2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
75 #include <opr/queue.h>
79 #include "rx_atomic.h"
80 #include "rx_globals.h"
82 #include "rx_internal.h"
89 #include "rx_packet.h"
90 #include "rx_server.h"
92 #include <afs/rxgen_consts.h>
95 #ifdef AFS_PTHREAD_ENV
97 int (*registerProgram) (pid_t, char *) = 0;
98 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
101 int (*registerProgram) (PROCESS, char *) = 0;
102 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
106 /* Local static routines */
107 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
108 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
109 struct rx_call *, struct rx_peer *,
111 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
113 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
114 void *dummy, int dummy2);
115 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
116 void *dummy, int dummy2);
117 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
118 void *unused, int unused2);
119 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
120 void *unused2, int unused3);
121 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
122 struct rx_packet *packet,
123 int istack, int force);
124 static void rxi_AckAll(struct rx_call *call);
125 static struct rx_connection
126 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
127 u_short serviceId, afs_uint32 cid,
128 afs_uint32 epoch, int type, u_int securityIndex);
129 static struct rx_packet
130 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
131 int istack, osi_socket socket,
132 afs_uint32 host, u_short port, int *tnop,
133 struct rx_call **newcallp);
134 static struct rx_packet
135 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
137 static struct rx_packet
138 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
139 struct rx_packet *np, int istack);
140 static struct rx_packet
141 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
142 struct rx_packet *np, int istack);
143 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
144 int *tnop, struct rx_call **newcallp);
145 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
146 static void rxi_ClearReceiveQueue(struct rx_call *call);
147 static void rxi_ResetCall(struct rx_call *call, int newcall);
148 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
149 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
150 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
151 static void rxi_KeepAliveOn(struct rx_call *call);
152 static void rxi_GrowMTUOn(struct rx_call *call);
153 static void rxi_ChallengeOn(struct rx_connection *conn);
154 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
155 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
157 #ifdef RX_ENABLE_LOCKS
159 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
160 rx_atomic_t rxi_start_in_error;
162 #endif /* RX_ENABLE_LOCKS */
164 /* Constant delay time before sending an acknowledge of the last packet
165 * received. This is to avoid sending an extra acknowledge when the
166 * client is about to make another call, anyway, or the server is
169 * The lastAckDelay may not exceeed 400ms without causing peers to
170 * unecessarily timeout.
172 struct clock rx_lastAckDelay = {0, 400000};
174 /* Constant delay time before sending a soft ack when none was requested.
175 * This is to make sure we send soft acks before the sender times out,
176 * Normally we wait and send a hard ack when the receiver consumes the packet
178 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
179 * will require changes to the peer's RTT calculations.
181 struct clock rx_softAckDelay = {0, 100000};
184 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
185 * currently allocated within rx. This number is used to allocate the
186 * memory required to return the statistics when queried.
187 * Protected by the rx_rpc_stats mutex.
190 static unsigned int rxi_rpc_peer_stat_cnt;
193 * rxi_rpc_process_stat_cnt counts the total number of local process stat
194 * structures currently allocated within rx. The number is used to allocate
195 * the memory required to return the statistics when queried.
196 * Protected by the rx_rpc_stats mutex.
199 static unsigned int rxi_rpc_process_stat_cnt;
202 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
203 * errors should be reported to the application when a call channel appears busy
204 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
205 * and there are other call channels in the connection that are not busy.
206 * If 0, we do not return errors upon receiving busy packets; we just keep
207 * trying on the same call channel until we hit a timeout.
209 static afs_int32 rxi_busyChannelError = 0;
211 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
212 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
214 /* Incoming calls wait on this queue when there are no available
215 * server processes */
216 struct opr_queue rx_incomingCallQueue;
218 /* Server processes wait on this queue when there are no appropriate
219 * calls to process */
220 struct opr_queue rx_idleServerQueue;
222 #if !defined(offsetof)
223 #include <stddef.h> /* for definition of offsetof() */
226 #ifdef RX_ENABLE_LOCKS
227 afs_kmutex_t rx_atomic_mutex;
230 /* Forward prototypes */
231 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
234 putConnection (struct rx_connection *conn) {
235 MUTEX_ENTER(&rx_refcnt_mutex);
237 MUTEX_EXIT(&rx_refcnt_mutex);
240 #ifdef AFS_PTHREAD_ENV
243 * Use procedural initialization of mutexes/condition variables
247 extern afs_kmutex_t rx_quota_mutex;
248 extern afs_kmutex_t rx_pthread_mutex;
249 extern afs_kmutex_t rx_packets_mutex;
250 extern afs_kmutex_t rx_refcnt_mutex;
251 extern afs_kmutex_t des_init_mutex;
252 extern afs_kmutex_t des_random_mutex;
253 extern afs_kmutex_t rx_clock_mutex;
254 extern afs_kmutex_t rxi_connCacheMutex;
255 extern afs_kmutex_t event_handler_mutex;
256 extern afs_kmutex_t listener_mutex;
257 extern afs_kmutex_t rx_if_init_mutex;
258 extern afs_kmutex_t rx_if_mutex;
260 extern afs_kcondvar_t rx_event_handler_cond;
261 extern afs_kcondvar_t rx_listener_cond;
263 static afs_kmutex_t epoch_mutex;
264 static afs_kmutex_t rx_init_mutex;
265 static afs_kmutex_t rx_debug_mutex;
266 static afs_kmutex_t rx_rpc_stats;
269 rxi_InitPthread(void)
271 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
273 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
287 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
288 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
290 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
291 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
293 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
294 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
295 #ifdef RX_ENABLE_LOCKS
298 #endif /* RX_LOCKS_DB */
299 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
300 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
302 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
304 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
306 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
308 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
309 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
310 #endif /* RX_ENABLE_LOCKS */
313 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
314 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
316 * The rx_stats_mutex mutex protects the following global variables:
317 * rxi_lowConnRefCount
318 * rxi_lowPeerRefCount
327 * The rx_quota_mutex mutex protects the following global variables:
335 * The rx_freePktQ_lock protects the following global variables:
340 * The rx_packets_mutex mutex protects the following global variables:
348 * The rx_pthread_mutex mutex protects the following global variables:
349 * rxi_fcfs_thread_num
352 #define INIT_PTHREAD_LOCKS
356 /* Variables for handling the minProcs implementation. availProcs gives the
357 * number of threads available in the pool at this moment (not counting dudes
358 * executing right now). totalMin gives the total number of procs required
359 * for handling all minProcs requests. minDeficit is a dynamic variable
360 * tracking the # of procs required to satisfy all of the remaining minProcs
362 * For fine grain locking to work, the quota check and the reservation of
363 * a server thread has to come while rxi_availProcs and rxi_minDeficit
364 * are locked. To this end, the code has been modified under #ifdef
365 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
366 * same time. A new function, ReturnToServerPool() returns the allocation.
368 * A call can be on several queue's (but only one at a time). When
369 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
370 * that no one else is touching the queue. To this end, we store the address
371 * of the queue lock in the call structure (under the call lock) when we
372 * put the call on a queue, and we clear the call_queue_lock when the
373 * call is removed from a queue (once the call lock has been obtained).
374 * This allows rxi_ResetCall to safely synchronize with others wishing
375 * to manipulate the queue.
378 #if defined(RX_ENABLE_LOCKS)
379 static afs_kmutex_t rx_rpc_stats;
382 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
383 ** pretty good that the next packet coming in is from the same connection
384 ** as the last packet, since we're send multiple packets in a transmit window.
386 struct rx_connection *rxLastConn = 0;
388 #ifdef RX_ENABLE_LOCKS
389 /* The locking hierarchy for rx fine grain locking is composed of these
392 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
393 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
394 * call->lock - locks call data fields.
395 * These are independent of each other:
396 * rx_freeCallQueue_lock
401 * serverQueueEntry->lock
402 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
404 * peer->lock - locks peer data fields.
405 * conn_data_lock - that more than one thread is not updating a conn data
406 * field at the same time.
417 * Do we need a lock to protect the peer field in the conn structure?
418 * conn->peer was previously a constant for all intents and so has no
419 * lock protecting this field. The multihomed client delta introduced
420 * a RX code change : change the peer field in the connection structure
421 * to that remote interface from which the last packet for this
422 * connection was sent out. This may become an issue if further changes
425 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
426 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
428 /* rxdb_fileID is used to identify the lock location, along with line#. */
429 static int rxdb_fileID = RXDB_FILE_RX;
430 #endif /* RX_LOCKS_DB */
431 #else /* RX_ENABLE_LOCKS */
432 #define SET_CALL_QUEUE_LOCK(C, L)
433 #define CLEAR_CALL_QUEUE_LOCK(C)
434 #endif /* RX_ENABLE_LOCKS */
435 struct rx_serverQueueEntry *rx_waitForPacket = 0;
436 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
438 /* ------------Exported Interfaces------------- */
440 /* This function allows rxkad to set the epoch to a suitably random number
441 * which rx_NewConnection will use in the future. The principle purpose is to
442 * get rxnull connections to use the same epoch as the rxkad connections do, at
443 * least once the first rxkad connection is established. This is important now
444 * that the host/port addresses aren't used in FindConnection: the uniqueness
445 * of epoch/cid matters and the start time won't do. */
447 #ifdef AFS_PTHREAD_ENV
449 * This mutex protects the following global variables:
453 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
454 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
458 #endif /* AFS_PTHREAD_ENV */
461 rx_SetEpoch(afs_uint32 epoch)
468 /* Initialize rx. A port number may be mentioned, in which case this
469 * becomes the default port number for any service installed later.
470 * If 0 is provided for the port number, a random port will be chosen
471 * by the kernel. Whether this will ever overlap anything in
472 * /etc/services is anybody's guess... Returns 0 on success, -1 on
477 int rxinit_status = 1;
478 #ifdef AFS_PTHREAD_ENV
480 * This mutex protects the following global variables:
484 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
485 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
488 #define UNLOCK_RX_INIT
492 rx_InitHost(u_int host, u_int port)
499 char *htable, *ptable;
506 if (rxinit_status == 0) {
507 tmp_status = rxinit_status;
509 return tmp_status; /* Already started; return previous error code. */
515 if (afs_winsockInit() < 0)
521 * Initialize anything necessary to provide a non-premptive threading
524 rxi_InitializeThreadSupport();
527 /* Allocate and initialize a socket for client and perhaps server
530 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
531 if (rx_socket == OSI_NULLSOCKET) {
535 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
538 #endif /* RX_LOCKS_DB */
539 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
540 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
541 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
542 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
543 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
544 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
545 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
546 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
547 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
548 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
550 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
552 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
554 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
556 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
557 #if defined(AFS_HPUX110_ENV)
559 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
560 #endif /* AFS_HPUX110_ENV */
561 #endif /* RX_ENABLE_LOCKS && KERNEL */
564 rx_connDeadTime = 12;
565 rx_tranquil = 0; /* reset flag */
566 rxi_ResetStatistics();
567 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
568 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
569 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
570 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
571 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
572 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
574 /* Malloc up a bunch of packets & buffers */
576 opr_queue_Init(&rx_freePacketQueue);
577 rxi_NeedMorePackets = FALSE;
578 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
580 /* enforce a minimum number of allocated packets */
581 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
582 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
584 /* allocate the initial free packet pool */
585 #ifdef RX_ENABLE_TSFPQ
586 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
587 #else /* RX_ENABLE_TSFPQ */
588 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
589 #endif /* RX_ENABLE_TSFPQ */
596 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
597 tv.tv_sec = clock_now.sec;
598 tv.tv_usec = clock_now.usec;
599 srand((unsigned int)tv.tv_usec);
606 #if defined(KERNEL) && !defined(UKERNEL)
607 /* Really, this should never happen in a real kernel */
610 struct sockaddr_in addr;
612 int addrlen = sizeof(addr);
614 socklen_t addrlen = sizeof(addr);
616 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
618 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
621 rx_port = addr.sin_port;
624 rx_stats.minRtt.sec = 9999999;
626 rx_SetEpoch(tv.tv_sec | 0x80000000);
628 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
629 * will provide a randomer value. */
631 MUTEX_ENTER(&rx_quota_mutex);
632 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
633 MUTEX_EXIT(&rx_quota_mutex);
634 /* *Slightly* random start time for the cid. This is just to help
635 * out with the hashing function at the peer */
636 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
637 rx_connHashTable = (struct rx_connection **)htable;
638 rx_peerHashTable = (struct rx_peer **)ptable;
640 rx_hardAckDelay.sec = 0;
641 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
643 rxevent_Init(20, rxi_ReScheduleEvents);
645 /* Initialize various global queues */
646 opr_queue_Init(&rx_idleServerQueue);
647 opr_queue_Init(&rx_incomingCallQueue);
648 opr_queue_Init(&rx_freeCallQueue);
650 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
651 /* Initialize our list of usable IP addresses. */
655 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
656 /* Start listener process (exact function is dependent on the
657 * implementation environment--kernel or user space) */
662 tmp_status = rxinit_status = 0;
670 return rx_InitHost(htonl(INADDR_ANY), port);
676 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
677 * maintaing the round trip timer.
682 * Start a new RTT timer for a given call and packet.
684 * There must be no resendEvent already listed for this call, otherwise this
685 * will leak events - intended for internal use within the RTO code only
688 * the RX call to start the timer for
689 * @param[in] lastPacket
690 * a flag indicating whether the last packet has been sent or not
692 * @pre call must be locked before calling this function
696 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
698 struct clock now, retryTime;
703 clock_Add(&retryTime, &call->rto);
705 /* If we're sending the last packet, and we're the client, then the server
706 * may wait for an additional 400ms before returning the ACK, wait for it
707 * rather than hitting a timeout */
708 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
709 clock_Addmsec(&retryTime, 400);
711 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
712 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
717 * Cancel an RTT timer for a given call.
721 * the RX call to cancel the timer for
723 * @pre call must be locked before calling this function
728 rxi_rto_cancel(struct rx_call *call)
730 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
734 * Tell the RTO timer that we have sent a packet.
736 * If the timer isn't already running, then start it. If the timer is running,
740 * the RX call that the packet has been sent on
741 * @param[in] lastPacket
742 * A flag which is true if this is the last packet for the call
744 * @pre The call must be locked before calling this function
749 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
751 if (call->resendEvent)
754 rxi_rto_startTimer(call, lastPacket, istack);
758 * Tell the RTO timer that we have received an new ACK message
760 * This function should be called whenever a call receives an ACK that
761 * acknowledges new packets. Whatever happens, we stop the current timer.
762 * If there are unacked packets in the queue which have been sent, then
763 * we restart the timer from now. Otherwise, we leave it stopped.
766 * the RX call that the ACK has been received on
770 rxi_rto_packet_acked(struct rx_call *call, int istack)
772 struct opr_queue *cursor;
774 rxi_rto_cancel(call);
776 if (opr_queue_IsEmpty(&call->tq))
779 for (opr_queue_Scan(&call->tq, cursor)) {
780 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
781 if (p->header.seq > call->tfirst + call->twind)
784 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
785 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
793 * Set an initial round trip timeout for a peer connection
795 * @param[in] secs The timeout to set in seconds
799 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
800 peer->rtt = secs * 8000;
804 * Enables or disables the busy call channel error (RX_CALL_BUSY).
806 * @param[in] onoff Non-zero to enable busy call channel errors.
808 * @pre Neither rx_Init nor rx_InitHost have been called yet
811 rx_SetBusyChannelError(afs_int32 onoff)
813 osi_Assert(rxinit_status != 0);
814 rxi_busyChannelError = onoff ? 1 : 0;
818 * Set a delayed ack event on the specified call for the given time
820 * @param[in] call - the call on which to set the event
821 * @param[in] offset - the delay from now after which the event fires
824 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
826 struct clock now, when;
830 clock_Add(&when, offset);
832 if (!call->delayedAckEvent
833 || clock_Gt(&call->delayedAckTime, &when)) {
835 rxevent_Cancel(&call->delayedAckEvent, call,
836 RX_CALL_REFCOUNT_DELAY);
837 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
839 call->delayedAckEvent = rxevent_Post(&when, &now,
842 call->delayedAckTime = when;
846 /* called with unincremented nRequestsRunning to see if it is OK to start
847 * a new thread in this service. Could be "no" for two reasons: over the
848 * max quota, or would prevent others from reaching their min quota.
850 #ifdef RX_ENABLE_LOCKS
851 /* This verion of QuotaOK reserves quota if it's ok while the
852 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
855 QuotaOK(struct rx_service *aservice)
857 /* check if over max quota */
858 if (aservice->nRequestsRunning >= aservice->maxProcs) {
862 /* under min quota, we're OK */
863 /* otherwise, can use only if there are enough to allow everyone
864 * to go to their min quota after this guy starts.
867 MUTEX_ENTER(&rx_quota_mutex);
868 if ((aservice->nRequestsRunning < aservice->minProcs)
869 || (rxi_availProcs > rxi_minDeficit)) {
870 aservice->nRequestsRunning++;
871 /* just started call in minProcs pool, need fewer to maintain
873 if (aservice->nRequestsRunning <= aservice->minProcs)
876 MUTEX_EXIT(&rx_quota_mutex);
879 MUTEX_EXIT(&rx_quota_mutex);
885 ReturnToServerPool(struct rx_service *aservice)
887 aservice->nRequestsRunning--;
888 MUTEX_ENTER(&rx_quota_mutex);
889 if (aservice->nRequestsRunning < aservice->minProcs)
892 MUTEX_EXIT(&rx_quota_mutex);
895 #else /* RX_ENABLE_LOCKS */
897 QuotaOK(struct rx_service *aservice)
900 /* under min quota, we're OK */
901 if (aservice->nRequestsRunning < aservice->minProcs)
904 /* check if over max quota */
905 if (aservice->nRequestsRunning >= aservice->maxProcs)
908 /* otherwise, can use only if there are enough to allow everyone
909 * to go to their min quota after this guy starts.
911 MUTEX_ENTER(&rx_quota_mutex);
912 if (rxi_availProcs > rxi_minDeficit)
914 MUTEX_EXIT(&rx_quota_mutex);
917 #endif /* RX_ENABLE_LOCKS */
920 /* Called by rx_StartServer to start up lwp's to service calls.
921 NExistingProcs gives the number of procs already existing, and which
922 therefore needn't be created. */
924 rxi_StartServerProcs(int nExistingProcs)
926 struct rx_service *service;
931 /* For each service, reserve N processes, where N is the "minimum"
932 * number of processes that MUST be able to execute a request in parallel,
933 * at any time, for that process. Also compute the maximum difference
934 * between any service's maximum number of processes that can run
935 * (i.e. the maximum number that ever will be run, and a guarantee
936 * that this number will run if other services aren't running), and its
937 * minimum number. The result is the extra number of processes that
938 * we need in order to provide the latter guarantee */
939 for (i = 0; i < RX_MAX_SERVICES; i++) {
941 service = rx_services[i];
942 if (service == (struct rx_service *)0)
944 nProcs += service->minProcs;
945 diff = service->maxProcs - service->minProcs;
949 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
950 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
951 for (i = 0; i < nProcs; i++) {
952 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
958 /* This routine is only required on Windows */
960 rx_StartClientThread(void)
962 #ifdef AFS_PTHREAD_ENV
964 pid = pthread_self();
965 #endif /* AFS_PTHREAD_ENV */
967 #endif /* AFS_NT40_ENV */
969 /* This routine must be called if any services are exported. If the
970 * donateMe flag is set, the calling process is donated to the server
973 rx_StartServer(int donateMe)
975 struct rx_service *service;
981 /* Start server processes, if necessary (exact function is dependent
982 * on the implementation environment--kernel or user space). DonateMe
983 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
984 * case, one less new proc will be created rx_StartServerProcs.
986 rxi_StartServerProcs(donateMe);
988 /* count up the # of threads in minProcs, and add set the min deficit to
989 * be that value, too.
991 for (i = 0; i < RX_MAX_SERVICES; i++) {
992 service = rx_services[i];
993 if (service == (struct rx_service *)0)
995 MUTEX_ENTER(&rx_quota_mutex);
996 rxi_totalMin += service->minProcs;
997 /* below works even if a thread is running, since minDeficit would
998 * still have been decremented and later re-incremented.
1000 rxi_minDeficit += service->minProcs;
1001 MUTEX_EXIT(&rx_quota_mutex);
1004 /* Turn on reaping of idle server connections */
1005 rxi_ReapConnections(NULL, NULL, NULL, 0);
1010 #ifndef AFS_NT40_ENV
1014 #ifdef AFS_PTHREAD_ENV
1016 pid = afs_pointer_to_int(pthread_self());
1017 #else /* AFS_PTHREAD_ENV */
1019 LWP_CurrentProcess(&pid);
1020 #endif /* AFS_PTHREAD_ENV */
1022 sprintf(name, "srv_%d", ++nProcs);
1023 if (registerProgram)
1024 (*registerProgram) (pid, name);
1026 #endif /* AFS_NT40_ENV */
1027 rx_ServerProc(NULL); /* Never returns */
1029 #ifdef RX_ENABLE_TSFPQ
1030 /* no use leaving packets around in this thread's local queue if
1031 * it isn't getting donated to the server thread pool.
