
8029 static struct rx_queue peerStats = { &peerStats, &peerStats };
8032 * rxi_monitor_processStats is used to turn process wide stat collection
8036 static int rxi_monitor_processStats = 0;
8039 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8042 static int rxi_monitor_peerStats = 0;
8046 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8048 rpc_stat->invocations = 0;
8049 rpc_stat->bytes_sent = 0;
8050 rpc_stat->bytes_rcvd = 0;
8051 rpc_stat->queue_time_sum.sec = 0;
8052 rpc_stat->queue_time_sum.usec = 0;
8053 rpc_stat->queue_time_sum_sqr.sec = 0;
8054 rpc_stat->queue_time_sum_sqr.usec = 0;
8055 rpc_stat->queue_time_min.sec = 9999999;
8056 rpc_stat->queue_time_min.usec = 9999999;
8057 rpc_stat->queue_time_max.sec = 0;
8058 rpc_stat->queue_time_max.usec = 0;
8059 rpc_stat->execution_time_sum.sec = 0;
8060 rpc_stat->execution_time_sum.usec = 0;
8061 rpc_stat->execution_time_sum_sqr.sec = 0;
8062 rpc_stat->execution_time_sum_sqr.usec = 0;
8063 rpc_stat->execution_time_min.sec = 9999999;
8064 rpc_stat->execution_time_min.usec = 9999999;
8065 rpc_stat->execution_time_max.sec = 0;
8066 rpc_stat->execution_time_max.usec = 0;
8070 * Given all of the information for a particular rpc
8071 * call, find or create (if requested) the stat structure for the rpc.
8074 * the queue of stats that will be updated with the new value
8076 * @param rxInterface
8077 * a unique number that identifies the rpc interface
8080 * the total number of functions in this interface. this is only
8081 * required if create is true
8084 * if true, this invocation was made to a server
8087 * the ip address of the remote host. this is only required if create
8088 * and addToPeerList are true
8091 * the port of the remote host. this is only required if create
8092 * and addToPeerList are true
8094 * @param addToPeerList
8095 * if != 0, add newly created stat to the global peer list
8098 * if a new stats structure is allocated, the counter will
8099 * be updated with the new number of allocated stat structures.
8100 * only required if create is true
8103 * if no stats structure exists, allocate one
8107 static rx_interface_stat_p
8108 rxi_FindRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8109 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8110 afs_uint32 remotePort, int addToPeerList,
8111 unsigned int *counter, int create)
8113 rx_interface_stat_p rpc_stat, nrpc_stat;
8115 if (queue_IsEmpty(stats) && !create)
8119 * See if there's already a structure for this interface
8122 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8123 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8124 && (rpc_stat->stats[0].remote_is_server == isServer))
8128 /* if they didn't ask us to create, we're done */
8132 /* can't proceed without these */
8133 if (!totalFunc || !counter)
8137 * Didn't find a match so allocate a new structure and add it to the
8141 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8142 || (rpc_stat->stats[0].interfaceId != rxInterface)
8143 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8148 sizeof(rx_interface_stat_t) +
8149 totalFunc * sizeof(rx_function_entry_v1_t);
8151 rpc_stat = rxi_Alloc(space);
8152 if (rpc_stat == NULL)
8155 *counter += totalFunc;
8156 for (i = 0; i < totalFunc; i++) {
8157 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8158 rpc_stat->stats[i].remote_peer = remoteHost;
8159 rpc_stat->stats[i].remote_port = remotePort;
8160 rpc_stat->stats[i].remote_is_server = isServer;
8161 rpc_stat->stats[i].interfaceId = rxInterface;
8162 rpc_stat->stats[i].func_total = totalFunc;
8163 rpc_stat->stats[i].func_index = i;
8165 queue_Prepend(stats, rpc_stat);
8166 if (addToPeerList) {
8167 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8174 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8176 rx_interface_stat_p rpc_stat;
8179 if (rxInterface == -1)
8182 MUTEX_ENTER(&rx_rpc_stats);
8183 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8186 totalFunc = rpc_stat->stats[0].func_total;
8187 for (i = 0; i < totalFunc; i++)
8188 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8190 MUTEX_EXIT(&rx_rpc_stats);
8195 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8197 rx_interface_stat_p rpc_stat;
8199 struct rx_peer * peer;
8201 if (rxInterface == -1)
8204 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8208 MUTEX_ENTER(&rx_rpc_stats);
8209 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8212 totalFunc = rpc_stat->stats[0].func_total;
8213 for (i = 0; i < totalFunc; i++)
8214 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8216 MUTEX_EXIT(&rx_rpc_stats);
8221 rx_CopyProcessRPCStats(afs_uint64 op)
8223 rx_interface_stat_p rpc_stat;
8224 rx_function_entry_v1_p rpcop_stat =
8225 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8226 int currentFunc = (op & MAX_AFS_UINT32);
8227 afs_int32 rxInterface = (op >> 32);
8229 if (!rxi_monitor_processStats)
8232 if (rxInterface == -1)
8235 if (rpcop_stat == NULL)
8238 MUTEX_ENTER(&rx_rpc_stats);
8239 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8242 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8243 sizeof(rx_function_entry_v1_t));
8244 MUTEX_EXIT(&rx_rpc_stats);
8246 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8253 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8255 rx_interface_stat_p rpc_stat;
8256 rx_function_entry_v1_p rpcop_stat =
8257 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8258 int currentFunc = (op & MAX_AFS_UINT32);
8259 afs_int32 rxInterface = (op >> 32);
8260 struct rx_peer *peer;
8262 if (!rxi_monitor_peerStats)
8265 if (rxInterface == -1)
8268 if (rpcop_stat == NULL)
8271 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8275 MUTEX_ENTER(&rx_rpc_stats);
8276 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8279 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8280 sizeof(rx_function_entry_v1_t));
8281 MUTEX_EXIT(&rx_rpc_stats);
8283 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8290 rx_ReleaseRPCStats(void *stats)
8293 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8297 * Given all of the information for a particular rpc
8298 * call, create (if needed) and update the stat totals for the rpc.
8301 * the queue of stats that will be updated with the new value
8303 * @param rxInterface
8304 * a unique number that identifies the rpc interface
8306 * @param currentFunc
8307 * the index of the function being invoked
8310 * the total number of functions in this interface
8313 * the amount of time this function waited for a thread
8316 * the amount of time this function invocation took to execute
8319 * the number bytes sent by this invocation
8322 * the number bytes received by this invocation
8325 * if true, this invocation was made to a server
8328 * the ip address of the remote host
8331 * the port of the remote host
8333 * @param addToPeerList
8334 * if != 0, add newly created stat to the global peer list
8337 * if a new stats structure is allocated, the counter will
8338 * be updated with the new number of allocated stat structures
8343 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8344 afs_uint32 currentFunc, afs_uint32 totalFunc,
8345 struct clock *queueTime, struct clock *execTime,
8346 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8347 afs_uint32 remoteHost, afs_uint32 remotePort,
8348 int addToPeerList, unsigned int *counter)
8351 rx_interface_stat_p rpc_stat;
8353 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8354 remoteHost, remotePort, addToPeerList, counter,
8362 * Increment the stats for this function
8365 rpc_stat->stats[currentFunc].invocations++;
8366 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8367 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8368 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8369 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8370 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8371 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8373 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8374 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8376 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8377 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8379 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8380 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8382 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8383 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8391 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8392 afs_uint32 currentFunc, afs_uint32 totalFunc,
8393 struct clock *queueTime, struct clock *execTime,
8394 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8398 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8401 MUTEX_ENTER(&rx_rpc_stats);
8403 if (rxi_monitor_peerStats) {
8404 MUTEX_ENTER(&peer->peer_lock);
8405 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8406 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8407 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8408 MUTEX_EXIT(&peer->peer_lock);
8411 if (rxi_monitor_processStats) {
8412 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8413 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8414 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8417 MUTEX_EXIT(&rx_rpc_stats);
8421 * Increment the times and count for a particular rpc function.
8423 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8424 * call rx_RecordCallStatistics instead, so the public version of this
8425 * function is left purely for legacy callers.