1033 rxi_FlushLocalPacketsTSFPQ();
1034 #endif /* RX_ENABLE_TSFPQ */
1038 /* Create a new client connection to the specified service, using the
1039 * specified security object to implement the security model for this
1041 struct rx_connection *
1042 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1043 struct rx_securityClass *securityObject,
1044 int serviceSecurityIndex)
1048 struct rx_connection *conn;
1053 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1054 "serviceSecurityIndex %d)\n",
1055 ntohl(shost), ntohs(sport), sservice, securityObject,
1056 serviceSecurityIndex));
1058 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1059 * the case of kmem_alloc? */
1060 conn = rxi_AllocConnection();
1061 #ifdef RX_ENABLE_LOCKS
1062 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1063 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1064 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1067 MUTEX_ENTER(&rx_connHashTable_lock);
1068 cid = (rx_nextCid += RX_MAXCALLS);
1069 conn->type = RX_CLIENT_CONNECTION;
1071 conn->epoch = rx_epoch;
1072 conn->peer = rxi_FindPeer(shost, sport, 1);
1073 conn->serviceId = sservice;
1074 conn->securityObject = securityObject;
1075 conn->securityData = (void *) 0;
1076 conn->securityIndex = serviceSecurityIndex;
1077 rx_SetConnDeadTime(conn, rx_connDeadTime);
1078 rx_SetConnSecondsUntilNatPing(conn, 0);
1079 conn->ackRate = RX_FAST_ACK_RATE;
1080 conn->nSpecific = 0;
1081 conn->specific = NULL;
1082 conn->challengeEvent = NULL;
1083 conn->delayedAbortEvent = NULL;
1084 conn->abortCount = 0;
1086 for (i = 0; i < RX_MAXCALLS; i++) {
1087 conn->twind[i] = rx_initSendWindow;
1088 conn->rwind[i] = rx_initReceiveWindow;
1089 conn->lastBusy[i] = 0;
1092 RXS_NewConnection(securityObject, conn);
1094 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1096 conn->refCount++; /* no lock required since only this thread knows... */
1097 conn->next = rx_connHashTable[hashindex];
1098 rx_connHashTable[hashindex] = conn;
1099 if (rx_stats_active)
1100 rx_atomic_inc(&rx_stats.nClientConns);
1101 MUTEX_EXIT(&rx_connHashTable_lock);
1107 * Ensure a connection's timeout values are valid.
1109 * @param[in] conn The connection to check
1111 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1112 * unless idleDeadTime and/or hardDeadTime are not set
1116 rxi_CheckConnTimeouts(struct rx_connection *conn)
1118 /* a connection's timeouts must have the relationship
1119 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1120 * total loss of network to a peer may cause an idle timeout instead of a
1121 * dead timeout, simply because the idle timeout gets hit first. Also set
1122 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1123 /* this logic is slightly complicated by the fact that
1124 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1126 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1127 if (conn->idleDeadTime) {
1128 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1130 if (conn->hardDeadTime) {
1131 if (conn->idleDeadTime) {
1132 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1134 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1140 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1142 /* The idea is to set the dead time to a value that allows several
1143 * keepalives to be dropped without timing out the connection. */
1144 conn->secondsUntilDead = seconds;
1145 rxi_CheckConnTimeouts(conn);
1146 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1150 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1152 conn->hardDeadTime = seconds;
1153 rxi_CheckConnTimeouts(conn);
1157 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1159 conn->idleDeadTime = seconds;
1160 conn->idleDeadDetection = (seconds ? 1 : 0);
1161 rxi_CheckConnTimeouts(conn);
1164 int rxi_lowPeerRefCount = 0;
1165 int rxi_lowConnRefCount = 0;
1168 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1169 * NOTE: must not be called with rx_connHashTable_lock held.
1172 rxi_CleanupConnection(struct rx_connection *conn)
1174 /* Notify the service exporter, if requested, that this connection
1175 * is being destroyed */
1176 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1177 (*conn->service->destroyConnProc) (conn);
1179 /* Notify the security module that this connection is being destroyed */
1180 RXS_DestroyConnection(conn->securityObject, conn);
1182 /* If this is the last connection using the rx_peer struct, set its
1183 * idle time to now. rxi_ReapConnections will reap it if it's still
1184 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1186 MUTEX_ENTER(&rx_peerHashTable_lock);
1187 if (conn->peer->refCount < 2) {
1188 conn->peer->idleWhen = clock_Sec();
1189 if (conn->peer->refCount < 1) {
1190 conn->peer->refCount = 1;
1191 if (rx_stats_active) {
1192 MUTEX_ENTER(&rx_stats_mutex);
1193 rxi_lowPeerRefCount++;
1194 MUTEX_EXIT(&rx_stats_mutex);
1198 conn->peer->refCount--;
1199 MUTEX_EXIT(&rx_peerHashTable_lock);
1201 if (rx_stats_active)
1203 if (conn->type == RX_SERVER_CONNECTION)
1204 rx_atomic_dec(&rx_stats.nServerConns);
1206 rx_atomic_dec(&rx_stats.nClientConns);
1209 if (conn->specific) {
1211 for (i = 0; i < conn->nSpecific; i++) {
1212 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1213 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1214 conn->specific[i] = NULL;
1216 free(conn->specific);
1218 conn->specific = NULL;
1219 conn->nSpecific = 0;
1220 #endif /* !KERNEL */
1222 MUTEX_DESTROY(&conn->conn_call_lock);
1223 MUTEX_DESTROY(&conn->conn_data_lock);
1224 CV_DESTROY(&conn->conn_call_cv);
1226 rxi_FreeConnection(conn);
1229 /* Destroy the specified connection */
1231 rxi_DestroyConnection(struct rx_connection *conn)
1233 MUTEX_ENTER(&rx_connHashTable_lock);
1234 rxi_DestroyConnectionNoLock(conn);
1235 /* conn should be at the head of the cleanup list */
1236 if (conn == rx_connCleanup_list) {
1237 rx_connCleanup_list = rx_connCleanup_list->next;
1238 MUTEX_EXIT(&rx_connHashTable_lock);
1239 rxi_CleanupConnection(conn);
1241 #ifdef RX_ENABLE_LOCKS
1243 MUTEX_EXIT(&rx_connHashTable_lock);
1245 #endif /* RX_ENABLE_LOCKS */
1249 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1251 struct rx_connection **conn_ptr;
1253 struct rx_packet *packet;
1260 MUTEX_ENTER(&conn->conn_data_lock);
1261 MUTEX_ENTER(&rx_refcnt_mutex);
1262 if (conn->refCount > 0)
1265 if (rx_stats_active) {
1266 MUTEX_ENTER(&rx_stats_mutex);
1267 rxi_lowConnRefCount++;
1268 MUTEX_EXIT(&rx_stats_mutex);
1272 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1273 /* Busy; wait till the last guy before proceeding */
1274 MUTEX_EXIT(&rx_refcnt_mutex);
1275 MUTEX_EXIT(&conn->conn_data_lock);
1280 /* If the client previously called rx_NewCall, but it is still
1281 * waiting, treat this as a running call, and wait to destroy the
1282 * connection later when the call completes. */
1283 if ((conn->type == RX_CLIENT_CONNECTION)
1284 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1285 conn->flags |= RX_CONN_DESTROY_ME;
1286 MUTEX_EXIT(&conn->conn_data_lock);
1290 MUTEX_EXIT(&rx_refcnt_mutex);
1291 MUTEX_EXIT(&conn->conn_data_lock);
1293 /* Check for extant references to this connection */
1294 MUTEX_ENTER(&conn->conn_call_lock);
1295 for (i = 0; i < RX_MAXCALLS; i++) {
1296 struct rx_call *call = conn->call[i];
1299 if (conn->type == RX_CLIENT_CONNECTION) {
1300 MUTEX_ENTER(&call->lock);
1301 if (call->delayedAckEvent) {
1302 /* Push the final acknowledgment out now--there
1303 * won't be a subsequent call to acknowledge the
1304 * last reply packets */
1305 rxevent_Cancel(&call->delayedAckEvent, call,
1306 RX_CALL_REFCOUNT_DELAY);
1307 if (call->state == RX_STATE_PRECALL
1308 || call->state == RX_STATE_ACTIVE) {
1309 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1314 MUTEX_EXIT(&call->lock);
1318 MUTEX_EXIT(&conn->conn_call_lock);
1320 #ifdef RX_ENABLE_LOCKS
1322 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1323 MUTEX_EXIT(&conn->conn_data_lock);
1325 /* Someone is accessing a packet right now. */
1329 #endif /* RX_ENABLE_LOCKS */
1332 /* Don't destroy the connection if there are any call
1333 * structures still in use */
1334 MUTEX_ENTER(&conn->conn_data_lock);
1335 conn->flags |= RX_CONN_DESTROY_ME;
1336 MUTEX_EXIT(&conn->conn_data_lock);
1341 if (conn->natKeepAliveEvent) {
1342 rxi_NatKeepAliveOff(conn);
1345 if (conn->delayedAbortEvent) {
1346 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1347 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1349 MUTEX_ENTER(&conn->conn_data_lock);
1350 rxi_SendConnectionAbort(conn, packet, 0, 1);
1351 MUTEX_EXIT(&conn->conn_data_lock);
1352 rxi_FreePacket(packet);
1356 /* Remove from connection hash table before proceeding */
1358 &rx_connHashTable[CONN_HASH
1359 (peer->host, peer->port, conn->cid, conn->epoch,
1361 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1362 if (*conn_ptr == conn) {
1363 *conn_ptr = conn->next;
1367 /* if the conn that we are destroying was the last connection, then we
1368 * clear rxLastConn as well */
1369 if (rxLastConn == conn)
1372 /* Make sure the connection is completely reset before deleting it. */
1373 /* get rid of pending events that could zap us later */
1374 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1375 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1376 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1378 /* Add the connection to the list of destroyed connections that
1379 * need to be cleaned up. This is necessary to avoid deadlocks
1380 * in the routines we call to inform others that this connection is
1381 * being destroyed. */
1382 conn->next = rx_connCleanup_list;
1383 rx_connCleanup_list = conn;
1386 /* Externally available version */
1388 rx_DestroyConnection(struct rx_connection *conn)
1393 rxi_DestroyConnection(conn);
1398 rx_GetConnection(struct rx_connection *conn)
1403 MUTEX_ENTER(&rx_refcnt_mutex);
1405 MUTEX_EXIT(&rx_refcnt_mutex);
1409 #ifdef RX_ENABLE_LOCKS
1410 /* Wait for the transmit queue to no longer be busy.
1411 * requires the call->lock to be held */
1413 rxi_WaitforTQBusy(struct rx_call *call) {
1414 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1415 call->flags |= RX_CALL_TQ_WAIT;
1417 MUTEX_ASSERT(&call->lock);
1418 CV_WAIT(&call->cv_tq, &call->lock);
1420 if (call->tqWaiters == 0) {
1421 call->flags &= ~RX_CALL_TQ_WAIT;
1428 rxi_WakeUpTransmitQueue(struct rx_call *call)
1430 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1431 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1432 call, call->tqWaiters, call->flags));
1433 #ifdef RX_ENABLE_LOCKS
1434 MUTEX_ASSERT(&call->lock);
1435 CV_BROADCAST(&call->cv_tq);
1436 #else /* RX_ENABLE_LOCKS */
1437 osi_rxWakeup(&call->tq);
1438 #endif /* RX_ENABLE_LOCKS */
1442 /* Start a new rx remote procedure call, on the specified connection.
1443 * If wait is set to 1, wait for a free call channel; otherwise return
1444 * 0. Maxtime gives the maximum number of seconds this call may take,
1445 * after rx_NewCall returns. After this time interval, a call to any
1446 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1447 * For fine grain locking, we hold the conn_call_lock in order to
1448 * to ensure that we don't get signalle after we found a call in an active
1449 * state and before we go to sleep.
1452 rx_NewCall(struct rx_connection *conn)
1454 int i, wait, ignoreBusy = 1;
1455 struct rx_call *call;
1456 struct clock queueTime;
1457 afs_uint32 leastBusy = 0;
1461 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1464 clock_GetTime(&queueTime);
1466 * Check if there are others waiting for a new call.
1467 * If so, let them go first to avoid starving them.
1468 * This is a fairly simple scheme, and might not be
1469 * a complete solution for large numbers of waiters.
1471 * makeCallWaiters keeps track of the number of
1472 * threads waiting to make calls and the
1473 * RX_CONN_MAKECALL_WAITING flag bit is used to
1474 * indicate that there are indeed calls waiting.
1475 * The flag is set when the waiter is incremented.
1476 * It is only cleared when makeCallWaiters is 0.
1477 * This prevents us from accidently destroying the
1478 * connection while it is potentially about to be used.
1480 MUTEX_ENTER(&conn->conn_call_lock);
1481 MUTEX_ENTER(&conn->conn_data_lock);
1482 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1483 conn->flags |= RX_CONN_MAKECALL_WAITING;
1484 conn->makeCallWaiters++;
1485 MUTEX_EXIT(&conn->conn_data_lock);
1487 #ifdef RX_ENABLE_LOCKS
1488 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1492 MUTEX_ENTER(&conn->conn_data_lock);
1493 conn->makeCallWaiters--;
1494 if (conn->makeCallWaiters == 0)
1495 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1498 /* We are now the active thread in rx_NewCall */
1499 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1500 MUTEX_EXIT(&conn->conn_data_lock);
1505 for (i = 0; i < RX_MAXCALLS; i++) {
1506 call = conn->call[i];
1508 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1509 /* we're not ignoring busy call slots; only look at the
1510 * call slot that is the "least" busy */
1514 if (call->state == RX_STATE_DALLY) {
1515 MUTEX_ENTER(&call->lock);
1516 if (call->state == RX_STATE_DALLY) {
1517 if (ignoreBusy && conn->lastBusy[i]) {
1518 /* if we're ignoring busy call slots, skip any ones that
1519 * have lastBusy set */
1520 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1521 leastBusy = conn->lastBusy[i];
1523 MUTEX_EXIT(&call->lock);
1528 * We are setting the state to RX_STATE_RESET to
1529 * ensure that no one else will attempt to use this
1530 * call once we drop the conn->conn_call_lock and
1531 * call->lock. We must drop the conn->conn_call_lock
1532 * before calling rxi_ResetCall because the process
1533 * of clearing the transmit queue can block for an
1534 * extended period of time. If we block while holding
1535 * the conn->conn_call_lock, then all rx_EndCall
1536 * processing will block as well. This has a detrimental
1537 * effect on overall system performance.
1539 call->state = RX_STATE_RESET;
1540 (*call->callNumber)++;
1541 MUTEX_EXIT(&conn->conn_call_lock);
1542 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1543 rxi_ResetCall(call, 0);
1544 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1548 * If we failed to be able to safely obtain the
1549 * conn->conn_call_lock we will have to drop the
1550 * call->lock to avoid a deadlock. When the call->lock
1551 * is released the state of the call can change. If it
1552 * is no longer RX_STATE_RESET then some other thread is
1555 MUTEX_EXIT(&call->lock);
1556 MUTEX_ENTER(&conn->conn_call_lock);
1557 MUTEX_ENTER(&call->lock);
1559 if (call->state == RX_STATE_RESET)
1563 * If we get here it means that after dropping
1564 * the conn->conn_call_lock and call->lock that
1565 * the call is no longer ours. If we can't find
1566 * a free call in the remaining slots we should
1567 * not go immediately to RX_CONN_MAKECALL_WAITING
1568 * because by dropping the conn->conn_call_lock
1569 * we have given up synchronization with rx_EndCall.
1570 * Instead, cycle through one more time to see if
1571 * we can find a call that can call our own.
1573 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1576 MUTEX_EXIT(&call->lock);
1579 if (ignoreBusy && conn->lastBusy[i]) {
1580 /* if we're ignoring busy call slots, skip any ones that
1581 * have lastBusy set */
1582 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1583 leastBusy = conn->lastBusy[i];
1588 /* rxi_NewCall returns with mutex locked */
1589 call = rxi_NewCall(conn, i);
1590 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1594 if (i < RX_MAXCALLS) {
1595 conn->lastBusy[i] = 0;
1596 call->flags &= ~RX_CALL_PEER_BUSY;
1601 if (leastBusy && ignoreBusy) {
1602 /* we didn't find a useable call slot, but we did see at least one
1603 * 'busy' slot; look again and only use a slot with the 'least
1609 MUTEX_ENTER(&conn->conn_data_lock);
1610 conn->flags |= RX_CONN_MAKECALL_WAITING;
1611 conn->makeCallWaiters++;
1612 MUTEX_EXIT(&conn->conn_data_lock);
1614 #ifdef RX_ENABLE_LOCKS
1615 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1619 MUTEX_ENTER(&conn->conn_data_lock);
1620 conn->makeCallWaiters--;
1621 if (conn->makeCallWaiters == 0)
1622 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1623 MUTEX_EXIT(&conn->conn_data_lock);
1625 /* Client is initially in send mode */
1626 call->state = RX_STATE_ACTIVE;
1627 call->error = conn->error;
1629 call->app.mode = RX_MODE_ERROR;
1631 call->app.mode = RX_MODE_SENDING;
1633 #ifdef AFS_RXERRQ_ENV
1634 /* remember how many network errors the peer has when we started, so if
1635 * more errors are encountered after the call starts, we know the other endpoint won't be
1636 * responding to us */
1637 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1640 /* remember start time for call in case we have hard dead time limit */
1641 call->queueTime = queueTime;
1642 clock_GetTime(&call->startTime);
1643 call->app.bytesSent = 0;
1644 call->app.bytesRcvd = 0;
1646 /* Turn on busy protocol. */
1647 rxi_KeepAliveOn(call);
1649 /* Attempt MTU discovery */
1650 rxi_GrowMTUOn(call);
1653 * We are no longer the active thread in rx_NewCall
1655 MUTEX_ENTER(&conn->conn_data_lock);
1656 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1657 MUTEX_EXIT(&conn->conn_data_lock);
1660 * Wake up anyone else who might be giving us a chance to
1661 * run (see code above that avoids resource starvation).
1663 #ifdef RX_ENABLE_LOCKS
1664 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1665 osi_Panic("rx_NewCall call about to be used without an empty tq");
1668 CV_BROADCAST(&conn->conn_call_cv);
1672 MUTEX_EXIT(&conn->conn_call_lock);
1673 MUTEX_EXIT(&call->lock);
1676 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1681 rxi_HasActiveCalls(struct rx_connection *aconn)
1684 struct rx_call *tcall;
1688 for (i = 0; i < RX_MAXCALLS; i++) {
1689 if ((tcall = aconn->call[i])) {
1690 if ((tcall->state == RX_STATE_ACTIVE)
1691 || (tcall->state == RX_STATE_PRECALL)) {
1702 rxi_GetCallNumberVector(struct rx_connection *aconn,
1703 afs_int32 * aint32s)
1706 struct rx_call *tcall;
1710 MUTEX_ENTER(&aconn->conn_call_lock);
1711 for (i = 0; i < RX_MAXCALLS; i++) {
1712 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1713 aint32s[i] = aconn->callNumber[i] + 1;
1715 aint32s[i] = aconn->callNumber[i];
1717 MUTEX_EXIT(&aconn->conn_call_lock);
1723 rxi_SetCallNumberVector(struct rx_connection *aconn,
1724 afs_int32 * aint32s)
1727 struct rx_call *tcall;
1731 MUTEX_ENTER(&aconn->conn_call_lock);
1732 for (i = 0; i < RX_MAXCALLS; i++) {
1733 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1734 aconn->callNumber[i] = aint32s[i] - 1;
1736 aconn->callNumber[i] = aint32s[i];
1738 MUTEX_EXIT(&aconn->conn_call_lock);
1743 /* Advertise a new service. A service is named locally by a UDP port
1744 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1747 char *serviceName; Name for identification purposes (e.g. the
1748 service name might be used for probing for
1751 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1752 char *serviceName, struct rx_securityClass **securityObjects,
1753 int nSecurityObjects,
1754 afs_int32(*serviceProc) (struct rx_call * acall))
1756 osi_socket socket = OSI_NULLSOCKET;
1757 struct rx_service *tservice;
1763 if (serviceId == 0) {
1765 "rx_NewService: service id for service %s is not non-zero.\n",
1772 "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",
1780 tservice = rxi_AllocService();
1783 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1785 for (i = 0; i < RX_MAX_SERVICES; i++) {
1786 struct rx_service *service = rx_services[i];
1788 if (port == service->servicePort && host == service->serviceHost) {
1789 if (service->serviceId == serviceId) {
1790 /* The identical service has already been
1791 * installed; if the caller was intending to
1792 * change the security classes used by this
1793 * service, he/she loses. */
1795 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1796 serviceName, serviceId, service->serviceName);
1798 rxi_FreeService(tservice);
1801 /* Different service, same port: re-use the socket
1802 * which is bound to the same port */
1803 socket = service->socket;
1806 if (socket == OSI_NULLSOCKET) {
1807 /* If we don't already have a socket (from another
1808 * service on same port) get a new one */
1809 socket = rxi_GetHostUDPSocket(host, port);
1810 if (socket == OSI_NULLSOCKET) {
1812 rxi_FreeService(tservice);
1817 service->socket = socket;
1818 service->serviceHost = host;
1819 service->servicePort = port;
1820 service->serviceId = serviceId;
1821 service->serviceName = serviceName;
1822 service->nSecurityObjects = nSecurityObjects;
1823 service->securityObjects = securityObjects;
1824 service->minProcs = 0;
1825 service->maxProcs = 1;
1826 service->idleDeadTime = 60;
1827 service->idleDeadErr = 0;
1828 service->connDeadTime = rx_connDeadTime;
1829 service->executeRequestProc = serviceProc;
1830 service->checkReach = 0;
1831 service->nSpecific = 0;
1832 service->specific = NULL;
1833 rx_services[i] = service; /* not visible until now */
1839 rxi_FreeService(tservice);
1840 (osi_Msg "rx_NewService: cannot support > %d services\n",
1845 /* Set configuration options for all of a service's security objects */
1848 rx_SetSecurityConfiguration(struct rx_service *service,
1849 rx_securityConfigVariables type,
1853 for (i = 0; i<service->nSecurityObjects; i++) {
1854 if (service->securityObjects[i]) {
1855 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1863 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1864 struct rx_securityClass **securityObjects, int nSecurityObjects,
1865 afs_int32(*serviceProc) (struct rx_call * acall))
1867 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1870 /* Generic request processing loop. This routine should be called
1871 * by the implementation dependent rx_ServerProc. If socketp is
1872 * non-null, it will be set to the file descriptor that this thread
1873 * is now listening on. If socketp is null, this routine will never
1876 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1878 struct rx_call *call;
1880 struct rx_service *tservice = NULL;
1887 call = rx_GetCall(threadID, tservice, socketp);
1888 if (socketp && *socketp != OSI_NULLSOCKET) {
1889 /* We are now a listener thread */
1895 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1896 #ifdef RX_ENABLE_LOCKS
1898 #endif /* RX_ENABLE_LOCKS */
1899 afs_termState = AFSOP_STOP_AFS;
1900 afs_osi_Wakeup(&afs_termState);
1901 #ifdef RX_ENABLE_LOCKS
1903 #endif /* RX_ENABLE_LOCKS */
1908 /* if server is restarting( typically smooth shutdown) then do not
1909 * allow any new calls.