8428 * The peer who invoked the rpc
8430 * @param rxInterface
8431 * A unique number that identifies the rpc interface
8433 * @param currentFunc
8434 * The index of the function being invoked
8437 * The total number of functions in this interface
8440 * The amount of time this function waited for a thread
8443 * The amount of time this function invocation took to execute
8446 * The number bytes sent by this invocation
8449 * The number bytes received by this invocation
8452 * If true, this invocation was made to a server
8456 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8457 afs_uint32 currentFunc, afs_uint32 totalFunc,
8458 struct clock *queueTime, struct clock *execTime,
8459 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8465 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8466 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8468 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8469 queueTime, execTime, sent64, rcvd64,
8476 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8480 * IN callerVersion - the rpc stat version of the caller.
8482 * IN count - the number of entries to marshall.
8484 * IN stats - pointer to stats to be marshalled.
8486 * OUT ptr - Where to store the marshalled data.
8493 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8494 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8500 * We only support the first version
8502 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8503 *(ptr++) = stats->remote_peer;
8504 *(ptr++) = stats->remote_port;
8505 *(ptr++) = stats->remote_is_server;
8506 *(ptr++) = stats->interfaceId;
8507 *(ptr++) = stats->func_total;
8508 *(ptr++) = stats->func_index;
8509 *(ptr++) = stats->invocations >> 32;
8510 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8511 *(ptr++) = stats->bytes_sent >> 32;
8512 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8513 *(ptr++) = stats->bytes_rcvd >> 32;
8514 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8515 *(ptr++) = stats->queue_time_sum.sec;
8516 *(ptr++) = stats->queue_time_sum.usec;
8517 *(ptr++) = stats->queue_time_sum_sqr.sec;
8518 *(ptr++) = stats->queue_time_sum_sqr.usec;
8519 *(ptr++) = stats->queue_time_min.sec;
8520 *(ptr++) = stats->queue_time_min.usec;
8521 *(ptr++) = stats->queue_time_max.sec;
8522 *(ptr++) = stats->queue_time_max.usec;
8523 *(ptr++) = stats->execution_time_sum.sec;
8524 *(ptr++) = stats->execution_time_sum.usec;
8525 *(ptr++) = stats->execution_time_sum_sqr.sec;
8526 *(ptr++) = stats->execution_time_sum_sqr.usec;
8527 *(ptr++) = stats->execution_time_min.sec;
8528 *(ptr++) = stats->execution_time_min.usec;
8529 *(ptr++) = stats->execution_time_max.sec;
8530 *(ptr++) = stats->execution_time_max.usec;
8536 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8541 * IN callerVersion - the rpc stat version of the caller
8543 * OUT myVersion - the rpc stat version of this function
8545 * OUT clock_sec - local time seconds
8547 * OUT clock_usec - local time microseconds
8549 * OUT allocSize - the number of bytes allocated to contain stats
8551 * OUT statCount - the number stats retrieved from this process.
8553 * OUT stats - the actual stats retrieved from this process.
8557 * Returns void. If successful, stats will != NULL.
8561 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8562 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8563 size_t * allocSize, afs_uint32 * statCount,
8564 afs_uint32 ** stats)
8574 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8577 * Check to see if stats are enabled
8580 MUTEX_ENTER(&rx_rpc_stats);
8581 if (!rxi_monitor_processStats) {
8582 MUTEX_EXIT(&rx_rpc_stats);
8586 clock_GetTime(&now);
8587 *clock_sec = now.sec;
8588 *clock_usec = now.usec;
8591 * Allocate the space based upon the caller version
8593 * If the client is at an older version than we are,
8594 * we return the statistic data in the older data format, but
8595 * we still return our version number so the client knows we
8596 * are maintaining more data than it can retrieve.