1912 if (rx_tranquil && (call != NULL)) {
1916 MUTEX_ENTER(&call->lock);
1918 rxi_CallError(call, RX_RESTARTING);
1919 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1921 MUTEX_EXIT(&call->lock);
1926 tservice = call->conn->service;
1928 if (tservice->beforeProc)
1929 (*tservice->beforeProc) (call);
1931 code = tservice->executeRequestProc(call);
1933 if (tservice->afterProc)
1934 (*tservice->afterProc) (call, code);
1936 rx_EndCall(call, code);
1938 if (tservice->postProc)
1939 (*tservice->postProc) (code);
1941 if (rx_stats_active) {
1942 MUTEX_ENTER(&rx_stats_mutex);
1944 MUTEX_EXIT(&rx_stats_mutex);
1951 rx_WakeupServerProcs(void)
1953 struct rx_serverQueueEntry *np, *tqp;
1954 struct opr_queue *cursor;
1958 MUTEX_ENTER(&rx_serverPool_lock);
1960 #ifdef RX_ENABLE_LOCKS
1961 if (rx_waitForPacket)
1962 CV_BROADCAST(&rx_waitForPacket->cv);
1963 #else /* RX_ENABLE_LOCKS */
1964 if (rx_waitForPacket)
1965 osi_rxWakeup(rx_waitForPacket);
1966 #endif /* RX_ENABLE_LOCKS */
1967 MUTEX_ENTER(&freeSQEList_lock);
1968 for (np = rx_FreeSQEList; np; np = tqp) {
1969 tqp = *(struct rx_serverQueueEntry **)np;
1970 #ifdef RX_ENABLE_LOCKS
1971 CV_BROADCAST(&np->cv);
1972 #else /* RX_ENABLE_LOCKS */
1974 #endif /* RX_ENABLE_LOCKS */
1976 MUTEX_EXIT(&freeSQEList_lock);
1977 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1978 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1979 #ifdef RX_ENABLE_LOCKS
1980 CV_BROADCAST(&np->cv);
1981 #else /* RX_ENABLE_LOCKS */
1983 #endif /* RX_ENABLE_LOCKS */
1985 MUTEX_EXIT(&rx_serverPool_lock);
1990 * One thing that seems to happen is that all the server threads get
1991 * tied up on some empty or slow call, and then a whole bunch of calls
1992 * arrive at once, using up the packet pool, so now there are more
1993 * empty calls. The most critical resources here are server threads
1994 * and the free packet pool. The "doreclaim" code seems to help in
1995 * general. I think that eventually we arrive in this state: there
1996 * are lots of pending calls which do have all their packets present,
1997 * so they won't be reclaimed, are multi-packet calls, so they won't
1998 * be scheduled until later, and thus are tying up most of the free
1999 * packet pool for a very long time.
2001 * 1. schedule multi-packet calls if all the packets are present.
2002 * Probably CPU-bound operation, useful to return packets to pool.
2003 * Do what if there is a full window, but the last packet isn't here?
2004 * 3. preserve one thread which *only* runs "best" calls, otherwise
2005 * it sleeps and waits for that type of call.
2006 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2007 * the current dataquota business is badly broken. The quota isn't adjusted
2008 * to reflect how many packets are presently queued for a running call.
2009 * So, when we schedule a queued call with a full window of packets queued
2010 * up for it, that *should* free up a window full of packets for other 2d-class
2011 * calls to be able to use from the packet pool. But it doesn't.
2013 * NB. Most of the time, this code doesn't run -- since idle server threads
2014 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2015 * as a new call arrives.
2017 /* Sleep until a call arrives. Returns a pointer to the call, ready
2018 * for an rx_Read. */
2019 #ifdef RX_ENABLE_LOCKS
2021 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2023 struct rx_serverQueueEntry *sq;
2024 struct rx_call *call = (struct rx_call *)0;
2025 struct rx_service *service = NULL;
2027 MUTEX_ENTER(&freeSQEList_lock);
2029 if ((sq = rx_FreeSQEList)) {
2030 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2031 MUTEX_EXIT(&freeSQEList_lock);
2032 } else { /* otherwise allocate a new one and return that */
2033 MUTEX_EXIT(&freeSQEList_lock);
2034 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2035 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2036 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2039 MUTEX_ENTER(&rx_serverPool_lock);
2040 if (cur_service != NULL) {
2041 ReturnToServerPool(cur_service);
2044 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2045 struct rx_call *tcall, *choice2 = NULL;
2046 struct opr_queue *cursor;
2048 /* Scan for eligible incoming calls. A call is not eligible
2049 * if the maximum number of calls for its service type are
2050 * already executing */
2051 /* One thread will process calls FCFS (to prevent starvation),
2052 * while the other threads may run ahead looking for calls which
2053 * have all their input data available immediately. This helps
2054 * keep threads from blocking, waiting for data from the client. */
2055 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2056 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2058 service = tcall->conn->service;
2059 if (!QuotaOK(service)) {
2062 MUTEX_ENTER(&rx_pthread_mutex);
2063 if (tno == rxi_fcfs_thread_num
2064 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2065 MUTEX_EXIT(&rx_pthread_mutex);
2066 /* If we're the fcfs thread , then we'll just use
2067 * this call. If we haven't been able to find an optimal
2068 * choice, and we're at the end of the list, then use a
2069 * 2d choice if one has been identified. Otherwise... */
2070 call = (choice2 ? choice2 : tcall);
2071 service = call->conn->service;
2073 MUTEX_EXIT(&rx_pthread_mutex);
2074 if (!opr_queue_IsEmpty(&tcall->rq)) {
2075 struct rx_packet *rp;
2076 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2078 if (rp->header.seq == 1) {
2080 || (rp->header.flags & RX_LAST_PACKET)) {
2082 } else if (rxi_2dchoice && !choice2
2083 && !(tcall->flags & RX_CALL_CLEARED)
2084 && (tcall->rprev > rxi_HardAckRate)) {
2094 ReturnToServerPool(service);
2100 opr_queue_Remove(&call->entry);
2101 MUTEX_EXIT(&rx_serverPool_lock);
2102 MUTEX_ENTER(&call->lock);
2104 if (call->flags & RX_CALL_WAIT_PROC) {
2105 call->flags &= ~RX_CALL_WAIT_PROC;
2106 rx_atomic_dec(&rx_nWaiting);
2109 if (call->state != RX_STATE_PRECALL || call->error) {
2110 MUTEX_EXIT(&call->lock);
2111 MUTEX_ENTER(&rx_serverPool_lock);
2112 ReturnToServerPool(service);
2117 if (opr_queue_IsEmpty(&call->rq)
2118 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2119 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2121 CLEAR_CALL_QUEUE_LOCK(call);
2124 /* If there are no eligible incoming calls, add this process
2125 * to the idle server queue, to wait for one */
2129 *socketp = OSI_NULLSOCKET;
2131 sq->socketp = socketp;
2132 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2133 #ifndef AFS_AIX41_ENV
2134 rx_waitForPacket = sq;
2136 rx_waitingForPacket = sq;
2137 #endif /* AFS_AIX41_ENV */
2139 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2141 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2142 MUTEX_EXIT(&rx_serverPool_lock);
2143 return (struct rx_call *)0;
2146 } while (!(call = sq->newcall)
2147 && !(socketp && *socketp != OSI_NULLSOCKET));
2148 MUTEX_EXIT(&rx_serverPool_lock);
2150 MUTEX_ENTER(&call->lock);
2156 MUTEX_ENTER(&freeSQEList_lock);
2157 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2158 rx_FreeSQEList = sq;
2159 MUTEX_EXIT(&freeSQEList_lock);
2162 clock_GetTime(&call->startTime);
2163 call->state = RX_STATE_ACTIVE;
2164 call->app.mode = RX_MODE_RECEIVING;
2165 #ifdef RX_KERNEL_TRACE
2166 if (ICL_SETACTIVE(afs_iclSetp)) {
2167 int glockOwner = ISAFS_GLOCK();
2170 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2171 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2178 rxi_calltrace(RX_CALL_START, call);
2179 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2180 call->conn->service->servicePort, call->conn->service->serviceId,
2183 MUTEX_EXIT(&call->lock);
2184 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2186 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2191 #else /* RX_ENABLE_LOCKS */
2193 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2195 struct rx_serverQueueEntry *sq;
2196 struct rx_call *call = (struct rx_call *)0, *choice2;
2197 struct rx_service *service = NULL;
2201 MUTEX_ENTER(&freeSQEList_lock);
2203 if ((sq = rx_FreeSQEList)) {
2204 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2205 MUTEX_EXIT(&freeSQEList_lock);
2206 } else { /* otherwise allocate a new one and return that */
2207 MUTEX_EXIT(&freeSQEList_lock);
2208 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2209 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2210 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2212 MUTEX_ENTER(&sq->lock);
2214 if (cur_service != NULL) {
2215 cur_service->nRequestsRunning--;
2216 MUTEX_ENTER(&rx_quota_mutex);
2217 if (cur_service->nRequestsRunning < cur_service->minProcs)
2220 MUTEX_EXIT(&rx_quota_mutex);
2222 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2223 struct rx_call *tcall;
2224 struct opr_queue *cursor;
2225 /* Scan for eligible incoming calls. A call is not eligible
2226 * if the maximum number of calls for its service type are
2227 * already executing */
2228 /* One thread will process calls FCFS (to prevent starvation),
2229 * while the other threads may run ahead looking for calls which
2230 * have all their input data available immediately. This helps
2231 * keep threads from blocking, waiting for data from the client. */
2232 choice2 = (struct rx_call *)0;
2233 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2234 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2235 service = tcall->conn->service;
2236 if (QuotaOK(service)) {
2237 MUTEX_ENTER(&rx_pthread_mutex);
2238 /* XXX - If tcall->entry.next is NULL, then we're no longer
2239 * on a queue at all. This shouldn't happen. */
2240 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2241 MUTEX_EXIT(&rx_pthread_mutex);
2242 /* If we're the fcfs thread, then we'll just use
2243 * this call. If we haven't been able to find an optimal
2244 * choice, and we're at the end of the list, then use a
2245 * 2d choice if one has been identified. Otherwise... */
2246 call = (choice2 ? choice2 : tcall);
2247 service = call->conn->service;
2249 MUTEX_EXIT(&rx_pthread_mutex);
2250 if (!opr_queue_IsEmpty(&tcall->rq)) {
2251 struct rx_packet *rp;
2252 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2254 if (rp->header.seq == 1
2256 || (rp->header.flags & RX_LAST_PACKET))) {
2258 } else if (rxi_2dchoice && !choice2
2259 && !(tcall->flags & RX_CALL_CLEARED)
2260 && (tcall->rprev > rxi_HardAckRate)) {
2273 opr_queue_Remove(&call->entry);
2274 /* we can't schedule a call if there's no data!!! */
2275 /* send an ack if there's no data, if we're missing the
2276 * first packet, or we're missing something between first
2277 * and last -- there's a "hole" in the incoming data. */
2278 if (opr_queue_IsEmpty(&call->rq)
2279 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2280 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2281 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2283 call->flags &= (~RX_CALL_WAIT_PROC);
2284 service->nRequestsRunning++;
2285 /* just started call in minProcs pool, need fewer to maintain
2287 MUTEX_ENTER(&rx_quota_mutex);
2288 if (service->nRequestsRunning <= service->minProcs)
2291 MUTEX_EXIT(&rx_quota_mutex);
2292 rx_atomic_dec(&rx_nWaiting);
2293 /* MUTEX_EXIT(&call->lock); */
2295 /* If there are no eligible incoming calls, add this process
2296 * to the idle server queue, to wait for one */
2299 *socketp = OSI_NULLSOCKET;
2301 sq->socketp = socketp;
2302 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2306 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2308 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2309 return (struct rx_call *)0;
2312 } while (!(call = sq->newcall)
2313 && !(socketp && *socketp != OSI_NULLSOCKET));
2315 MUTEX_EXIT(&sq->lock);
2317 MUTEX_ENTER(&freeSQEList_lock);
2318 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2319 rx_FreeSQEList = sq;
2320 MUTEX_EXIT(&freeSQEList_lock);
2323 clock_GetTime(&call->startTime);
2324 call->state = RX_STATE_ACTIVE;
2325 call->app.mode = RX_MODE_RECEIVING;
2326 #ifdef RX_KERNEL_TRACE
2327 if (ICL_SETACTIVE(afs_iclSetp)) {
2328 int glockOwner = ISAFS_GLOCK();
2331 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2332 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2339 rxi_calltrace(RX_CALL_START, call);
2340 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2341 call->conn->service->servicePort, call->conn->service->serviceId,
2344 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2351 #endif /* RX_ENABLE_LOCKS */
2355 /* Establish a procedure to be called when a packet arrives for a
2356 * call. This routine will be called at most once after each call,
2357 * and will also be called if there is an error condition on the or
2358 * the call is complete. Used by multi rx to build a selection
2359 * function which determines which of several calls is likely to be a
2360 * good one to read from.
2361 * NOTE: the way this is currently implemented it is probably only a
2362 * good idea to (1) use it immediately after a newcall (clients only)
2363 * and (2) only use it once. Other uses currently void your warranty
2366 rx_SetArrivalProc(struct rx_call *call,
2367 void (*proc) (struct rx_call * call,
2370 void * handle, int arg)
2372 call->arrivalProc = proc;
2373 call->arrivalProcHandle = handle;
2374 call->arrivalProcArg = arg;
2377 /* Call is finished (possibly prematurely). Return rc to the peer, if
2378 * appropriate, and return the final error code from the conversation
2382 rx_EndCall(struct rx_call *call, afs_int32 rc)
2384 struct rx_connection *conn = call->conn;
2388 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2389 call, rc, call->error, call->abortCode));
2392 MUTEX_ENTER(&call->lock);
2394 if (rc == 0 && call->error == 0) {
2395 call->abortCode = 0;
2396 call->abortCount = 0;
2399 call->arrivalProc = (void (*)())0;
2400 if (rc && call->error == 0) {
2401 rxi_CallError(call, rc);
2402 call->app.mode = RX_MODE_ERROR;
2403 /* Send an abort message to the peer if this error code has
2404 * only just been set. If it was set previously, assume the
2405 * peer has already been sent the error code or will request it
2407 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2409 if (conn->type == RX_SERVER_CONNECTION) {
2410 /* Make sure reply or at least dummy reply is sent */
2411 if (call->app.mode == RX_MODE_RECEIVING) {
2412 MUTEX_EXIT(&call->lock);
2413 rxi_WriteProc(call, 0, 0);
2414 MUTEX_ENTER(&call->lock);
2416 if (call->app.mode == RX_MODE_SENDING) {
2417 MUTEX_EXIT(&call->lock);
2418 rxi_FlushWrite(call);
2419 MUTEX_ENTER(&call->lock);
2421 rxi_calltrace(RX_CALL_END, call);
2422 /* Call goes to hold state until reply packets are acknowledged */
2423 if (call->tfirst + call->nSoftAcked < call->tnext) {
2424 call->state = RX_STATE_HOLD;
2426 call->state = RX_STATE_DALLY;
2427 rxi_ClearTransmitQueue(call, 0);
2428 rxi_rto_cancel(call);
2429 rxevent_Cancel(&call->keepAliveEvent, call,
2430 RX_CALL_REFCOUNT_ALIVE);
2432 } else { /* Client connection */
2434 /* Make sure server receives input packets, in the case where
2435 * no reply arguments are expected */
2437 if ((call->app.mode == RX_MODE_SENDING)
2438 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2439 MUTEX_EXIT(&call->lock);
2440 (void)rxi_ReadProc(call, &dummy, 1);
2441 MUTEX_ENTER(&call->lock);
2444 /* If we had an outstanding delayed ack, be nice to the server
2445 * and force-send it now.
2447 if (call->delayedAckEvent) {
2448 rxevent_Cancel(&call->delayedAckEvent, call,
2449 RX_CALL_REFCOUNT_DELAY);
2450 rxi_SendDelayedAck(NULL, call, NULL, 0);
2453 /* We need to release the call lock since it's lower than the
2454 * conn_call_lock and we don't want to hold the conn_call_lock
2455 * over the rx_ReadProc call. The conn_call_lock needs to be held
2456 * here for the case where rx_NewCall is perusing the calls on
2457 * the connection structure. We don't want to signal until
2458 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2459 * have checked this call, found it active and by the time it
2460 * goes to sleep, will have missed the signal.
2462 MUTEX_EXIT(&call->lock);
2463 MUTEX_ENTER(&conn->conn_call_lock);
2464 MUTEX_ENTER(&call->lock);
2466 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2467 conn->lastBusy[call->channel] = 0;
2470 MUTEX_ENTER(&conn->conn_data_lock);
2471 conn->flags |= RX_CONN_BUSY;
2472 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2473 MUTEX_EXIT(&conn->conn_data_lock);
2474 #ifdef RX_ENABLE_LOCKS
2475 CV_BROADCAST(&conn->conn_call_cv);
2480 #ifdef RX_ENABLE_LOCKS
2482 MUTEX_EXIT(&conn->conn_data_lock);
2484 #endif /* RX_ENABLE_LOCKS */
2485 call->state = RX_STATE_DALLY;
2487 error = call->error;
2489 /* currentPacket, nLeft, and NFree must be zeroed here, because
2490 * ResetCall cannot: ResetCall may be called at splnet(), in the
2491 * kernel version, and may interrupt the macros rx_Read or
2492 * rx_Write, which run at normal priority for efficiency. */
2493 if (call->app.currentPacket) {
2494 #ifdef RX_TRACK_PACKETS
2495 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2497 rxi_FreePacket(call->app.currentPacket);
2498 call->app.currentPacket = (struct rx_packet *)0;
2501 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2503 /* Free any packets from the last call to ReadvProc/WritevProc */
2504 #ifdef RXDEBUG_PACKET
2506 #endif /* RXDEBUG_PACKET */
2507 rxi_FreePackets(0, &call->app.iovq);
2508 MUTEX_EXIT(&call->lock);
2510 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2511 if (conn->type == RX_CLIENT_CONNECTION) {
2512 MUTEX_ENTER(&conn->conn_data_lock);
2513 conn->flags &= ~RX_CONN_BUSY;
2514 MUTEX_EXIT(&conn->conn_data_lock);
2515 MUTEX_EXIT(&conn->conn_call_lock);
2519 * Map errors to the local host's errno.h format.
2521 error = ntoh_syserr_conv(error);
2525 #if !defined(KERNEL)
2527 /* Call this routine when shutting down a server or client (especially
2528 * clients). This will allow Rx to gracefully garbage collect server
2529 * connections, and reduce the number of retries that a server might
2530 * make to a dead client.
2531 * This is not quite right, since some calls may still be ongoing and
2532 * we can't lock them to destroy them. */
2536 struct rx_connection **conn_ptr, **conn_end;
2540 if (rxinit_status == 1) {
2542 return; /* Already shutdown. */
2544 rxi_DeleteCachedConnections();
2545 if (rx_connHashTable) {
2546 MUTEX_ENTER(&rx_connHashTable_lock);
2547 for (conn_ptr = &rx_connHashTable[0], conn_end =
2548 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2550 struct rx_connection *conn, *next;
2551 for (conn = *conn_ptr; conn; conn = next) {
2553 if (conn->type == RX_CLIENT_CONNECTION) {
2554 MUTEX_ENTER(&rx_refcnt_mutex);
2556 MUTEX_EXIT(&rx_refcnt_mutex);
2557 #ifdef RX_ENABLE_LOCKS
2558 rxi_DestroyConnectionNoLock(conn);
2559 #else /* RX_ENABLE_LOCKS */
2560 rxi_DestroyConnection(conn);
2561 #endif /* RX_ENABLE_LOCKS */
2565 #ifdef RX_ENABLE_LOCKS
2566 while (rx_connCleanup_list) {
2567 struct rx_connection *conn;
2568 conn = rx_connCleanup_list;
2569 rx_connCleanup_list = rx_connCleanup_list->next;
2570 MUTEX_EXIT(&rx_connHashTable_lock);
2571 rxi_CleanupConnection(conn);
2572 MUTEX_ENTER(&rx_connHashTable_lock);
2574 MUTEX_EXIT(&rx_connHashTable_lock);
2575 #endif /* RX_ENABLE_LOCKS */
2580 afs_winsockCleanup();
2588 /* if we wakeup packet waiter too often, can get in loop with two
2589 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2591 rxi_PacketsUnWait(void)
2593 if (!rx_waitingForPackets) {
2597 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2598 return; /* still over quota */
2601 rx_waitingForPackets = 0;
2602 #ifdef RX_ENABLE_LOCKS
2603 CV_BROADCAST(&rx_waitingForPackets_cv);
2605 osi_rxWakeup(&rx_waitingForPackets);
2611 /* ------------------Internal interfaces------------------------- */
2613 /* Return this process's service structure for the
2614 * specified socket and service */
2615 static struct rx_service *
2616 rxi_FindService(osi_socket socket, u_short serviceId)
2618 struct rx_service **sp;
2619 for (sp = &rx_services[0]; *sp; sp++) {
2620 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2626 #ifdef RXDEBUG_PACKET
2627 #ifdef KDUMP_RX_LOCK
2628 static struct rx_call_rx_lock *rx_allCallsp = 0;
2630 static struct rx_call *rx_allCallsp = 0;
2632 #endif /* RXDEBUG_PACKET */
2634 /* Allocate a call structure, for the indicated channel of the
2635 * supplied connection. The mode and state of the call must be set by
2636 * the caller. Returns the call with mutex locked. */
2637 static struct rx_call *
2638 rxi_NewCall(struct rx_connection *conn, int channel)
2640 struct rx_call *call;
2641 #ifdef RX_ENABLE_LOCKS
2642 struct rx_call *cp; /* Call pointer temp */
2643 struct opr_queue *cursor;
2646 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2648 /* Grab an existing call structure, or allocate a new one.
2649 * Existing call structures are assumed to have been left reset by
2651 MUTEX_ENTER(&rx_freeCallQueue_lock);
2653 #ifdef RX_ENABLE_LOCKS
2655 * EXCEPT that the TQ might not yet be cleared out.
2656 * Skip over those with in-use TQs.
2659 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2660 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2661 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2667 #else /* RX_ENABLE_LOCKS */
2668 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2669 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2670 #endif /* RX_ENABLE_LOCKS */
2671 opr_queue_Remove(&call->entry);
2672 if (rx_stats_active)
2673 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2674 MUTEX_EXIT(&rx_freeCallQueue_lock);
2675 MUTEX_ENTER(&call->lock);
2676 CLEAR_CALL_QUEUE_LOCK(call);
2677 #ifdef RX_ENABLE_LOCKS
2678 /* Now, if TQ wasn't cleared earlier, do it now. */
2679 rxi_WaitforTQBusy(call);
2680 if (call->flags & RX_CALL_TQ_CLEARME) {
2681 rxi_ClearTransmitQueue(call, 1);
2682 /*queue_Init(&call->tq);*/
2684 #endif /* RX_ENABLE_LOCKS */
2685 /* Bind the call to its connection structure */
2687 rxi_ResetCall(call, 1);
2690 call = rxi_Alloc(sizeof(struct rx_call));
2691 #ifdef RXDEBUG_PACKET
2692 call->allNextp = rx_allCallsp;
2693 rx_allCallsp = call;
2695 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2696 #else /* RXDEBUG_PACKET */
2697 rx_atomic_inc(&rx_stats.nCallStructs);
2698 #endif /* RXDEBUG_PACKET */
2700 MUTEX_EXIT(&rx_freeCallQueue_lock);
2701 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2702 MUTEX_ENTER(&call->lock);
2703 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2704 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2705 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2707 /* Initialize once-only items */
2708 opr_queue_Init(&call->tq);
2709 opr_queue_Init(&call->rq);
2710 opr_queue_Init(&call->app.iovq);
2711 #ifdef RXDEBUG_PACKET
2712 call->rqc = call->tqc = call->iovqc = 0;
2713 #endif /* RXDEBUG_PACKET */
2714 /* Bind the call to its connection structure (prereq for reset) */
2716 rxi_ResetCall(call, 1);
2718 call->channel = channel;
2719 call->callNumber = &conn->callNumber[channel];
2720 call->rwind = conn->rwind[channel];
2721 call->twind = conn->twind[channel];
2722 /* Note that the next expected call number is retained (in
2723 * conn->callNumber[i]), even if we reallocate the call structure
2725 conn->call[channel] = call;
2726 /* if the channel's never been used (== 0), we should start at 1, otherwise
2727 * the call number is valid from the last time this channel was used */
2728 if (*call->callNumber == 0)
2729 *call->callNumber = 1;
2734 /* A call has been inactive long enough that so we can throw away
2735 * state, including the call structure, which is placed on the call
2738 * call->lock amd rx_refcnt_mutex are held upon entry.
2739 * haveCTLock is set when called from rxi_ReapConnections.
2741 * return 1 if the call is freed, 0 if not.
2744 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2746 int channel = call->channel;
2747 struct rx_connection *conn = call->conn;
2748 u_char state = call->state;
2751 * We are setting the state to RX_STATE_RESET to
2752 * ensure that no one else will attempt to use this
2753 * call once we drop the refcnt lock. We must drop
2754 * the refcnt lock before calling rxi_ResetCall
2755 * because it cannot be held across acquiring the
2756 * freepktQ lock. NewCall does the same.
2758 call->state = RX_STATE_RESET;
2759 MUTEX_EXIT(&rx_refcnt_mutex);
2760 rxi_ResetCall(call, 0);
2762 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2764 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2765 (*call->callNumber)++;
2767 if (call->conn->call[channel] == call)
2768 call->conn->call[channel] = 0;
2769 MUTEX_EXIT(&conn->conn_call_lock);
2772 * We couldn't obtain the conn_call_lock so we can't
2773 * disconnect the call from the connection. Set the
2774 * call state to dally so that the call can be reused.
2776 MUTEX_ENTER(&rx_refcnt_mutex);
2777 call->state = RX_STATE_DALLY;
2781 MUTEX_ENTER(&rx_freeCallQueue_lock);
2782 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2783 #ifdef RX_ENABLE_LOCKS
2784 /* A call may be free even though its transmit queue is still in use.
2785 * Since we search the call list from head to tail, put busy calls at
2786 * the head of the list, and idle calls at the tail.
2788 if (call->flags & RX_CALL_TQ_BUSY)
2789 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2791 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2792 #else /* RX_ENABLE_LOCKS */
2793 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2794 #endif /* RX_ENABLE_LOCKS */
2795 if (rx_stats_active)
2796 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2797 MUTEX_EXIT(&rx_freeCallQueue_lock);
2799 /* Destroy the connection if it was previously slated for
2800 * destruction, i.e. the Rx client code previously called
2801 * rx_DestroyConnection (client connections), or
2802 * rxi_ReapConnections called the same routine (server
2803 * connections). Only do this, however, if there are no
2804 * outstanding calls. Note that for fine grain locking, there appears
2805 * to be a deadlock in that rxi_FreeCall has a call locked and
2806 * DestroyConnectionNoLock locks each call in the conn. But note a
2807 * few lines up where we have removed this call from the conn.
2808 * If someone else destroys a connection, they either have no
2809 * call lock held or are going through this section of code.
2811 MUTEX_ENTER(&conn->conn_data_lock);
2812 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2813 MUTEX_ENTER(&rx_refcnt_mutex);
2815 MUTEX_EXIT(&rx_refcnt_mutex);
2816 MUTEX_EXIT(&conn->conn_data_lock);
2817 #ifdef RX_ENABLE_LOCKS
2819 rxi_DestroyConnectionNoLock(conn);
2821 rxi_DestroyConnection(conn);
2822 #else /* RX_ENABLE_LOCKS */
2823 rxi_DestroyConnection(conn);
2824 #endif /* RX_ENABLE_LOCKS */
2826 MUTEX_EXIT(&conn->conn_data_lock);
2828 MUTEX_ENTER(&rx_refcnt_mutex);
2832 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2833 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2836 rxi_Alloc(size_t size)
2840 if (rx_stats_active) {
2841 rx_atomic_add(&rxi_Allocsize, (int) size);
2842 rx_atomic_inc(&rxi_Alloccnt);
2846 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2847 afs_osi_Alloc_NoSleep(size);
2852 osi_Panic("rxi_Alloc error");
2858 rxi_Free(void *addr, size_t size)
2860 if (rx_stats_active) {
2861 rx_atomic_sub(&rxi_Allocsize, (int) size);
2862 rx_atomic_dec(&rxi_Alloccnt);
2864 osi_Free(addr, size);
2868 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2870 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2871 struct rx_peer *next = NULL;
2875 MUTEX_ENTER(&rx_peerHashTable_lock);
2877 peer_ptr = &rx_peerHashTable[0];
2878 peer_end = &rx_peerHashTable[rx_hashTableSize];
2881 for ( ; peer_ptr < peer_end; peer_ptr++) {
2884 for ( ; peer; peer = next) {
2886 if (host == peer->host)
2891 hashIndex = PEER_HASH(host, port);
2892 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2893 if ((peer->host == host) && (peer->port == port))
2898 MUTEX_ENTER(&rx_peerHashTable_lock);
2903 MUTEX_EXIT(&rx_peerHashTable_lock);
2905 MUTEX_ENTER(&peer->peer_lock);
2906 /* We don't handle dropping below min, so don't */
2907 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2908 peer->ifMTU=MIN(mtu, peer->ifMTU);
2909 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2910 /* if we tweaked this down, need to tune our peer MTU too */
2911 peer->MTU = MIN(peer->MTU, peer->natMTU);
2912 /* if we discovered a sub-1500 mtu, degrade */
2913 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2914 peer->maxDgramPackets = 1;
2915 /* We no longer have valid peer packet information */
2916 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2917 peer->maxPacketSize = 0;
2918 MUTEX_EXIT(&peer->peer_lock);
2920 MUTEX_ENTER(&rx_peerHashTable_lock);
2922 if (host && !port) {
2924 /* pick up where we left off */
2928 MUTEX_EXIT(&rx_peerHashTable_lock);
2931 #ifdef AFS_RXERRQ_ENV
2933 rxi_SetPeerDead(afs_uint32 host, afs_uint16 port)
2935 int hashIndex = PEER_HASH(host, port);
2936 struct rx_peer *peer;
2938 MUTEX_ENTER(&rx_peerHashTable_lock);
2940 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2941 if (peer->host == host && peer->port == port) {
2947 rx_atomic_inc(&peer->neterrs);
2950 MUTEX_EXIT(&rx_peerHashTable_lock);
2954 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2956 # ifdef AFS_ADAPT_PMTU
2957 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2958 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2962 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2963 switch (err->ee_code) {
2964 case ICMP_NET_UNREACH:
2965 case ICMP_HOST_UNREACH:
2966 case ICMP_PORT_UNREACH:
2969 rxi_SetPeerDead(addr, port);
2974 #endif /* AFS_RXERRQ_ENV */
2976 /* Find the peer process represented by the supplied (host,port)
2977 * combination. If there is no appropriate active peer structure, a
2978 * new one will be allocated and initialized
2981 rxi_FindPeer(afs_uint32 host, u_short port, int create)
2985 hashIndex = PEER_HASH(host, port);
2986 MUTEX_ENTER(&rx_peerHashTable_lock);
2987 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2988 if ((pp->host == host) && (pp->port == port))
2993 pp = rxi_AllocPeer(); /* This bzero's *pp */
2994 pp->host = host; /* set here or in InitPeerParams is zero */
2996 #ifdef AFS_RXERRQ_ENV
2997 rx_atomic_set(&pp->neterrs, 0);
2999 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3000 opr_queue_Init(&pp->rpcStats);
3001 pp->next = rx_peerHashTable[hashIndex];
3002 rx_peerHashTable[hashIndex] = pp;
3003 rxi_InitPeerParams(pp);
3004 if (rx_stats_active)
3005 rx_atomic_inc(&rx_stats.nPeerStructs);
3011 MUTEX_EXIT(&rx_peerHashTable_lock);
3016 /* Find the connection at (host, port) started at epoch, and with the
3017 * given connection id. Creates the server connection if necessary.
3018 * The type specifies whether a client connection or a server
3019 * connection is desired. In both cases, (host, port) specify the
3020 * peer's (host, pair) pair. Client connections are not made
3021 * automatically by this routine. The parameter socket gives the
3022 * socket descriptor on which the packet was received. This is used,
3023 * in the case of server connections, to check that *new* connections
3024 * come via a valid (port, serviceId). Finally, the securityIndex
3025 * parameter must match the existing index for the connection. If a
3026 * server connection is created, it will be created using the supplied
3027 * index, if the index is valid for this service */
3028 static struct rx_connection *
3029 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3030 u_short port, u_short serviceId, afs_uint32 cid,
3031 afs_uint32 epoch, int type, u_int securityIndex)
3033 int hashindex, flag, i;
3034 struct rx_connection *conn;
3035 hashindex = CONN_HASH(host, port, cid, epoch, type);
3036 MUTEX_ENTER(&rx_connHashTable_lock);
3037 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3038 rx_connHashTable[hashindex],
3041 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3042 && (epoch == conn->epoch)) {
3043 struct rx_peer *pp = conn->peer;
3044 if (securityIndex != conn->securityIndex) {
3045 /* this isn't supposed to happen, but someone could forge a packet
3046 * like this, and there seems to be some CM bug that makes this
3047 * happen from time to time -- in which case, the fileserver
3049 MUTEX_EXIT(&rx_connHashTable_lock);
3050 return (struct rx_connection *)0;
3052 if (pp->host == host && pp->port == port)
3054 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3056 /* So what happens when it's a callback connection? */
3057 if ( /*type == RX_CLIENT_CONNECTION && */
3058 (conn->epoch & 0x80000000))
3062 /* the connection rxLastConn that was used the last time is not the
3063 ** one we are looking for now. Hence, start searching in the hash */
3065 conn = rx_connHashTable[hashindex];
3070 struct rx_service *service;
3071 if (type == RX_CLIENT_CONNECTION) {
3072 MUTEX_EXIT(&rx_connHashTable_lock);
3073 return (struct rx_connection *)0;
3075 service = rxi_FindService(socket, serviceId);
3076 if (!service || (securityIndex >= service->nSecurityObjects)
3077 || (service->securityObjects[securityIndex] == 0)) {
3078 MUTEX_EXIT(&rx_connHashTable_lock);
3079 return (struct rx_connection *)0;
3081 conn = rxi_AllocConnection(); /* This bzero's the connection */
3082 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3083 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3084 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3085 conn->next = rx_connHashTable[hashindex];
3086 rx_connHashTable[hashindex] = conn;
3087 conn->peer = rxi_FindPeer(host, port, 1);
3088 conn->type = RX_SERVER_CONNECTION;
3089 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3090 conn->epoch = epoch;
3091 conn->cid = cid & RX_CIDMASK;
3092 conn->ackRate = RX_FAST_ACK_RATE;
3093 conn->service = service;
3094 conn->serviceId = serviceId;
3095 conn->securityIndex = securityIndex;
3096 conn->securityObject = service->securityObjects[securityIndex];
3097 conn->nSpecific = 0;
3098 conn->specific = NULL;
3099 rx_SetConnDeadTime(conn, service->connDeadTime);
3100 conn->idleDeadTime = service->idleDeadTime;
3101 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3102 for (i = 0; i < RX_MAXCALLS; i++) {
3103 conn->twind[i] = rx_initSendWindow;
3104 conn->rwind[i] = rx_initReceiveWindow;
3106 /* Notify security object of the new connection */
3107 RXS_NewConnection(conn->securityObject, conn);
3108 /* XXXX Connection timeout? */
3109 if (service->newConnProc)
3110 (*service->newConnProc) (conn);
3111 if (rx_stats_active)
3112 rx_atomic_inc(&rx_stats.nServerConns);
3115 MUTEX_ENTER(&rx_refcnt_mutex);
3117 MUTEX_EXIT(&rx_refcnt_mutex);
3119 rxLastConn = conn; /* store this connection as the last conn used */
3120 MUTEX_EXIT(&rx_connHashTable_lock);
3125 * Timeout a call on a busy call channel if appropriate.
3127 * @param[in] call The busy call.
3129 * @pre 'call' is marked as busy (namely,
3130 * call->conn->lastBusy[call->channel] != 0)
3132 * @pre call->lock is held
3133 * @pre rxi_busyChannelError is nonzero
3135 * @note call->lock is dropped and reacquired
3138 rxi_CheckBusy(struct rx_call *call)
3140 struct rx_connection *conn = call->conn;
3141 int channel = call->channel;
3142 int freechannel = 0;
3145 MUTEX_EXIT(&call->lock);
3147 MUTEX_ENTER(&conn->conn_call_lock);
3149 /* Are there any other call slots on this conn that we should try? Look for
3150 * slots that are empty and are either non-busy, or were marked as busy
3151 * longer than conn->secondsUntilDead seconds before this call started. */
3153 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3155 /* only look at channels that aren't us */
3159 if (conn->lastBusy[i]) {
3160 /* if this channel looked busy too recently, don't look at it */
3161 if (conn->lastBusy[i] >= call->startTime.sec) {
3164 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3169 if (conn->call[i]) {
3170 struct rx_call *tcall = conn->call[i];
3171 MUTEX_ENTER(&tcall->lock);
3172 if (tcall->state == RX_STATE_DALLY) {
3175 MUTEX_EXIT(&tcall->lock);
3181 MUTEX_ENTER(&call->lock);
3183 /* Since the call->lock has been released it is possible that the call may
3184 * no longer be busy (the call channel cannot have been reallocated as we
3185 * haven't dropped the conn_call_lock) Therefore, we must confirm
3186 * that the call state has not changed when deciding whether or not to
3187 * force this application thread to retry by forcing a Timeout error. */
3189 if (freechannel && (call->flags & RX_CALL_PEER_BUSY)) {
3190 /* Since 'freechannel' is set, there exists another channel in this
3191 * rx_conn that the application thread might be able to use. We know
3192 * that we have the correct call since callNumber is unchanged, and we
3193 * know that the call is still busy. So, set the call error state to
3194 * rxi_busyChannelError so the application can retry the request,
3195 * presumably on a less-busy call channel. */
3197 rxi_CallError(call, RX_CALL_BUSY);
3199 MUTEX_EXIT(&conn->conn_call_lock);
3203 * Abort the call if the server is over the busy threshold. This
3204 * can be used without requiring a call structure be initialised,
3205 * or connected to a particular channel
3208 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3209 struct rx_packet *np)
3211 if ((rx_BusyThreshold > 0) &&
3212 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3213 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3214 rx_BusyError, np, 0);
3215 if (rx_stats_active)
3216 rx_atomic_inc(&rx_stats.nBusies);
3223 static_inline struct rx_call *
3224 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3227 struct rx_call *call;
3229 channel = np->header.cid & RX_CHANNELMASK;
3230 MUTEX_ENTER(&conn->conn_call_lock);
3231 call = conn->call[channel];
3232 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3233 MUTEX_EXIT(&conn->conn_call_lock);
3234 if (rx_stats_active)
3235 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3239 MUTEX_ENTER(&call->lock);
3240 MUTEX_EXIT(&conn->conn_call_lock);
3242 if ((call->state == RX_STATE_DALLY)
3243 && np->header.type == RX_PACKET_TYPE_ACK) {
3244 if (rx_stats_active)
3245 rx_atomic_inc(&rx_stats.ignorePacketDally);
3246 MUTEX_EXIT(&call->lock);
3253 static_inline struct rx_call *
3254 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3255 struct rx_connection *conn)
3258 struct rx_call *call;
3260 channel = np->header.cid & RX_CHANNELMASK;
3261 MUTEX_ENTER(&conn->conn_call_lock);
3262 call = conn->call[channel];
3265 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3266 MUTEX_EXIT(&conn->conn_call_lock);
3270 call = rxi_NewCall(conn, channel); /* returns locked call */
3271 *call->callNumber = np->header.callNumber;
3272 MUTEX_EXIT(&conn->conn_call_lock);
3274 call->state = RX_STATE_PRECALL;
3275 clock_GetTime(&call->queueTime);
3276 call->app.bytesSent = 0;
3277 call->app.bytesRcvd = 0;
3278 rxi_KeepAliveOn(call);
3283 if (np->header.callNumber == conn->callNumber[channel]) {
3284 MUTEX_ENTER(&call->lock);
3285 MUTEX_EXIT(&conn->conn_call_lock);
3289 if (np->header.callNumber < conn->callNumber[channel]) {
3290 MUTEX_EXIT(&conn->conn_call_lock);
3291 if (rx_stats_active)
3292 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3296 MUTEX_ENTER(&call->lock);
3297 MUTEX_EXIT(&conn->conn_call_lock);
3299 /* Wait until the transmit queue is idle before deciding
3300 * whether to reset the current call. Chances are that the
3301 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3304 #ifdef RX_ENABLE_LOCKS
3305 if (call->state == RX_STATE_ACTIVE) {
3306 int old_error = call->error;
3307 rxi_WaitforTQBusy(call);
3308 /* If we entered error state while waiting,
3309 * must call rxi_CallError to permit rxi_ResetCall
3310 * to processed when the tqWaiter count hits zero.