8599 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8600 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8601 *statCount = rxi_rpc_process_stat_cnt;
8604 * This can't happen yet, but in the future version changes
8605 * can be handled by adding additional code here
8609 if (space > (size_t) 0) {
8611 ptr = *stats = rxi_Alloc(space);
8614 rx_interface_stat_p rpc_stat, nrpc_stat;
8618 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8620 * Copy the data based upon the caller version
8622 rx_MarshallProcessRPCStats(callerVersion,
8623 rpc_stat->stats[0].func_total,
8624 rpc_stat->stats, &ptr);
8630 MUTEX_EXIT(&rx_rpc_stats);
8635 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8639 * IN callerVersion - the rpc stat version of the caller
8641 * OUT myVersion - the rpc stat version of this function
8643 * OUT clock_sec - local time seconds
8645 * OUT clock_usec - local time microseconds
8647 * OUT allocSize - the number of bytes allocated to contain stats
8649 * OUT statCount - the number of stats retrieved from the individual
8652 * OUT stats - the actual stats retrieved from the individual peer structures.
8656 * Returns void. If successful, stats will != NULL.
8660 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8661 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8662 size_t * allocSize, afs_uint32 * statCount,
8663 afs_uint32 ** stats)
8673 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8676 * Check to see if stats are enabled
8679 MUTEX_ENTER(&rx_rpc_stats);
8680 if (!rxi_monitor_peerStats) {
8681 MUTEX_EXIT(&rx_rpc_stats);
8685 clock_GetTime(&now);
8686 *clock_sec = now.sec;
8687 *clock_usec = now.usec;
8690 * Allocate the space based upon the caller version
8692 * If the client is at an older version than we are,
8693 * we return the statistic data in the older data format, but
8694 * we still return our version number so the client knows we
8695 * are maintaining more data than it can retrieve.
8698 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8699 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8700 *statCount = rxi_rpc_peer_stat_cnt;
8703 * This can't happen yet, but in the future version changes
8704 * can be handled by adding additional code here
8708 if (space > (size_t) 0) {
8710 ptr = *stats = rxi_Alloc(space);
8713 rx_interface_stat_p rpc_stat, nrpc_stat;
8717 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8719 * We have to fix the offset of rpc_stat since we are
8720 * keeping this structure on two rx_queues. The rx_queue
8721 * package assumes that the rx_queue member is the first
8722 * member of the structure. That is, rx_queue assumes that
8723 * any one item is only on one queue at a time. We are
8724 * breaking that assumption and so we have to do a little
8725 * math to fix our pointers.
8728 fix_offset = (char *)rpc_stat;
8729 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8730 rpc_stat = (rx_interface_stat_p) fix_offset;
8733 * Copy the data based upon the caller version
8735 rx_MarshallProcessRPCStats(callerVersion,
8736 rpc_stat->stats[0].func_total,
8737 rpc_stat->stats, &ptr);
8743 MUTEX_EXIT(&rx_rpc_stats);
8748 * rx_FreeRPCStats - free memory allocated by
8749 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8753 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8754 * rx_RetrievePeerRPCStats
8756 * IN allocSize - the number of bytes in stats.
8764 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8766 rxi_Free(stats, allocSize);
8770 * rx_queryProcessRPCStats - see if process rpc stat collection is
8771 * currently enabled.