3312 if (call->error && call->error != old_error) {
3313 rxi_CallError(call, call->error);
3314 MUTEX_EXIT(&call->lock);
3318 #endif /* RX_ENABLE_LOCKS */
3319 /* If the new call cannot be taken right now send a busy and set
3320 * the error condition in this call, so that it terminates as
3321 * quickly as possible */
3322 if (call->state == RX_STATE_ACTIVE) {
3323 rxi_CallError(call, RX_CALL_DEAD);
3324 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3326 MUTEX_EXIT(&call->lock);
3330 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3331 MUTEX_EXIT(&call->lock);
3335 rxi_ResetCall(call, 0);
3336 /* The conn_call_lock is not held but no one else should be
3337 * using this call channel while we are processing this incoming
3338 * packet. This assignment should be safe.
3340 *call->callNumber = np->header.callNumber;
3341 call->state = RX_STATE_PRECALL;
3342 clock_GetTime(&call->queueTime);
3343 call->app.bytesSent = 0;
3344 call->app.bytesRcvd = 0;
3345 rxi_KeepAliveOn(call);
3351 /* There are two packet tracing routines available for testing and monitoring
3352 * Rx. One is called just after every packet is received and the other is
3353 * called just before every packet is sent. Received packets, have had their
3354 * headers decoded, and packets to be sent have not yet had their headers
3355 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3356 * containing the network address. Both can be modified. The return value, if
3357 * non-zero, indicates that the packet should be dropped. */
3359 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3360 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3362 /* A packet has been received off the interface. Np is the packet, socket is
3363 * the socket number it was received from (useful in determining which service
3364 * this packet corresponds to), and (host, port) reflect the host,port of the
3365 * sender. This call returns the packet to the caller if it is finished with
3366 * it, rather than de-allocating it, just as a small performance hack */
3369 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3370 afs_uint32 host, u_short port, int *tnop,
3371 struct rx_call **newcallp)
3373 struct rx_call *call;
3374 struct rx_connection *conn;
3379 struct rx_packet *tnp;
3382 /* We don't print out the packet until now because (1) the time may not be
3383 * accurate enough until now in the lwp implementation (rx_Listener only gets
3384 * the time after the packet is read) and (2) from a protocol point of view,
3385 * this is the first time the packet has been seen */
3386 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3387 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3388 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3389 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3390 np->header.epoch, np->header.cid, np->header.callNumber,
3391 np->header.seq, np->header.flags, np));
3394 /* Account for connectionless packets */
3395 if (rx_stats_active &&
3396 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3397 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3398 struct rx_peer *peer;
3400 /* Try to look up the peer structure, but don't create one */
3401 peer = rxi_FindPeer(host, port, 0);
3403 /* Since this may not be associated with a connection, it may have
3404 * no refCount, meaning we could race with ReapConnections
3407 if (peer && (peer->refCount > 0)) {
3408 #ifdef AFS_RXERRQ_ENV
3409 if (rx_atomic_read(&peer->neterrs)) {
3410 rx_atomic_set(&peer->neterrs, 0);
3413 MUTEX_ENTER(&peer->peer_lock);
3414 peer->bytesReceived += np->length;
3415 MUTEX_EXIT(&peer->peer_lock);
3419 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3420 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3423 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3424 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3427 /* If an input tracer function is defined, call it with the packet and
3428 * network address. Note this function may modify its arguments. */
3429 if (rx_justReceived) {
3430 struct sockaddr_in addr;
3432 addr.sin_family = AF_INET;
3433 addr.sin_port = port;
3434 addr.sin_addr.s_addr = host;
3435 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3436 addr.sin_len = sizeof(addr);
3437 #endif /* AFS_OSF_ENV */
3438 drop = (*rx_justReceived) (np, &addr);
3439 /* drop packet if return value is non-zero */
3442 port = addr.sin_port; /* in case fcn changed addr */
3443 host = addr.sin_addr.s_addr;
3447 /* If packet was not sent by the client, then *we* must be the client */
3448 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3449 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3451 /* Find the connection (or fabricate one, if we're the server & if
3452 * necessary) associated with this packet */
3454 rxi_FindConnection(socket, host, port, np->header.serviceId,
3455 np->header.cid, np->header.epoch, type,
3456 np->header.securityIndex);
3458 /* To avoid having 2 connections just abort at each other,
3459 don't abort an abort. */
3461 if (np->header.type != RX_PACKET_TYPE_ABORT)
3462 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3467 #ifdef AFS_RXERRQ_ENV
3468 if (rx_atomic_read(&conn->peer->neterrs)) {
3469 rx_atomic_set(&conn->peer->neterrs, 0);
3473 /* If we're doing statistics, then account for the incoming packet */
3474 if (rx_stats_active) {
3475 MUTEX_ENTER(&conn->peer->peer_lock);
3476 conn->peer->bytesReceived += np->length;
3477 MUTEX_EXIT(&conn->peer->peer_lock);
3480 /* If the connection is in an error state, send an abort packet and ignore
3481 * the incoming packet */
3483 /* Don't respond to an abort packet--we don't want loops! */
3484 MUTEX_ENTER(&conn->conn_data_lock);
3485 if (np->header.type != RX_PACKET_TYPE_ABORT)
3486 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3487 putConnection(conn);
3488 MUTEX_EXIT(&conn->conn_data_lock);
3492 /* Check for connection-only requests (i.e. not call specific). */
3493 if (np->header.callNumber == 0) {
3494 switch (np->header.type) {
3495 case RX_PACKET_TYPE_ABORT: {
3496 /* What if the supplied error is zero? */
3497 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3498 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3499 rxi_ConnectionError(conn, errcode);
3500 putConnection(conn);
3503 case RX_PACKET_TYPE_CHALLENGE:
3504 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3505 putConnection(conn);
3507 case RX_PACKET_TYPE_RESPONSE:
3508 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3509 putConnection(conn);
3511 case RX_PACKET_TYPE_PARAMS:
3512 case RX_PACKET_TYPE_PARAMS + 1:
3513 case RX_PACKET_TYPE_PARAMS + 2:
3514 /* ignore these packet types for now */
3515 putConnection(conn);
3519 /* Should not reach here, unless the peer is broken: send an
3521 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3522 MUTEX_ENTER(&conn->conn_data_lock);
3523 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3524 putConnection(conn);
3525 MUTEX_EXIT(&conn->conn_data_lock);
3530 if (type == RX_SERVER_CONNECTION)
3531 call = rxi_ReceiveServerCall(socket, np, conn);
3533 call = rxi_ReceiveClientCall(np, conn);
3536 putConnection(conn);
3540 MUTEX_ASSERT(&call->lock);
3541 /* Set remote user defined status from packet */
3542 call->remoteStatus = np->header.userStatus;
3544 /* Now do packet type-specific processing */
3545 switch (np->header.type) {
3546 case RX_PACKET_TYPE_DATA:
3547 /* If we're a client, and receiving a response, then all the packets
3548 * we transmitted packets are implicitly acknowledged. */
3549 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3550 rxi_AckAllInTransmitQueue(call);
3552 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3555 case RX_PACKET_TYPE_ACK:
3556 /* Respond immediately to ack packets requesting acknowledgement
3558 if (np->header.flags & RX_REQUEST_ACK) {
3560 (void)rxi_SendCallAbort(call, 0, 1, 0);
3562 (void)rxi_SendAck(call, 0, np->header.serial,
3563 RX_ACK_PING_RESPONSE, 1);
3565 np = rxi_ReceiveAckPacket(call, np, 1);
3567 case RX_PACKET_TYPE_ABORT: {
3568 /* An abort packet: reset the call, passing the error up to the user. */
3569 /* What if error is zero? */
3570 /* What if the error is -1? the application will treat it as a timeout. */
3571 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3572 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3573 rxi_CallError(call, errdata);
3574 MUTEX_EXIT(&call->lock);
3575 putConnection(conn);
3576 return np; /* xmitting; drop packet */
3578 case RX_PACKET_TYPE_BUSY: {
3579 struct clock busyTime;
3581 clock_GetTime(&busyTime);
3583 MUTEX_EXIT(&call->lock);
3585 MUTEX_ENTER(&conn->conn_call_lock);
3586 MUTEX_ENTER(&call->lock);
3587 conn->lastBusy[call->channel] = busyTime.sec;
3588 call->flags |= RX_CALL_PEER_BUSY;
3589 MUTEX_EXIT(&call->lock);
3590 MUTEX_EXIT(&conn->conn_call_lock);
3592 putConnection(conn);
3596 case RX_PACKET_TYPE_ACKALL:
3597 /* All packets acknowledged, so we can drop all packets previously
3598 * readied for sending */
3599 rxi_AckAllInTransmitQueue(call);
3602 /* Should not reach here, unless the peer is broken: send an abort
3604 rxi_CallError(call, RX_PROTOCOL_ERROR);
3605 np = rxi_SendCallAbort(call, np, 1, 0);
3608 /* Note when this last legitimate packet was received, for keep-alive
3609 * processing. Note, we delay getting the time until now in the hope that
3610 * the packet will be delivered to the user before any get time is required
3611 * (if not, then the time won't actually be re-evaluated here). */
3612 call->lastReceiveTime = clock_Sec();
3613 /* we've received a legit packet, so the channel is not busy */
3614 call->flags &= ~RX_CALL_PEER_BUSY;
3615 MUTEX_EXIT(&call->lock);
3616 putConnection(conn);
3620 /* return true if this is an "interesting" connection from the point of view
3621 of someone trying to debug the system */
3623 rxi_IsConnInteresting(struct rx_connection *aconn)
3626 struct rx_call *tcall;
3628 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3631 for (i = 0; i < RX_MAXCALLS; i++) {
3632 tcall = aconn->call[i];
3634 if ((tcall->state == RX_STATE_PRECALL)
3635 || (tcall->state == RX_STATE_ACTIVE))
3637 if ((tcall->app.mode == RX_MODE_SENDING)
3638 || (tcall->app.mode == RX_MODE_RECEIVING))
3646 /* if this is one of the last few packets AND it wouldn't be used by the
3647 receiving call to immediately satisfy a read request, then drop it on
3648 the floor, since accepting it might prevent a lock-holding thread from
3649 making progress in its reading. If a call has been cleared while in
3650 the precall state then ignore all subsequent packets until the call
3651 is assigned to a thread. */
3654 TooLow(struct rx_packet *ap, struct rx_call *acall)
3658 MUTEX_ENTER(&rx_quota_mutex);
3659 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3660 && (acall->state == RX_STATE_PRECALL))
3661 || ((rx_nFreePackets < rxi_dataQuota + 2)
3662 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3663 && (acall->flags & RX_CALL_READER_WAIT)))) {
3666 MUTEX_EXIT(&rx_quota_mutex);
3672 * Clear the attach wait flag on a connection and proceed.
3674 * Any processing waiting for a connection to be attached should be
3675 * unblocked. We clear the flag and do any other needed tasks.
3678 * the conn to unmark waiting for attach
3680 * @pre conn's conn_data_lock must be locked before calling this function
3684 rxi_ConnClearAttachWait(struct rx_connection *conn)
3686 /* Indicate that rxi_CheckReachEvent is no longer running by
3687 * clearing the flag. Must be atomic under conn_data_lock to
3688 * avoid a new call slipping by: rxi_CheckConnReach holds
3689 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3691 conn->flags &= ~RX_CONN_ATTACHWAIT;
3692 if (conn->flags & RX_CONN_NAT_PING) {
3693 conn->flags &= ~RX_CONN_NAT_PING;
3694 rxi_ScheduleNatKeepAliveEvent(conn);
3699 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3701 struct rx_connection *conn = arg1;
3702 struct rx_call *acall = arg2;
3703 struct rx_call *call = acall;
3704 struct clock when, now;
3707 MUTEX_ENTER(&conn->conn_data_lock);
3710 rxevent_Put(conn->checkReachEvent);
3711 conn->checkReachEvent = NULL;
3714 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3716 putConnection(conn);
3718 MUTEX_EXIT(&conn->conn_data_lock);
3722 MUTEX_ENTER(&conn->conn_call_lock);
3723 MUTEX_ENTER(&conn->conn_data_lock);
3724 for (i = 0; i < RX_MAXCALLS; i++) {
3725 struct rx_call *tc = conn->call[i];
3726 if (tc && tc->state == RX_STATE_PRECALL) {
3732 rxi_ConnClearAttachWait(conn);
3733 MUTEX_EXIT(&conn->conn_data_lock);
3734 MUTEX_EXIT(&conn->conn_call_lock);
3739 MUTEX_ENTER(&call->lock);
3740 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3742 MUTEX_EXIT(&call->lock);
3744 clock_GetTime(&now);
3746 when.sec += RX_CHECKREACH_TIMEOUT;
3747 MUTEX_ENTER(&conn->conn_data_lock);
3748 if (!conn->checkReachEvent) {
3749 MUTEX_ENTER(&rx_refcnt_mutex);
3751 MUTEX_EXIT(&rx_refcnt_mutex);
3752 conn->checkReachEvent = rxevent_Post(&when, &now,
3753 rxi_CheckReachEvent, conn,
3756 MUTEX_EXIT(&conn->conn_data_lock);
3762 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3764 struct rx_service *service = conn->service;
3765 struct rx_peer *peer = conn->peer;
3766 afs_uint32 now, lastReach;
3768 if (service->checkReach == 0)
3772 MUTEX_ENTER(&peer->peer_lock);
3773 lastReach = peer->lastReachTime;
3774 MUTEX_EXIT(&peer->peer_lock);
3775 if (now - lastReach < RX_CHECKREACH_TTL)
3778 MUTEX_ENTER(&conn->conn_data_lock);
3779 if (conn->flags & RX_CONN_ATTACHWAIT) {
3780 MUTEX_EXIT(&conn->conn_data_lock);
3783 conn->flags |= RX_CONN_ATTACHWAIT;
3784 MUTEX_EXIT(&conn->conn_data_lock);
3785 if (!conn->checkReachEvent)
3786 rxi_CheckReachEvent(NULL, conn, call, 0);
3791 /* try to attach call, if authentication is complete */
3793 TryAttach(struct rx_call *acall, osi_socket socket,
3794 int *tnop, struct rx_call **newcallp,
3797 struct rx_connection *conn = acall->conn;
3799 if (conn->type == RX_SERVER_CONNECTION
3800 && acall->state == RX_STATE_PRECALL) {
3801 /* Don't attach until we have any req'd. authentication. */
3802 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3803 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3804 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3805 /* Note: this does not necessarily succeed; there
3806 * may not any proc available
3809 rxi_ChallengeOn(acall->conn);
3814 /* A data packet has been received off the interface. This packet is
3815 * appropriate to the call (the call is in the right state, etc.). This
3816 * routine can return a packet to the caller, for re-use */
3818 static struct rx_packet *
3819 rxi_ReceiveDataPacket(struct rx_call *call,
3820 struct rx_packet *np, int istack,
3821 osi_socket socket, afs_uint32 host, u_short port,
3822 int *tnop, struct rx_call **newcallp)
3824 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3829 afs_uint32 serial=0, flags=0;
3831 struct rx_packet *tnp;
3832 if (rx_stats_active)
3833 rx_atomic_inc(&rx_stats.dataPacketsRead);
3836 /* If there are no packet buffers, drop this new packet, unless we can find
3837 * packet buffers from inactive calls */
3839 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3840 MUTEX_ENTER(&rx_freePktQ_lock);
3841 rxi_NeedMorePackets = TRUE;
3842 MUTEX_EXIT(&rx_freePktQ_lock);
3843 if (rx_stats_active)
3844 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3845 rxi_calltrace(RX_TRACE_DROP, call);
3846 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3847 /* We used to clear the receive queue here, in an attempt to free
3848 * packets. However this is unsafe if the queue has received a
3849 * soft ACK for the final packet */
3850 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3856 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3857 * packet is one of several packets transmitted as a single
3858 * datagram. Do not send any soft or hard acks until all packets
3859 * in a jumbogram have been processed. Send negative acks right away.
3861 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3862 /* tnp is non-null when there are more packets in the
3863 * current jumbo gram */
3870 seq = np->header.seq;
3871 serial = np->header.serial;
3872 flags = np->header.flags;
3874 /* If the call is in an error state, send an abort message */
3876 return rxi_SendCallAbort(call, np, istack, 0);
3878 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3879 * AFS 3.5 jumbogram. */
3880 if (flags & RX_JUMBO_PACKET) {
3881 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3886 if (np->header.spare != 0) {
3887 MUTEX_ENTER(&call->conn->conn_data_lock);
3888 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3889 MUTEX_EXIT(&call->conn->conn_data_lock);
3892 /* The usual case is that this is the expected next packet */
3893 if (seq == call->rnext) {
3895 /* Check to make sure it is not a duplicate of one already queued */
3896 if (!opr_queue_IsEmpty(&call->rq)
3897 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3898 if (rx_stats_active)
3899 rx_atomic_inc(&rx_stats.dupPacketsRead);
3900 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3901 rxevent_Cancel(&call->delayedAckEvent, call,
3902 RX_CALL_REFCOUNT_DELAY);
3903 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3909 /* It's the next packet. Stick it on the receive queue
3910 * for this call. Set newPackets to make sure we wake
3911 * the reader once all packets have been processed */
3912 #ifdef RX_TRACK_PACKETS
3913 np->flags |= RX_PKTFLAG_RQ;
3915 opr_queue_Prepend(&call->rq, &np->entry);
3916 #ifdef RXDEBUG_PACKET
3918 #endif /* RXDEBUG_PACKET */
3920 np = NULL; /* We can't use this anymore */
3923 /* If an ack is requested then set a flag to make sure we
3924 * send an acknowledgement for this packet */
3925 if (flags & RX_REQUEST_ACK) {
3926 ackNeeded = RX_ACK_REQUESTED;
3929 /* Keep track of whether we have received the last packet */
3930 if (flags & RX_LAST_PACKET) {
3931 call->flags |= RX_CALL_HAVE_LAST;
3935 /* Check whether we have all of the packets for this call */
3936 if (call->flags & RX_CALL_HAVE_LAST) {
3937 afs_uint32 tseq; /* temporary sequence number */
3938 struct opr_queue *cursor;
3940 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
3941 struct rx_packet *tp;
3943 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
3944 if (tseq != tp->header.seq)
3946 if (tp->header.flags & RX_LAST_PACKET) {
3947 call->flags |= RX_CALL_RECEIVE_DONE;
3954 /* Provide asynchronous notification for those who want it
3955 * (e.g. multi rx) */
3956 if (call->arrivalProc) {
3957 (*call->arrivalProc) (call, call->arrivalProcHandle,
3958 call->arrivalProcArg);
3959 call->arrivalProc = (void (*)())0;
3962 /* Update last packet received */
3965 /* If there is no server process serving this call, grab
3966 * one, if available. We only need to do this once. If a
3967 * server thread is available, this thread becomes a server
3968 * thread and the server thread becomes a listener thread. */
3970 TryAttach(call, socket, tnop, newcallp, 0);
3973 /* This is not the expected next packet. */
3975 /* Determine whether this is a new or old packet, and if it's
3976 * a new one, whether it fits into the current receive window.
3977 * Also figure out whether the packet was delivered in sequence.
3978 * We use the prev variable to determine whether the new packet
3979 * is the successor of its immediate predecessor in the
3980 * receive queue, and the missing flag to determine whether
3981 * any of this packets predecessors are missing. */
3983 afs_uint32 prev; /* "Previous packet" sequence number */
3984 struct opr_queue *cursor;
3985 int missing; /* Are any predecessors missing? */
3987 /* If the new packet's sequence number has been sent to the
3988 * application already, then this is a duplicate */
3989 if (seq < call->rnext) {
3990 if (rx_stats_active)
3991 rx_atomic_inc(&rx_stats.dupPacketsRead);
3992 rxevent_Cancel(&call->delayedAckEvent, call,
3993 RX_CALL_REFCOUNT_DELAY);
3994 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4000 /* If the sequence number is greater than what can be
4001 * accomodated by the current window, then send a negative
4002 * acknowledge and drop the packet */
4003 if ((call->rnext + call->rwind) <= seq) {
4004 rxevent_Cancel(&call->delayedAckEvent, call,
4005 RX_CALL_REFCOUNT_DELAY);
4006 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4013 /* Look for the packet in the queue of old received packets */
4014 prev = call->rnext - 1;
4016 for (opr_queue_Scan(&call->rq, cursor)) {
4017 struct rx_packet *tp
4018 = opr_queue_Entry(cursor, struct rx_packet, entry);
4020 /*Check for duplicate packet */
4021 if (seq == tp->header.seq) {
4022 if (rx_stats_active)
4023 rx_atomic_inc(&rx_stats.dupPacketsRead);
4024 rxevent_Cancel(&call->delayedAckEvent, call,
4025 RX_CALL_REFCOUNT_DELAY);
4026 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4032 /* If we find a higher sequence packet, break out and
4033 * insert the new packet here. */
4034 if (seq < tp->header.seq)
4036 /* Check for missing packet */
4037 if (tp->header.seq != prev + 1) {
4041 prev = tp->header.seq;
4044 /* Keep track of whether we have received the last packet. */
4045 if (flags & RX_LAST_PACKET) {
4046 call->flags |= RX_CALL_HAVE_LAST;
4049 /* It's within the window: add it to the the receive queue.
4050 * tp is left by the previous loop either pointing at the
4051 * packet before which to insert the new packet, or at the
4052 * queue head if the queue is empty or the packet should be
4054 #ifdef RX_TRACK_PACKETS
4055 np->flags |= RX_PKTFLAG_RQ;
4057 #ifdef RXDEBUG_PACKET
4059 #endif /* RXDEBUG_PACKET */
4060 opr_queue_InsertBefore(cursor, &np->entry);
4064 /* Check whether we have all of the packets for this call */
4065 if ((call->flags & RX_CALL_HAVE_LAST)
4066 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4067 afs_uint32 tseq; /* temporary sequence number */
4070 for (opr_queue_Scan(&call->rq, cursor)) {
4071 struct rx_packet *tp
4072 = opr_queue_Entry(cursor, struct rx_packet, entry);
4073 if (tseq != tp->header.seq)
4075 if (tp->header.flags & RX_LAST_PACKET) {
4076 call->flags |= RX_CALL_RECEIVE_DONE;
4083 /* We need to send an ack of the packet is out of sequence,
4084 * or if an ack was requested by the peer. */
4085 if (seq != prev + 1 || missing) {
4086 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4087 } else if (flags & RX_REQUEST_ACK) {
4088 ackNeeded = RX_ACK_REQUESTED;
4091 /* Acknowledge the last packet for each call */
4092 if (flags & RX_LAST_PACKET) {
4103 * If the receiver is waiting for an iovec, fill the iovec
4104 * using the data from the receive queue */
4105 if (call->flags & RX_CALL_IOVEC_WAIT) {
4106 didHardAck = rxi_FillReadVec(call, serial);
4107 /* the call may have been aborted */
4116 /* Wakeup the reader if any */
4117 if ((call->flags & RX_CALL_READER_WAIT)
4118 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4119 || (call->iovNext >= call->iovMax)
4120 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4121 call->flags &= ~RX_CALL_READER_WAIT;
4122 #ifdef RX_ENABLE_LOCKS
4123 CV_BROADCAST(&call->cv_rq);
4125 osi_rxWakeup(&call->rq);
4131 * Send an ack when requested by the peer, or once every
4132 * rxi_SoftAckRate packets until the last packet has been
4133 * received. Always send a soft ack for the last packet in
4134 * the server's reply. */
4136 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4137 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4138 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4139 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4140 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4141 } else if (call->nSoftAcks) {
4142 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4143 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4145 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4146 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4147 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4154 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4156 struct rx_peer *peer = conn->peer;
4158 MUTEX_ENTER(&peer->peer_lock);
4159 peer->lastReachTime = clock_Sec();
4160 MUTEX_EXIT(&peer->peer_lock);
4162 MUTEX_ENTER(&conn->conn_data_lock);
4163 if (conn->flags & RX_CONN_ATTACHWAIT) {
4166 rxi_ConnClearAttachWait(conn);
4167 MUTEX_EXIT(&conn->conn_data_lock);
4169 for (i = 0; i < RX_MAXCALLS; i++) {
4170 struct rx_call *call = conn->call[i];
4173 MUTEX_ENTER(&call->lock);
4174 /* tnop can be null if newcallp is null */
4175 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4177 MUTEX_EXIT(&call->lock);
4181 MUTEX_EXIT(&conn->conn_data_lock);
4184 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4186 rx_ack_reason(int reason)
4189 case RX_ACK_REQUESTED:
4191 case RX_ACK_DUPLICATE:
4193 case RX_ACK_OUT_OF_SEQUENCE:
4195 case RX_ACK_EXCEEDS_WINDOW:
4197 case RX_ACK_NOSPACE:
4201 case RX_ACK_PING_RESPONSE:
4214 /* The real smarts of the whole thing. */
4215 static struct rx_packet *
4216 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4219 struct rx_ackPacket *ap;
4221 struct rx_packet *tp;
4222 struct rx_connection *conn = call->conn;
4223 struct rx_peer *peer = conn->peer;
4224 struct opr_queue *cursor;
4225 struct clock now; /* Current time, for RTT calculations */
4233 int newAckCount = 0;
4234 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4235 int pktsize = 0; /* Set if we need to update the peer mtu */
4236 int conn_data_locked = 0;
4238 if (rx_stats_active)
4239 rx_atomic_inc(&rx_stats.ackPacketsRead);
4240 ap = (struct rx_ackPacket *)rx_DataOf(np);
4241 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4243 return np; /* truncated ack packet */
4245 /* depends on ack packet struct */
4246 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4247 first = ntohl(ap->firstPacket);
4248 prev = ntohl(ap->previousPacket);
4249 serial = ntohl(ap->serial);
4252 * Ignore ack packets received out of order while protecting
4253 * against peers that set the previousPacket field to a packet
4254 * serial number instead of a sequence number.
4256 if (first < call->tfirst ||
4257 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4264 if (np->header.flags & RX_SLOW_START_OK) {
4265 call->flags |= RX_CALL_SLOW_START_OK;
4268 if (ap->reason == RX_ACK_PING_RESPONSE)
4269 rxi_UpdatePeerReach(conn, call);
4271 if (conn->lastPacketSizeSeq) {
4272 MUTEX_ENTER(&conn->conn_data_lock);
4273 conn_data_locked = 1;
4274 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4275 pktsize = conn->lastPacketSize;
4276 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4279 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4280 if (!conn_data_locked) {
4281 MUTEX_ENTER(&conn->conn_data_lock);
4282 conn_data_locked = 1;
4284 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4285 /* process mtu ping ack */
4286 pktsize = conn->lastPingSize;
4287 conn->lastPingSizeSer = conn->lastPingSize = 0;
4291 if (conn_data_locked) {
4292 MUTEX_EXIT(&conn->conn_data_lock);
4293 conn_data_locked = 0;
4297 if (rxdebug_active) {
4301 len = _snprintf(msg, sizeof(msg),
4302 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4303 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4304 ntohl(ap->serial), ntohl(ap->previousPacket),
4305 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4306 ap->nAcks, ntohs(ap->bufferSpace) );
4310 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4311 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4315 OutputDebugString(msg);
4317 #else /* AFS_NT40_ENV */
4320 "RACK: reason %x previous %u seq %u serial %u first %u",
4321 ap->reason, ntohl(ap->previousPacket),
4322 (unsigned int)np->header.seq, (unsigned int)serial,
4323 ntohl(ap->firstPacket));
4326 for (offset = 0; offset < nAcks; offset++)
4327 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4332 #endif /* AFS_NT40_ENV */
4335 MUTEX_ENTER(&peer->peer_lock);
4338 * Start somewhere. Can't assume we can send what we can receive,
4339 * but we are clearly receiving.
4341 if (!peer->maxPacketSize)
4342 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4344 if (pktsize > peer->maxPacketSize) {
4345 peer->maxPacketSize = pktsize;
4346 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4347 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4348 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4349 rxi_ScheduleGrowMTUEvent(call, 1);
4354 clock_GetTime(&now);
4356 /* The transmit queue splits into 4 sections.
4358 * The first section is packets which have now been acknowledged
4359 * by a window size change in the ack. These have reached the
4360 * application layer, and may be discarded. These are packets
4361 * with sequence numbers < ap->firstPacket.
4363 * The second section is packets which have sequence numbers in
4364 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4365 * contents of the packet's ack array determines whether these
4366 * packets are acknowledged or not.
4368 * The third section is packets which fall above the range
4369 * addressed in the ack packet. These have not yet been received
4372 * The four section is packets which have not yet been transmitted.
4373 * These packets will have a header.serial of 0.
4376 /* First section - implicitly acknowledged packets that can be
4380 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4381 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4382 struct rx_packet *next;
4384 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4385 call->tfirst = tp->header.seq + 1;
4387 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4389 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4392 #ifdef RX_ENABLE_LOCKS
4393 /* XXX Hack. Because we have to release the global call lock when sending
4394 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4395 * in rxi_Start sending packets out because packets may move to the
4396 * freePacketQueue as result of being here! So we drop these packets until
4397 * we're safely out of the traversing. Really ugly!
4398 * To make it even uglier, if we're using fine grain locking, we can
4399 * set the ack bits in the packets and have rxi_Start remove the packets
4400 * when it's done transmitting.
4402 if (call->flags & RX_CALL_TQ_BUSY) {
4403 tp->flags |= RX_PKTFLAG_ACKED;
4404 call->flags |= RX_CALL_TQ_SOME_ACKED;
4406 #endif /* RX_ENABLE_LOCKS */
4408 opr_queue_Remove(&tp->entry);
4409 #ifdef RX_TRACK_PACKETS
4410 tp->flags &= ~RX_PKTFLAG_TQ;
4412 #ifdef RXDEBUG_PACKET
4414 #endif /* RXDEBUG_PACKET */
4415 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4420 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4422 /* Second section of the queue - packets for which we are receiving
4425 * Go through the explicit acks/nacks and record the results in
4426 * the waiting packets. These are packets that can't be released
4427 * yet, even with a positive acknowledge. This positive
4428 * acknowledge only means the packet has been received by the
4429 * peer, not that it will be retained long enough to be sent to
4430 * the peer's upper level. In addition, reset the transmit timers
4431 * of any missing packets (those packets that must be missing
4432 * because this packet was out of sequence) */
4434 call->nSoftAcked = 0;
4436 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4437 && tp->header.seq < first + nAcks) {
4438 /* Set the acknowledge flag per packet based on the
4439 * information in the ack packet. An acknowlegded packet can
4440 * be downgraded when the server has discarded a packet it
4441 * soacked previously, or when an ack packet is received
4442 * out of sequence. */
4443 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4444 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4446 tp->flags |= RX_PKTFLAG_ACKED;
4447 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4454 } else /* RX_ACK_TYPE_NACK */ {
4455 tp->flags &= ~RX_PKTFLAG_ACKED;
4459 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4462 /* We don't need to take any action with the 3rd or 4th section in the
4463 * queue - they're not addressed by the contents of this ACK packet.
4466 /* If the window has been extended by this acknowledge packet,
4467 * then wakeup a sender waiting in alloc for window space, or try
4468 * sending packets now, if he's been sitting on packets due to
4469 * lack of window space */
4470 if (call->tnext < (call->tfirst + call->twind)) {
4471 #ifdef RX_ENABLE_LOCKS
4472 CV_SIGNAL(&call->cv_twind);
4474 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4475 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4476 osi_rxWakeup(&call->twind);
4479 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4480 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4484 /* if the ack packet has a receivelen field hanging off it,
4485 * update our state */
4486 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4489 /* If the ack packet has a "recommended" size that is less than
4490 * what I am using now, reduce my size to match */
4491 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4492 (int)sizeof(afs_int32), &tSize);
4493 tSize = (afs_uint32) ntohl(tSize);
4494 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4496 /* Get the maximum packet size to send to this peer */
4497 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4499 tSize = (afs_uint32) ntohl(tSize);
4500 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4501 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4503 /* sanity check - peer might have restarted with different params.
4504 * If peer says "send less", dammit, send less... Peer should never
4505 * be unable to accept packets of the size that prior AFS versions would
4506 * send without asking. */
4507 if (peer->maxMTU != tSize) {
4508 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4510 peer->maxMTU = tSize;
4511 peer->MTU = MIN(tSize, peer->MTU);
4512 call->MTU = MIN(call->MTU, tSize);
4515 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4518 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4519 (int)sizeof(afs_int32), &tSize);
4520 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4521 if (tSize < call->twind) { /* smaller than our send */
4522 call->twind = tSize; /* window, we must send less... */
4523 call->ssthresh = MIN(call->twind, call->ssthresh);
4524 call->conn->twind[call->channel] = call->twind;
4527 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4528 * network MTU confused with the loopback MTU. Calculate the
4529 * maximum MTU here for use in the slow start code below.
4531 /* Did peer restart with older RX version? */
4532 if (peer->maxDgramPackets > 1) {
4533 peer->maxDgramPackets = 1;
4535 } else if (np->length >=
4536 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4539 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4540 sizeof(afs_int32), &tSize);
4541 tSize = (afs_uint32) ntohl(tSize);
4543 * As of AFS 3.5 we set the send window to match the receive window.
4545 if (tSize < call->twind) {
4546 call->twind = tSize;
4547 call->conn->twind[call->channel] = call->twind;
4548 call->ssthresh = MIN(call->twind, call->ssthresh);
4549 } else if (tSize > call->twind) {
4550 call->twind = tSize;
4551 call->conn->twind[call->channel] = call->twind;
4555 * As of AFS 3.5, a jumbogram is more than one fixed size
4556 * packet transmitted in a single UDP datagram. If the remote
4557 * MTU is smaller than our local MTU then never send a datagram
4558 * larger than the natural MTU.
4561 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4562 (int)sizeof(afs_int32), &tSize);
4563 maxDgramPackets = (afs_uint32) ntohl(tSize);
4564 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4566 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4567 if (maxDgramPackets > 1) {
4568 peer->maxDgramPackets = maxDgramPackets;
4569 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4571 peer->maxDgramPackets = 1;
4572 call->MTU = peer->natMTU;
4574 } else if (peer->maxDgramPackets > 1) {
4575 /* Restarted with lower version of RX */
4576 peer->maxDgramPackets = 1;
4578 } else if (peer->maxDgramPackets > 1
4579 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4580 /* Restarted with lower version of RX */
4581 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4582 peer->natMTU = OLD_MAX_PACKET_SIZE;
4583 peer->MTU = OLD_MAX_PACKET_SIZE;
4584 peer->maxDgramPackets = 1;
4585 peer->nDgramPackets = 1;
4587 call->MTU = OLD_MAX_PACKET_SIZE;
4592 * Calculate how many datagrams were successfully received after
4593 * the first missing packet and adjust the negative ack counter
4598 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4599 if (call->nNacks < nNacked) {
4600 call->nNacks = nNacked;
4603 call->nAcks += newAckCount;
4607 /* If the packet contained new acknowledgements, rather than just
4608 * being a duplicate of one we have previously seen, then we can restart
4611 if (newAckCount > 0)
4612 rxi_rto_packet_acked(call, istack);
4614 if (call->flags & RX_CALL_FAST_RECOVER) {
4615 if (newAckCount == 0) {
4616 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4618 call->flags &= ~RX_CALL_FAST_RECOVER;
4619 call->cwind = call->nextCwind;
4620 call->nextCwind = 0;
4623 call->nCwindAcks = 0;
4624 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4625 /* Three negative acks in a row trigger congestion recovery */
4626 call->flags |= RX_CALL_FAST_RECOVER;
4627 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4629 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4630 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4631 call->nextCwind = call->ssthresh;
4634 peer->MTU = call->MTU;
4635 peer->cwind = call->nextCwind;
4636 peer->nDgramPackets = call->nDgramPackets;
4638 call->congestSeq = peer->congestSeq;
4640 /* Reset the resend times on the packets that were nacked
4641 * so we will retransmit as soon as the window permits
4645 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4646 struct rx_packet *tp =
4647 opr_queue_Entry(cursor, struct rx_packet, entry);
4649 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4650 tp->flags &= ~RX_PKTFLAG_SENT;
4652 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4657 /* If cwind is smaller than ssthresh, then increase
4658 * the window one packet for each ack we receive (exponential
4660 * If cwind is greater than or equal to ssthresh then increase
4661 * the congestion window by one packet for each cwind acks we
4662 * receive (linear growth). */
4663 if (call->cwind < call->ssthresh) {
4665 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4666 call->nCwindAcks = 0;
4668 call->nCwindAcks += newAckCount;
4669 if (call->nCwindAcks >= call->cwind) {
4670 call->nCwindAcks = 0;
4671 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4675 * If we have received several acknowledgements in a row then
4676 * it is time to increase the size of our datagrams
4678 if ((int)call->nAcks > rx_nDgramThreshold) {
4679 if (peer->maxDgramPackets > 1) {
4680 if (call->nDgramPackets < peer->maxDgramPackets) {
4681 call->nDgramPackets++;
4683 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4684 } else if (call->MTU < peer->maxMTU) {
4685 /* don't upgrade if we can't handle it */
4686 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4687 call->MTU = peer->ifMTU;
4689 call->MTU += peer->natMTU;
4690 call->MTU = MIN(call->MTU, peer->maxMTU);
4697 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4699 /* Servers need to hold the call until all response packets have
4700 * been acknowledged. Soft acks are good enough since clients
4701 * are not allowed to clear their receive queues. */
4702 if (call->state == RX_STATE_HOLD
4703 && call->tfirst + call->nSoftAcked >= call->tnext) {
4704 call->state = RX_STATE_DALLY;
4705 rxi_ClearTransmitQueue(call, 0);
4706 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4707 } else if (!opr_queue_IsEmpty(&call->tq)) {
4708 rxi_Start(call, istack);
4713 /* Received a response to a challenge packet */
4714 static struct rx_packet *
4715 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4716 struct rx_packet *np, int istack)
4720 /* Ignore the packet if we're the client */
4721 if (conn->type == RX_CLIENT_CONNECTION)
4724 /* If already authenticated, ignore the packet (it's probably a retry) */
4725 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4728 /* Otherwise, have the security object evaluate the response packet */
4729 error = RXS_CheckResponse(conn->securityObject, conn, np);
4731 /* If the response is invalid, reset the connection, sending
4732 * an abort to the peer */
4736 rxi_ConnectionError(conn, error);
4737 MUTEX_ENTER(&conn->conn_data_lock);
4738 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4739 MUTEX_EXIT(&conn->conn_data_lock);
4742 /* If the response is valid, any calls waiting to attach
4743 * servers can now do so */
4746 for (i = 0; i < RX_MAXCALLS; i++) {
4747 struct rx_call *call = conn->call[i];
4749 MUTEX_ENTER(&call->lock);
4750 if (call->state == RX_STATE_PRECALL)
4751 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4752 /* tnop can be null if newcallp is null */
4753 MUTEX_EXIT(&call->lock);
4757 /* Update the peer reachability information, just in case
4758 * some calls went into attach-wait while we were waiting
4759 * for authentication..
4761 rxi_UpdatePeerReach(conn, NULL);
4766 /* A client has received an authentication challenge: the security
4767 * object is asked to cough up a respectable response packet to send
4768 * back to the server. The server is responsible for retrying the
4769 * challenge if it fails to get a response. */
4771 static struct rx_packet *
4772 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4773 struct rx_packet *np, int istack)
4777 /* Ignore the challenge if we're the server */
4778 if (conn->type == RX_SERVER_CONNECTION)
4781 /* Ignore the challenge if the connection is otherwise idle; someone's
4782 * trying to use us as an oracle. */
4783 if (!rxi_HasActiveCalls(conn))
4786 /* Send the security object the challenge packet. It is expected to fill
4787 * in the response. */
4788 error = RXS_GetResponse(conn->securityObject, conn, np);
4790 /* If the security object is unable to return a valid response, reset the
4791 * connection and send an abort to the peer. Otherwise send the response
4792 * packet to the peer connection. */
4794 rxi_ConnectionError(conn, error);
4795 MUTEX_ENTER(&conn->conn_data_lock);
4796 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4797 MUTEX_EXIT(&conn->conn_data_lock);
4799 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4800 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4806 /* Find an available server process to service the current request in
4807 * the given call structure. If one isn't available, queue up this
4808 * call so it eventually gets one */
4810 rxi_AttachServerProc(struct rx_call *call,
4811 osi_socket socket, int *tnop,
4812 struct rx_call **newcallp)
4814 struct rx_serverQueueEntry *sq;
4815 struct rx_service *service = call->conn->service;
4818 /* May already be attached */
4819 if (call->state == RX_STATE_ACTIVE)
4822 MUTEX_ENTER(&rx_serverPool_lock);
4824 haveQuota = QuotaOK(service);
4825 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4826 /* If there are no processes available to service this call,
4827 * put the call on the incoming call queue (unless it's
4828 * already on the queue).
4830 #ifdef RX_ENABLE_LOCKS
4832 ReturnToServerPool(service);
4833 #endif /* RX_ENABLE_LOCKS */
4835 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4836 call->flags |= RX_CALL_WAIT_PROC;
4837 rx_atomic_inc(&rx_nWaiting);
4838 rx_atomic_inc(&rx_nWaited);
4839 rxi_calltrace(RX_CALL_ARRIVAL, call);
4840 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4841 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4844 sq = opr_queue_Last(&rx_idleServerQueue,
4845 struct rx_serverQueueEntry, entry);
4847 /* If hot threads are enabled, and both newcallp and sq->socketp
4848 * are non-null, then this thread will process the call, and the
4849 * idle server thread will start listening on this threads socket.