8777 * Returns 0 if stats are not enabled != 0 otherwise
8781 rx_queryProcessRPCStats(void)
8784 MUTEX_ENTER(&rx_rpc_stats);
8785 rc = rxi_monitor_processStats;
8786 MUTEX_EXIT(&rx_rpc_stats);
8791 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8797 * Returns 0 if stats are not enabled != 0 otherwise
8801 rx_queryPeerRPCStats(void)
8804 MUTEX_ENTER(&rx_rpc_stats);
8805 rc = rxi_monitor_peerStats;
8806 MUTEX_EXIT(&rx_rpc_stats);
8811 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8821 rx_enableProcessRPCStats(void)
8823 MUTEX_ENTER(&rx_rpc_stats);
8824 rx_enable_stats = 1;
8825 rxi_monitor_processStats = 1;
8826 MUTEX_EXIT(&rx_rpc_stats);
8830 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8840 rx_enablePeerRPCStats(void)
8842 MUTEX_ENTER(&rx_rpc_stats);
8843 rx_enable_stats = 1;
8844 rxi_monitor_peerStats = 1;
8845 MUTEX_EXIT(&rx_rpc_stats);
8849 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8859 rx_disableProcessRPCStats(void)
8861 rx_interface_stat_p rpc_stat, nrpc_stat;
8864 MUTEX_ENTER(&rx_rpc_stats);
8867 * Turn off process statistics and if peer stats is also off, turn
8871 rxi_monitor_processStats = 0;
8872 if (rxi_monitor_peerStats == 0) {
8873 rx_enable_stats = 0;
8876 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8877 unsigned int num_funcs = 0;
8880 queue_Remove(rpc_stat);
8881 num_funcs = rpc_stat->stats[0].func_total;
8883 sizeof(rx_interface_stat_t) +
8884 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8886 rxi_Free(rpc_stat, space);
8887 rxi_rpc_process_stat_cnt -= num_funcs;
8889 MUTEX_EXIT(&rx_rpc_stats);
8893 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8903 rx_disablePeerRPCStats(void)
8905 struct rx_peer **peer_ptr, **peer_end;
8909 * Turn off peer statistics and if process stats is also off, turn
8913 rxi_monitor_peerStats = 0;
8914 if (rxi_monitor_processStats == 0) {
8915 rx_enable_stats = 0;
8918 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8919 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8921 struct rx_peer *peer, *next, *prev;
8923 MUTEX_ENTER(&rx_peerHashTable_lock);
8924 MUTEX_ENTER(&rx_rpc_stats);
8925 for (prev = peer = *peer_ptr; peer; peer = next) {
8927 code = MUTEX_TRYENTER(&peer->peer_lock);
8929 rx_interface_stat_p rpc_stat, nrpc_stat;
8932 if (prev == *peer_ptr) {
8943 MUTEX_EXIT(&rx_peerHashTable_lock);
8946 (&peer->rpcStats, rpc_stat, nrpc_stat,
8947 rx_interface_stat)) {
8948 unsigned int num_funcs = 0;
8951 queue_Remove(&rpc_stat->queue_header);
8952 queue_Remove(&rpc_stat->all_peers);
8953 num_funcs = rpc_stat->stats[0].func_total;
8955 sizeof(rx_interface_stat_t) +
8956 rpc_stat->stats[0].func_total *
8957 sizeof(rx_function_entry_v1_t);
8959 rxi_Free(rpc_stat, space);
8960 rxi_rpc_peer_stat_cnt -= num_funcs;
8962 MUTEX_EXIT(&peer->peer_lock);
8964 MUTEX_ENTER(&rx_peerHashTable_lock);
8974 MUTEX_EXIT(&rx_rpc_stats);
8975 MUTEX_EXIT(&rx_peerHashTable_lock);
8980 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8985 * IN clearFlag - flag indicating which stats to clear
8993 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8995 rx_interface_stat_p rpc_stat, nrpc_stat;
8997 MUTEX_ENTER(&rx_rpc_stats);
8999 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
9000 unsigned int num_funcs = 0, i;
9001 num_funcs = rpc_stat->stats[0].func_total;
9002 for (i = 0; i < num_funcs; i++) {
9003 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9004 rpc_stat->stats[i].invocations = 0;
9006 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9007 rpc_stat->stats[i].bytes_sent = 0;
9009 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9010 rpc_stat->stats[i].bytes_rcvd = 0;
9012 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9013 rpc_stat->stats[i].queue_time_sum.sec = 0;
9014 rpc_stat->stats[i].queue_time_sum.usec = 0;
9016 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9017 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9018 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9020 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9021 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9022 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9024 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9025 rpc_stat->stats[i].queue_time_max.sec = 0;
9026 rpc_stat->stats[i].queue_time_max.usec = 0;
9028 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9029 rpc_stat->stats[i].execution_time_sum.sec = 0;
9030 rpc_stat->stats[i].execution_time_sum.usec = 0;
9032 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9033 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9034 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9036 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9037 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9038 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9040 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9041 rpc_stat->stats[i].execution_time_max.sec = 0;
9042 rpc_stat->stats[i].execution_time_max.usec = 0;
9047 MUTEX_EXIT(&rx_rpc_stats);
9051 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9056 * IN clearFlag - flag indicating which stats to clear
9064 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9066 rx_interface_stat_p rpc_stat, nrpc_stat;
9068 MUTEX_ENTER(&rx_rpc_stats);
9070 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
9071 unsigned int num_funcs = 0, i;
9074 * We have to fix the offset of rpc_stat since we are
9075 * keeping this structure on two rx_queues. The rx_queue
9076 * package assumes that the rx_queue member is the first
9077 * member of the structure. That is, rx_queue assumes that
9078 * any one item is only on one queue at a time. We are
9079 * breaking that assumption and so we have to do a little
9080 * math to fix our pointers.