4851 opr_queue_Remove(&sq->entry);
4853 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4856 *sq->socketp = socket;
4857 clock_GetTime(&call->startTime);
4858 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4862 if (call->flags & RX_CALL_WAIT_PROC) {
4863 /* Conservative: I don't think this should happen */
4864 call->flags &= ~RX_CALL_WAIT_PROC;
4865 rx_atomic_dec(&rx_nWaiting);
4866 if (opr_queue_IsOnQueue(&call->entry)) {
4867 opr_queue_Remove(&call->entry);
4870 call->state = RX_STATE_ACTIVE;
4871 call->app.mode = RX_MODE_RECEIVING;
4872 #ifdef RX_KERNEL_TRACE
4874 int glockOwner = ISAFS_GLOCK();
4877 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4878 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4884 if (call->flags & RX_CALL_CLEARED) {
4885 /* send an ack now to start the packet flow up again */
4886 call->flags &= ~RX_CALL_CLEARED;
4887 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4889 #ifdef RX_ENABLE_LOCKS
4892 service->nRequestsRunning++;
4893 MUTEX_ENTER(&rx_quota_mutex);
4894 if (service->nRequestsRunning <= service->minProcs)
4897 MUTEX_EXIT(&rx_quota_mutex);
4901 MUTEX_EXIT(&rx_serverPool_lock);
4904 /* Delay the sending of an acknowledge event for a short while, while
4905 * a new call is being prepared (in the case of a client) or a reply
4906 * is being prepared (in the case of a server). Rather than sending
4907 * an ack packet, an ACKALL packet is sent. */
4909 rxi_AckAll(struct rx_call *call)
4911 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4913 call->flags |= RX_CALL_ACKALL_SENT;
4917 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4920 struct rx_call *call = arg1;
4921 #ifdef RX_ENABLE_LOCKS
4923 MUTEX_ENTER(&call->lock);
4924 if (event == call->delayedAckEvent) {
4925 rxevent_Put(call->delayedAckEvent);
4926 call->delayedAckEvent = NULL;
4928 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4930 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4932 MUTEX_EXIT(&call->lock);
4933 #else /* RX_ENABLE_LOCKS */
4935 rxevent_Put(call->delayedAckEvent);
4936 call->delayedAckEvent = NULL;
4938 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4939 #endif /* RX_ENABLE_LOCKS */
4942 #ifdef RX_ENABLE_LOCKS
4943 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4944 * clearing them out.
4947 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4949 struct opr_queue *cursor;
4952 for (opr_queue_Scan(&call->tq, cursor)) {
4954 = opr_queue_Entry(cursor, struct rx_packet, entry);
4956 p->flags |= RX_PKTFLAG_ACKED;
4961 call->flags |= RX_CALL_TQ_CLEARME;
4962 call->flags |= RX_CALL_TQ_SOME_ACKED;
4965 rxi_rto_cancel(call);
4967 call->tfirst = call->tnext;
4968 call->nSoftAcked = 0;
4970 if (call->flags & RX_CALL_FAST_RECOVER) {
4971 call->flags &= ~RX_CALL_FAST_RECOVER;
4972 call->cwind = call->nextCwind;
4973 call->nextCwind = 0;
4976 CV_SIGNAL(&call->cv_twind);
4978 #endif /* RX_ENABLE_LOCKS */
4981 * Acknowledge the whole transmit queue.
4983 * If we're running without locks, or the transmit queue isn't busy, then
4984 * we can just clear the queue now. Otherwise, we have to mark all of the
4985 * packets as acknowledged, and let rxi_Start clear it later on
4988 rxi_AckAllInTransmitQueue(struct rx_call *call)
4990 #ifdef RX_ENABLE_LOCKS
4991 if (call->flags & RX_CALL_TQ_BUSY) {
4992 rxi_SetAcksInTransmitQueue(call);
4996 rxi_ClearTransmitQueue(call, 0);
4998 /* Clear out the transmit queue for the current call (all packets have
4999 * been received by peer) */
5001 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5003 #ifdef RX_ENABLE_LOCKS
5004 struct opr_queue *cursor;
5005 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5007 for (opr_queue_Scan(&call->tq, cursor)) {
5009 = opr_queue_Entry(cursor, struct rx_packet, entry);
5011 p->flags |= RX_PKTFLAG_ACKED;
5015 call->flags |= RX_CALL_TQ_CLEARME;
5016 call->flags |= RX_CALL_TQ_SOME_ACKED;
5019 #endif /* RX_ENABLE_LOCKS */
5020 #ifdef RXDEBUG_PACKET
5022 #endif /* RXDEBUG_PACKET */
5023 rxi_FreePackets(0, &call->tq);
5024 rxi_WakeUpTransmitQueue(call);
5025 #ifdef RX_ENABLE_LOCKS
5026 call->flags &= ~RX_CALL_TQ_CLEARME;
5030 rxi_rto_cancel(call);
5031 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5032 call->nSoftAcked = 0;
5034 if (call->flags & RX_CALL_FAST_RECOVER) {
5035 call->flags &= ~RX_CALL_FAST_RECOVER;
5036 call->cwind = call->nextCwind;
5038 #ifdef RX_ENABLE_LOCKS
5039 CV_SIGNAL(&call->cv_twind);
5041 osi_rxWakeup(&call->twind);
5046 rxi_ClearReceiveQueue(struct rx_call *call)
5048 if (!opr_queue_IsEmpty(&call->rq)) {
5051 count = rxi_FreePackets(0, &call->rq);
5052 rx_packetReclaims += count;
5053 #ifdef RXDEBUG_PACKET
5055 if ( call->rqc != 0 )
5056 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5058 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5060 if (call->state == RX_STATE_PRECALL) {
5061 call->flags |= RX_CALL_CLEARED;
5065 /* Send an abort packet for the specified call */
5066 static struct rx_packet *
5067 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5068 int istack, int force)
5070 afs_int32 error, cerror;
5071 struct clock when, now;
5076 switch (call->error) {
5079 cerror = RX_CALL_TIMEOUT;
5082 cerror = call->error;
5085 /* Clients should never delay abort messages */
5086 if (rx_IsClientConn(call->conn))
5089 if (call->abortCode != cerror) {
5090 call->abortCode = cerror;
5091 call->abortCount = 0;
5094 if (force || rxi_callAbortThreshhold == 0
5095 || call->abortCount < rxi_callAbortThreshhold) {
5096 if (call->delayedAbortEvent) {
5097 rxevent_Cancel(&call->delayedAbortEvent, call,
5098 RX_CALL_REFCOUNT_ABORT);
5100 error = htonl(cerror);
5103 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5104 (char *)&error, sizeof(error), istack);
5105 } else if (!call->delayedAbortEvent) {
5106 clock_GetTime(&now);
5108 clock_Addmsec(&when, rxi_callAbortDelay);
5109 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5110 call->delayedAbortEvent =
5111 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5116 /* Send an abort packet for the specified connection. Packet is an
5117 * optional pointer to a packet that can be used to send the abort.
5118 * Once the number of abort messages reaches the threshhold, an
5119 * event is scheduled to send the abort. Setting the force flag
5120 * overrides sending delayed abort messages.
5122 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5123 * to send the abort packet.
5126 rxi_SendConnectionAbort(struct rx_connection *conn,
5127 struct rx_packet *packet, int istack, int force)
5130 struct clock when, now;
5135 /* Clients should never delay abort messages */
5136 if (rx_IsClientConn(conn))
5139 if (force || rxi_connAbortThreshhold == 0
5140 || conn->abortCount < rxi_connAbortThreshhold) {
5142 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5143 error = htonl(conn->error);
5145 MUTEX_EXIT(&conn->conn_data_lock);
5147 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5148 RX_PACKET_TYPE_ABORT, (char *)&error,
5149 sizeof(error), istack);
5150 MUTEX_ENTER(&conn->conn_data_lock);
5151 } else if (!conn->delayedAbortEvent) {
5152 clock_GetTime(&now);
5154 clock_Addmsec(&when, rxi_connAbortDelay);
5155 conn->delayedAbortEvent =
5156 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5161 /* Associate an error all of the calls owned by a connection. Called
5162 * with error non-zero. This is only for really fatal things, like
5163 * bad authentication responses. The connection itself is set in
5164 * error at this point, so that future packets received will be
5167 rxi_ConnectionError(struct rx_connection *conn,
5173 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5175 MUTEX_ENTER(&conn->conn_data_lock);
5176 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5177 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5178 if (conn->checkReachEvent) {
5179 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5180 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5181 putConnection(conn);
5183 MUTEX_EXIT(&conn->conn_data_lock);
5184 for (i = 0; i < RX_MAXCALLS; i++) {
5185 struct rx_call *call = conn->call[i];
5187 MUTEX_ENTER(&call->lock);
5188 rxi_CallError(call, error);
5189 MUTEX_EXIT(&call->lock);
5192 conn->error = error;
5193 if (rx_stats_active)
5194 rx_atomic_inc(&rx_stats.fatalErrors);
5199 * Interrupt an in-progress call with the specified error and wakeup waiters.
5201 * @param[in] call The call to interrupt
5202 * @param[in] error The error code to send to the peer
5205 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5207 MUTEX_ENTER(&call->lock);
5208 rxi_CallError(call, error);
5209 rxi_SendCallAbort(call, NULL, 0, 1);
5210 MUTEX_EXIT(&call->lock);
5214 rxi_CallError(struct rx_call *call, afs_int32 error)
5216 MUTEX_ASSERT(&call->lock);
5217 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5219 error = call->error;
5221 #ifdef RX_ENABLE_LOCKS
5222 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5223 rxi_ResetCall(call, 0);
5226 rxi_ResetCall(call, 0);
5228 call->error = error;
5231 /* Reset various fields in a call structure, and wakeup waiting
5232 * processes. Some fields aren't changed: state & mode are not
5233 * touched (these must be set by the caller), and bufptr, nLeft, and
5234 * nFree are not reset, since these fields are manipulated by
5235 * unprotected macros, and may only be reset by non-interrupting code.
5239 rxi_ResetCall(struct rx_call *call, int newcall)
5242 struct rx_peer *peer;
5243 struct rx_packet *packet;
5245 MUTEX_ASSERT(&call->lock);
5246 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5248 /* Notify anyone who is waiting for asynchronous packet arrival */
5249 if (call->arrivalProc) {
5250 (*call->arrivalProc) (call, call->arrivalProcHandle,
5251 call->arrivalProcArg);
5252 call->arrivalProc = (void (*)())0;
5256 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5258 if (call->delayedAbortEvent) {
5259 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5260 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5262 rxi_SendCallAbort(call, packet, 0, 1);
5263 rxi_FreePacket(packet);
5268 * Update the peer with the congestion information in this call
5269 * so other calls on this connection can pick up where this call
5270 * left off. If the congestion sequence numbers don't match then
5271 * another call experienced a retransmission.
5273 peer = call->conn->peer;
5274 MUTEX_ENTER(&peer->peer_lock);
5276 if (call->congestSeq == peer->congestSeq) {
5277 peer->cwind = MAX(peer->cwind, call->cwind);
5278 peer->MTU = MAX(peer->MTU, call->MTU);
5279 peer->nDgramPackets =
5280 MAX(peer->nDgramPackets, call->nDgramPackets);
5283 call->abortCode = 0;
5284 call->abortCount = 0;
5286 if (peer->maxDgramPackets > 1) {
5287 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5289 call->MTU = peer->MTU;
5291 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5292 call->ssthresh = rx_maxSendWindow;
5293 call->nDgramPackets = peer->nDgramPackets;
5294 call->congestSeq = peer->congestSeq;
5295 call->rtt = peer->rtt;
5296 call->rtt_dev = peer->rtt_dev;
5297 clock_Zero(&call->rto);
5298 clock_Addmsec(&call->rto,
5299 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5300 MUTEX_EXIT(&peer->peer_lock);
5302 flags = call->flags;
5303 rxi_WaitforTQBusy(call);
5305 rxi_ClearTransmitQueue(call, 1);
5306 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5307 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5311 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5312 /* The call channel is still busy; resetting the call doesn't change
5313 * that. However, if 'newcall' is set, we are processing a call
5314 * structure that has either been recycled from the free list, or has
5315 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5316 * 'newcall' is set, since it describes a completely different call
5317 * channel which we do not care about. */
5318 call->flags |= RX_CALL_PEER_BUSY;
5321 rxi_ClearReceiveQueue(call);
5322 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5326 call->twind = call->conn->twind[call->channel];
5327 call->rwind = call->conn->rwind[call->channel];
5328 call->nSoftAcked = 0;
5329 call->nextCwind = 0;
5332 call->nCwindAcks = 0;
5333 call->nSoftAcks = 0;
5334 call->nHardAcks = 0;
5336 call->tfirst = call->rnext = call->tnext = 1;
5339 call->lastAcked = 0;
5340 call->localStatus = call->remoteStatus = 0;
5342 if (flags & RX_CALL_READER_WAIT) {
5343 #ifdef RX_ENABLE_LOCKS
5344 CV_BROADCAST(&call->cv_rq);
5346 osi_rxWakeup(&call->rq);
5349 if (flags & RX_CALL_WAIT_PACKETS) {
5350 MUTEX_ENTER(&rx_freePktQ_lock);
5351 rxi_PacketsUnWait(); /* XXX */
5352 MUTEX_EXIT(&rx_freePktQ_lock);
5354 #ifdef RX_ENABLE_LOCKS
5355 CV_SIGNAL(&call->cv_twind);
5357 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5358 osi_rxWakeup(&call->twind);
5361 if (flags & RX_CALL_WAIT_PROC) {
5362 rx_atomic_dec(&rx_nWaiting);
5364 #ifdef RX_ENABLE_LOCKS
5365 /* The following ensures that we don't mess with any queue while some
5366 * other thread might also be doing so. The call_queue_lock field is
5367 * is only modified under the call lock. If the call is in the process
5368 * of being removed from a queue, the call is not locked until the
5369 * the queue lock is dropped and only then is the call_queue_lock field
5370 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5371 * Note that any other routine which removes a call from a queue has to
5372 * obtain the queue lock before examing the queue and removing the call.
5374 if (call->call_queue_lock) {
5375 MUTEX_ENTER(call->call_queue_lock);
5376 if (opr_queue_IsOnQueue(&call->entry)) {
5377 opr_queue_Remove(&call->entry);
5379 MUTEX_EXIT(call->call_queue_lock);
5380 CLEAR_CALL_QUEUE_LOCK(call);
5382 #else /* RX_ENABLE_LOCKS */
5383 if (opr_queue_IsOnQueue(&call->entry)) {
5384 opr_queue_Remove(&call->entry);
5386 #endif /* RX_ENABLE_LOCKS */
5388 rxi_KeepAliveOff(call);
5389 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5392 /* Send an acknowledge for the indicated packet (seq,serial) of the
5393 * indicated call, for the indicated reason (reason). This
5394 * acknowledge will specifically acknowledge receiving the packet, and
5395 * will also specify which other packets for this call have been
5396 * received. This routine returns the packet that was used to the
5397 * caller. The caller is responsible for freeing it or re-using it.
5398 * This acknowledgement also returns the highest sequence number
5399 * actually read out by the higher level to the sender; the sender
5400 * promises to keep around packets that have not been read by the
5401 * higher level yet (unless, of course, the sender decides to abort
5402 * the call altogether). Any of p, seq, serial, pflags, or reason may
5403 * be set to zero without ill effect. That is, if they are zero, they
5404 * will not convey any information.
5405 * NOW there is a trailer field, after the ack where it will safely be
5406 * ignored by mundanes, which indicates the maximum size packet this
5407 * host can swallow. */
5409 struct rx_packet *optionalPacket; use to send ack (or null)
5410 int seq; Sequence number of the packet we are acking
5411 int serial; Serial number of the packet
5412 int pflags; Flags field from packet header
5413 int reason; Reason an acknowledge was prompted
5417 rxi_SendAck(struct rx_call *call,
5418 struct rx_packet *optionalPacket, int serial, int reason,
5421 struct rx_ackPacket *ap;
5422 struct rx_packet *p;
5423 struct opr_queue *cursor;
5426 afs_uint32 padbytes = 0;
5427 #ifdef RX_ENABLE_TSFPQ
5428 struct rx_ts_info_t * rx_ts_info;
5432 * Open the receive window once a thread starts reading packets
5434 if (call->rnext > 1) {
5435 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5438 /* Don't attempt to grow MTU if this is a critical ping */
5439 if (reason == RX_ACK_MTU) {
5440 /* keep track of per-call attempts, if we're over max, do in small
5441 * otherwise in larger? set a size to increment by, decrease
5444 if (call->conn->peer->maxPacketSize &&
5445 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5447 padbytes = call->conn->peer->maxPacketSize+16;
5449 padbytes = call->conn->peer->maxMTU + 128;
5451 /* do always try a minimum size ping */
5452 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5454 /* subtract the ack payload */
5455 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5456 reason = RX_ACK_PING;
5459 call->nHardAcks = 0;
5460 call->nSoftAcks = 0;
5461 if (call->rnext > call->lastAcked)
5462 call->lastAcked = call->rnext;
5466 rx_computelen(p, p->length); /* reset length, you never know */
5467 } /* where that's been... */
5468 #ifdef RX_ENABLE_TSFPQ
5470 RX_TS_INFO_GET(rx_ts_info);
5471 if ((p = rx_ts_info->local_special_packet)) {
5472 rx_computelen(p, p->length);
5473 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5474 rx_ts_info->local_special_packet = p;
5475 } else { /* We won't send the ack, but don't panic. */
5476 return optionalPacket;
5480 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5481 /* We won't send the ack, but don't panic. */
5482 return optionalPacket;
5487 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5490 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5491 #ifndef RX_ENABLE_TSFPQ
5492 if (!optionalPacket)
5495 return optionalPacket;
5497 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5498 if (rx_Contiguous(p) < templ) {
5499 #ifndef RX_ENABLE_TSFPQ
5500 if (!optionalPacket)
5503 return optionalPacket;
5508 /* MTUXXX failing to send an ack is very serious. We should */
5509 /* try as hard as possible to send even a partial ack; it's */
5510 /* better than nothing. */
5511 ap = (struct rx_ackPacket *)rx_DataOf(p);
5512 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5513 ap->reason = reason;
5515 /* The skew computation used to be bogus, I think it's better now. */
5516 /* We should start paying attention to skew. XXX */
5517 ap->serial = htonl(serial);
5518 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5521 * First packet not yet forwarded to reader. When ACKALL has been
5522 * sent the peer has been told that all received packets will be
5523 * delivered to the reader. The value 'rnext' is used internally
5524 * to refer to the next packet in the receive queue that must be
5525 * delivered to the reader. From the perspective of the peer it
5526 * already has so report the last sequence number plus one if there
5527 * are packets in the receive queue awaiting processing.
5529 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5530 !opr_queue_IsEmpty(&call->rq)) {
5531 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5533 ap->firstPacket = htonl(call->rnext);
5535 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5537 /* No fear of running out of ack packet here because there can only
5538 * be at most one window full of unacknowledged packets. The window
5539 * size must be constrained to be less than the maximum ack size,
5540 * of course. Also, an ack should always fit into a single packet
5541 * -- it should not ever be fragmented. */
5543 for (opr_queue_Scan(&call->rq, cursor)) {
5544 struct rx_packet *rqp
5545 = opr_queue_Entry(cursor, struct rx_packet, entry);
5547 if (!rqp || !call->rq.next
5548 || (rqp->header.seq > (call->rnext + call->rwind))) {
5549 #ifndef RX_ENABLE_TSFPQ
5550 if (!optionalPacket)
5553 rxi_CallError(call, RX_CALL_DEAD);
5554 return optionalPacket;
5557 while (rqp->header.seq > call->rnext + offset)
5558 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5559 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5561 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5562 #ifndef RX_ENABLE_TSFPQ
5563 if (!optionalPacket)
5566 rxi_CallError(call, RX_CALL_DEAD);
5567 return optionalPacket;
5573 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5575 /* these are new for AFS 3.3 */
5576 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5577 templ = htonl(templ);
5578 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5579 templ = htonl(call->conn->peer->ifMTU);
5580 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5581 sizeof(afs_int32), &templ);
5583 /* new for AFS 3.4 */
5584 templ = htonl(call->rwind);
5585 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5586 sizeof(afs_int32), &templ);
5588 /* new for AFS 3.5 */
5589 templ = htonl(call->conn->peer->ifDgramPackets);
5590 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5591 sizeof(afs_int32), &templ);
5593 p->header.serviceId = call->conn->serviceId;
5594 p->header.cid = (call->conn->cid | call->channel);
5595 p->header.callNumber = *call->callNumber;
5597 p->header.securityIndex = call->conn->securityIndex;
5598 p->header.epoch = call->conn->epoch;
5599 p->header.type = RX_PACKET_TYPE_ACK;
5600 p->header.flags = RX_SLOW_START_OK;
5601 if (reason == RX_ACK_PING) {
5602 p->header.flags |= RX_REQUEST_ACK;
5604 p->length = padbytes +
5605 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5608 /* not fast but we can potentially use this if truncated
5609 * fragments are delivered to figure out the mtu.