9083 fix_offset = (char *)rpc_stat;
9084 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
9085 rpc_stat = (rx_interface_stat_p) fix_offset;
9087 num_funcs = rpc_stat->stats[0].func_total;
9088 for (i = 0; i < num_funcs; i++) {
9089 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9090 rpc_stat->stats[i].invocations = 0;
9092 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9093 rpc_stat->stats[i].bytes_sent = 0;
9095 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9096 rpc_stat->stats[i].bytes_rcvd = 0;
9098 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9099 rpc_stat->stats[i].queue_time_sum.sec = 0;
9100 rpc_stat->stats[i].queue_time_sum.usec = 0;
9102 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9103 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9104 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9106 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9107 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9108 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9110 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9111 rpc_stat->stats[i].queue_time_max.sec = 0;
9112 rpc_stat->stats[i].queue_time_max.usec = 0;
9114 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9115 rpc_stat->stats[i].execution_time_sum.sec = 0;
9116 rpc_stat->stats[i].execution_time_sum.usec = 0;
9118 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9119 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9120 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9122 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9123 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9124 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9126 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9127 rpc_stat->stats[i].execution_time_max.sec = 0;
9128 rpc_stat->stats[i].execution_time_max.usec = 0;
9133 MUTEX_EXIT(&rx_rpc_stats);
9137 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9138 * is authorized to enable/disable/clear RX statistics.
9140 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9143 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9145 rxi_rxstat_userok = proc;
9149 rx_RxStatUserOk(struct rx_call *call)
9151 if (!rxi_rxstat_userok)
9153 return rxi_rxstat_userok(call);
9158 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9159 * function in the MSVC runtime DLL (msvcrt.dll).
9161 * Note: the system serializes calls to this function.
9164 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9165 DWORD reason, /* reason function is being called */
9166 LPVOID reserved) /* reserved for future use */
9169 case DLL_PROCESS_ATTACH:
9170 /* library is being attached to a process */
9174 case DLL_PROCESS_DETACH:
9181 #endif /* AFS_NT40_ENV */
9184 int rx_DumpCalls(FILE *outputFile, char *cookie)
9186 #ifdef RXDEBUG_PACKET
9187 #ifdef KDUMP_RX_LOCK
9188 struct rx_call_rx_lock *c;
9195 #define RXDPRINTF sprintf
9196 #define RXDPRINTOUT output
9198 #define RXDPRINTF fprintf
9199 #define RXDPRINTOUT outputFile
9202 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9204 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9207 for (c = rx_allCallsp; c; c = c->allNextp) {
9208 u_short rqc, tqc, iovqc;
9209 struct rx_packet *p, *np;
9211 MUTEX_ENTER(&c->lock);
9212 queue_Count(&c->rq, p, np, rx_packet, rqc);
9213 queue_Count(&c->tq, p, np, rx_packet, tqc);
9214 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9216 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, "
9217 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9218 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9219 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9220 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9221 #ifdef RX_ENABLE_LOCKS
9224 #ifdef RX_REFCOUNT_CHECK
9225 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9226 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9229 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
9230 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9231 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9232 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9233 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9234 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9235 #ifdef RX_ENABLE_LOCKS
9236 , (afs_uint32)c->refCount
9238 #ifdef RX_REFCOUNT_CHECK
9239 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9242 MUTEX_EXIT(&c->lock);
9245 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9248 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9250 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9252 #endif /* RXDEBUG_PACKET */