5611 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5612 sizeof(afs_int32), sizeof(afs_int32),
5616 if (call->conn->type == RX_CLIENT_CONNECTION)
5617 p->header.flags |= RX_CLIENT_INITIATED;
5621 if (rxdebug_active) {
5625 len = _snprintf(msg, sizeof(msg),
5626 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5627 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5628 ntohl(ap->serial), ntohl(ap->previousPacket),
5629 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5630 ap->nAcks, ntohs(ap->bufferSpace) );
5634 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5635 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5639 OutputDebugString(msg);
5641 #else /* AFS_NT40_ENV */
5643 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5644 ap->reason, ntohl(ap->previousPacket),
5645 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5647 for (offset = 0; offset < ap->nAcks; offset++)
5648 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5653 #endif /* AFS_NT40_ENV */
5656 int i, nbytes = p->length;
5658 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5659 if (nbytes <= p->wirevec[i].iov_len) {
5662 savelen = p->wirevec[i].iov_len;
5664 p->wirevec[i].iov_len = nbytes;
5666 rxi_Send(call, p, istack);
5667 p->wirevec[i].iov_len = savelen;
5671 nbytes -= p->wirevec[i].iov_len;
5674 if (rx_stats_active)
5675 rx_atomic_inc(&rx_stats.ackPacketsSent);
5676 #ifndef RX_ENABLE_TSFPQ
5677 if (!optionalPacket)
5680 return optionalPacket; /* Return packet for re-use by caller */
5684 struct rx_packet **list;
5689 /* Send all of the packets in the list in single datagram */
5691 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5692 int istack, int moreFlag)
5698 struct rx_connection *conn = call->conn;
5699 struct rx_peer *peer = conn->peer;
5701 MUTEX_ENTER(&peer->peer_lock);
5702 peer->nSent += xmit->len;
5703 if (xmit->resending)
5704 peer->reSends += xmit->len;
5705 MUTEX_EXIT(&peer->peer_lock);
5707 if (rx_stats_active) {
5708 if (xmit->resending)
5709 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5711 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5714 clock_GetTime(&now);
5716 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5720 /* Set the packet flags and schedule the resend events */
5721 /* Only request an ack for the last packet in the list */
5722 for (i = 0; i < xmit->len; i++) {
5723 struct rx_packet *packet = xmit->list[i];
5725 /* Record the time sent */
5726 packet->timeSent = now;
5727 packet->flags |= RX_PKTFLAG_SENT;
5729 /* Ask for an ack on retransmitted packets, on every other packet
5730 * if the peer doesn't support slow start. Ask for an ack on every
5731 * packet until the congestion window reaches the ack rate. */
5732 if (packet->header.serial) {
5735 packet->firstSent = now;
5736 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5737 || (!(call->flags & RX_CALL_SLOW_START_OK)
5738 && (packet->header.seq & 1)))) {
5743 /* Tag this packet as not being the last in this group,
5744 * for the receiver's benefit */
5745 if (i < xmit->len - 1 || moreFlag) {
5746 packet->header.flags |= RX_MORE_PACKETS;
5751 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5754 /* Since we're about to send a data packet to the peer, it's
5755 * safe to nuke any scheduled end-of-packets ack */
5756 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5758 MUTEX_EXIT(&call->lock);
5759 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5760 if (xmit->len > 1) {
5761 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5763 rxi_SendPacket(call, conn, xmit->list[0], istack);
5765 MUTEX_ENTER(&call->lock);
5766 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5768 /* Tell the RTO calculation engine that we have sent a packet, and
5769 * if it was the last one */
5770 rxi_rto_packet_sent(call, lastPacket, istack);
5772 /* Update last send time for this call (for keep-alive
5773 * processing), and for the connection (so that we can discover
5774 * idle connections) */
5775 conn->lastSendTime = call->lastSendTime = clock_Sec();
5776 /* Let a set of retransmits trigger an idle timeout */
5777 if (!xmit->resending)
5778 call->lastSendData = call->lastSendTime;
5781 /* When sending packets we need to follow these rules:
5782 * 1. Never send more than maxDgramPackets in a jumbogram.
5783 * 2. Never send a packet with more than two iovecs in a jumbogram.
5784 * 3. Never send a retransmitted packet in a jumbogram.
5785 * 4. Never send more than cwind/4 packets in a jumbogram
5786 * We always keep the last list we should have sent so we
5787 * can set the RX_MORE_PACKETS flags correctly.
5791 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5796 struct xmitlist working;
5797 struct xmitlist last;
5799 struct rx_peer *peer = call->conn->peer;
5800 int morePackets = 0;
5802 memset(&last, 0, sizeof(struct xmitlist));
5803 working.list = &list[0];
5805 working.resending = 0;
5807 recovery = call->flags & RX_CALL_FAST_RECOVER;
5809 for (i = 0; i < len; i++) {
5810 /* Does the current packet force us to flush the current list? */
5812 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5813 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5815 /* This sends the 'last' list and then rolls the current working
5816 * set into the 'last' one, and resets the working set */
5819 rxi_SendList(call, &last, istack, 1);
5820 /* If the call enters an error state stop sending, or if
5821 * we entered congestion recovery mode, stop sending */
5823 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5828 working.resending = 0;
5829 working.list = &list[i];
5831 /* Add the current packet to the list if it hasn't been acked.
5832 * Otherwise adjust the list pointer to skip the current packet. */
5833 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5836 if (list[i]->header.serial)
5837 working.resending = 1;
5839 /* Do we need to flush the list? */
5840 if (working.len >= (int)peer->maxDgramPackets
5841 || working.len >= (int)call->nDgramPackets
5842 || working.len >= (int)call->cwind
5843 || list[i]->header.serial
5844 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5846 rxi_SendList(call, &last, istack, 1);
5847 /* If the call enters an error state stop sending, or if
5848 * we entered congestion recovery mode, stop sending */
5850 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5855 working.resending = 0;
5856 working.list = &list[i + 1];
5859 if (working.len != 0) {
5860 osi_Panic("rxi_SendList error");
5862 working.list = &list[i + 1];
5866 /* Send the whole list when the call is in receive mode, when
5867 * the call is in eof mode, when we are in fast recovery mode,
5868 * and when we have the last packet */
5869 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5870 * the listener or event threads
5872 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5873 || (call->flags & RX_CALL_FLUSH)
5874 || (call->flags & RX_CALL_FAST_RECOVER)) {
5875 /* Check for the case where the current list contains
5876 * an acked packet. Since we always send retransmissions
5877 * in a separate packet, we only need to check the first
5878 * packet in the list */
5879 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5883 rxi_SendList(call, &last, istack, morePackets);
5884 /* If the call enters an error state stop sending, or if
5885 * we entered congestion recovery mode, stop sending */
5887 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5891 rxi_SendList(call, &working, istack, 0);
5893 } else if (last.len > 0) {
5894 rxi_SendList(call, &last, istack, 0);
5895 /* Packets which are in 'working' are not sent by this call */
5900 * Check if the peer for the given call is known to be dead
5902 * If the call's peer appears dead (it has encountered fatal network errors
5903 * since the call started) the call is killed with RX_CALL_DEAD if the call
5904 * is active. Otherwise, we do nothing.
5906 * @param[in] call The call to check
5909 * @retval 0 The call is fine, and we haven't done anything to the call
5910 * @retval nonzero The call's peer appears dead, and the call has been
5911 * terminated if it was active
5913 * @pre call->lock must be locked
5916 rxi_CheckPeerDead(struct rx_call *call)
5918 #ifdef AFS_RXERRQ_ENV
5921 if (call->state == RX_STATE_DALLY) {
5925 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5926 if (call->neterr_gen < peererrs) {
5927 /* we have received network errors since this call started; kill
5929 if (call->state == RX_STATE_ACTIVE) {
5930 rxi_CallError(call, RX_CALL_DEAD);
5934 if (call->neterr_gen > peererrs) {
5935 /* someone has reset the number of peer errors; set the call error gen
5936 * so we can detect if more errors are encountered */
5937 call->neterr_gen = peererrs;
5944 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5946 struct rx_call *call = arg0;
5947 struct rx_peer *peer;
5948 struct opr_queue *cursor;
5949 struct clock maxTimeout = { 60, 0 };
5951 MUTEX_ENTER(&call->lock);
5953 peer = call->conn->peer;
5955 /* Make sure that the event pointer is removed from the call
5956 * structure, since there is no longer a per-call retransmission
5958 if (event == call->resendEvent) {
5959 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5960 rxevent_Put(call->resendEvent);
5961 call->resendEvent = NULL;
5964 rxi_CheckPeerDead(call);
5966 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5967 rxi_CheckBusy(call);
5970 if (opr_queue_IsEmpty(&call->tq)) {
5971 /* Nothing to do. This means that we've been raced, and that an
5972 * ACK has come in between when we were triggered, and when we
5973 * actually got to run. */
5977 /* We're in loss recovery */
5978 call->flags |= RX_CALL_FAST_RECOVER;
5980 /* Mark all of the pending packets in the queue as being lost */
5981 for (opr_queue_Scan(&call->tq, cursor)) {
5982 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
5983 if (!(p->flags & RX_PKTFLAG_ACKED))
5984 p->flags &= ~RX_PKTFLAG_SENT;
5987 /* We're resending, so we double the timeout of the call. This will be
5988 * dropped back down by the first successful ACK that we receive.
5990 * We apply a maximum value here of 60 seconds
5992 clock_Add(&call->rto, &call->rto);
5993 if (clock_Gt(&call->rto, &maxTimeout))
5994 call->rto = maxTimeout;
5996 /* Packet loss is most likely due to congestion, so drop our window size
5997 * and start again from the beginning */
5998 if (peer->maxDgramPackets >1) {
5999 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6000 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6002 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6003 call->nDgramPackets = 1;
6005 call->nextCwind = 1;
6008 MUTEX_ENTER(&peer->peer_lock);
6009 peer->MTU = call->MTU;
6010 peer->cwind = call->cwind;
6011 peer->nDgramPackets = 1;
6013 call->congestSeq = peer->congestSeq;
6014 MUTEX_EXIT(&peer->peer_lock);
6016 rxi_Start(call, istack);
6019 MUTEX_EXIT(&call->lock);
6022 /* This routine is called when new packets are readied for
6023 * transmission and when retransmission may be necessary, or when the
6024 * transmission window or burst count are favourable. This should be
6025 * better optimized for new packets, the usual case, now that we've
6026 * got rid of queues of send packets. XXXXXXXXXXX */
6028 rxi_Start(struct rx_call *call, int istack)
6030 struct opr_queue *cursor;
6031 #ifdef RX_ENABLE_LOCKS
6032 struct opr_queue *store;
6038 #ifdef RX_ENABLE_LOCKS
6039 if (rx_stats_active)
6040 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6045 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6046 /* Send (or resend) any packets that need it, subject to
6047 * window restrictions and congestion burst control
6048 * restrictions. Ask for an ack on the last packet sent in
6049 * this burst. For now, we're relying upon the window being
6050 * considerably bigger than the largest number of packets that
6051 * are typically sent at once by one initial call to
6052 * rxi_Start. This is probably bogus (perhaps we should ask
6053 * for an ack when we're half way through the current
6054 * window?). Also, for non file transfer applications, this
6055 * may end up asking for an ack for every packet. Bogus. XXXX
6058 * But check whether we're here recursively, and let the other guy
6061 #ifdef RX_ENABLE_LOCKS
6062 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6063 call->flags |= RX_CALL_TQ_BUSY;
6065 #endif /* RX_ENABLE_LOCKS */
6067 #ifdef RX_ENABLE_LOCKS
6068 call->flags &= ~RX_CALL_NEED_START;
6069 #endif /* RX_ENABLE_LOCKS */
6071 maxXmitPackets = MIN(call->twind, call->cwind);
6072 for (opr_queue_Scan(&call->tq, cursor)) {
6074 = opr_queue_Entry(cursor, struct rx_packet, entry);
6076 if (p->flags & RX_PKTFLAG_ACKED) {
6077 /* Since we may block, don't trust this */
6078 if (rx_stats_active)
6079 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6080 continue; /* Ignore this packet if it has been acknowledged */
6083 /* Turn off all flags except these ones, which are the same
6084 * on each transmission */
6085 p->header.flags &= RX_PRESET_FLAGS;
6087 if (p->header.seq >=
6088 call->tfirst + MIN((int)call->twind,
6089 (int)(call->nSoftAcked +
6091 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6092 /* Note: if we're waiting for more window space, we can
6093 * still send retransmits; hence we don't return here, but
6094 * break out to schedule a retransmit event */
6095 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6096 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6101 /* Transmit the packet if it needs to be sent. */
6102 if (!(p->flags & RX_PKTFLAG_SENT)) {
6103 if (nXmitPackets == maxXmitPackets) {
6104 rxi_SendXmitList(call, call->xmitList,
6105 nXmitPackets, istack);
6108 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6109 *(call->callNumber), p));
6110 call->xmitList[nXmitPackets++] = p;
6112 } /* end of the queue_Scan */
6114 /* xmitList now hold pointers to all of the packets that are
6115 * ready to send. Now we loop to send the packets */
6116 if (nXmitPackets > 0) {
6117 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6121 #ifdef RX_ENABLE_LOCKS
6123 /* We went into the error state while sending packets. Now is
6124 * the time to reset the call. This will also inform the using
6125 * process that the call is in an error state.
6127 if (rx_stats_active)
6128 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6129 call->flags &= ~RX_CALL_TQ_BUSY;
6130 rxi_WakeUpTransmitQueue(call);
6131 rxi_CallError(call, call->error);
6135 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6137 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6138 /* Some packets have received acks. If they all have, we can clear
6139 * the transmit queue.
6142 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6144 = opr_queue_Entry(cursor, struct rx_packet, entry);
6146 if (p->header.seq < call->tfirst
6147 && (p->flags & RX_PKTFLAG_ACKED)) {
6148 opr_queue_Remove(&p->entry);
6149 #ifdef RX_TRACK_PACKETS
6150 p->flags &= ~RX_PKTFLAG_TQ;
6152 #ifdef RXDEBUG_PACKET
6160 call->flags |= RX_CALL_TQ_CLEARME;
6162 if (call->flags & RX_CALL_TQ_CLEARME)
6163 rxi_ClearTransmitQueue(call, 1);
6164 } while (call->flags & RX_CALL_NEED_START);
6166 * TQ references no longer protected by this flag; they must remain
6167 * protected by the call lock.
6169 call->flags &= ~RX_CALL_TQ_BUSY;
6170 rxi_WakeUpTransmitQueue(call);
6172 call->flags |= RX_CALL_NEED_START;
6174 #endif /* RX_ENABLE_LOCKS */
6176 rxi_rto_cancel(call);
6180 /* Also adjusts the keep alive parameters for the call, to reflect
6181 * that we have just sent a packet (so keep alives aren't sent
6184 rxi_Send(struct rx_call *call, struct rx_packet *p,
6187 struct rx_connection *conn = call->conn;
6189 /* Stamp each packet with the user supplied status */
6190 p->header.userStatus = call->localStatus;
6192 /* Allow the security object controlling this call's security to
6193 * make any last-minute changes to the packet */
6194 RXS_SendPacket(conn->securityObject, call, p);
6196 /* Since we're about to send SOME sort of packet to the peer, it's
6197 * safe to nuke any scheduled end-of-packets ack */
6198 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6200 /* Actually send the packet, filling in more connection-specific fields */
6201 MUTEX_EXIT(&call->lock);
6202 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6203 rxi_SendPacket(call, conn, p, istack);
6204 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6205 MUTEX_ENTER(&call->lock);
6207 /* Update last send time for this call (for keep-alive
6208 * processing), and for the connection (so that we can discover
6209 * idle connections) */
6210 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6211 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6212 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6214 conn->lastSendTime = call->lastSendTime = clock_Sec();
6215 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6216 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6217 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6218 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6219 RX_ACK_PING_RESPONSE)))
6220 call->lastSendData = call->lastSendTime;
6224 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6225 * that things are fine. Also called periodically to guarantee that nothing
6226 * falls through the cracks (e.g. (error + dally) connections have keepalive
6227 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6229 * haveCTLock Set if calling from rxi_ReapConnections
6232 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6234 struct rx_connection *conn = call->conn;
6236 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6237 afs_uint32 fudgeFactor;
6240 int idle_timeout = 0;
6241 afs_int32 clock_diff = 0;
6243 if (rxi_CheckPeerDead(call)) {
6249 /* Large swings in the clock can have a significant impact on
6250 * the performance of RX call processing. Forward clock shifts
6251 * will result in premature event triggering or timeouts.
6252 * Backward shifts can result in calls not completing until
6253 * the clock catches up with the original start clock value.
6255 * If a backward clock shift of more than five minutes is noticed,
6256 * just fail the call.
6258 if (now < call->lastSendTime)
6259 clock_diff = call->lastSendTime - now;
6260 if (now < call->startWait)
6261 clock_diff = MAX(clock_diff, call->startWait - now);
6262 if (now < call->lastReceiveTime)
6263 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6264 if (clock_diff > 5 * 60)
6266 if (call->state == RX_STATE_ACTIVE)
6267 rxi_CallError(call, RX_CALL_TIMEOUT);
6271 #ifdef RX_ENABLE_LOCKS
6272 if (call->flags & RX_CALL_TQ_BUSY) {
6273 /* Call is active and will be reset by rxi_Start if it's
6274 * in an error state.
6279 /* RTT + 8*MDEV, rounded up to the next second. */
6280 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6281 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6283 deadTime = conn->secondsUntilDead + fudgeFactor;
6284 /* These are computed to the second (+- 1 second). But that's
6285 * good enough for these values, which should be a significant
6286 * number of seconds. */
6287 if (now > (call->lastReceiveTime + deadTime)) {
6288 if (call->state == RX_STATE_ACTIVE) {
6289 #ifdef AFS_ADAPT_PMTU
6290 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6292 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6293 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6294 ip_stack_t *ipst = ns->netstack_ip;
6296 ire = ire_cache_lookup(conn->peer->host
6297 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6299 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6301 # if defined(GLOBAL_NETSTACKID)
6308 if (ire && ire->ire_max_frag > 0)
6309 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6311 # if defined(GLOBAL_NETSTACKID)
6315 #endif /* AFS_ADAPT_PMTU */
6316 cerror = RX_CALL_DEAD;
6319 #ifdef RX_ENABLE_LOCKS
6320 /* Cancel pending events */
6321 rxevent_Cancel(&call->delayedAckEvent, call,
6322 RX_CALL_REFCOUNT_DELAY);
6323 rxi_rto_cancel(call);
6324 rxevent_Cancel(&call->keepAliveEvent, call,
6325 RX_CALL_REFCOUNT_ALIVE);
6326 rxevent_Cancel(&call->growMTUEvent, call,
6327 RX_CALL_REFCOUNT_MTU);
6328 MUTEX_ENTER(&rx_refcnt_mutex);
6329 /* if rxi_FreeCall returns 1 it has freed the call */
6330 if (call->refCount == 0 &&
6331 rxi_FreeCall(call, haveCTLock))
6333 MUTEX_EXIT(&rx_refcnt_mutex);
6336 MUTEX_EXIT(&rx_refcnt_mutex);
6338 #else /* RX_ENABLE_LOCKS */
6339 rxi_FreeCall(call, 0);
6341 #endif /* RX_ENABLE_LOCKS */
6343 /* Non-active calls are destroyed if they are not responding
6344 * to pings; active calls are simply flagged in error, so the
6345 * attached process can die reasonably gracefully. */
6348 if (conn->idleDeadDetection) {
6349 if (conn->idleDeadTime) {
6350 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6354 /* see if we have a non-activity timeout */
6355 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6356 (call->flags & RX_CALL_READER_WAIT)) {
6357 if (call->state == RX_STATE_ACTIVE) {
6358 cerror = RX_CALL_TIMEOUT;
6363 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6364 if (call->state == RX_STATE_ACTIVE) {
6365 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6373 if (conn->hardDeadTime) {
6374 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6377 /* see if we have a hard timeout */
6379 && (now > (hardDeadTime + call->startTime.sec))) {
6380 if (call->state == RX_STATE_ACTIVE)
6381 rxi_CallError(call, RX_CALL_TIMEOUT);
6386 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6387 call->lastReceiveTime) {
6388 int oldMTU = conn->peer->ifMTU;
6390 /* if we thought we could send more, perhaps things got worse */
6391 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6392 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6393 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6394 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6396 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6398 /* minimum capped in SetPeerMtu */
6399 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6402 conn->lastPacketSize = 0;
6404 /* needed so ResetCall doesn't clobber us. */
6405 call->MTU = conn->peer->ifMTU;
6407 /* if we never succeeded, let the error pass out as-is */
6408 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6409 cerror = conn->msgsizeRetryErr;
6412 rxi_CallError(call, cerror);
6417 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6418 void *dummy, int dummy2)
6420 struct rx_connection *conn = arg1;
6421 struct rx_header theader;
6422 char tbuffer[1 + sizeof(struct rx_header)];
6423 struct sockaddr_in taddr;
6426 struct iovec tmpiov[2];
6429 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6432 tp = &tbuffer[sizeof(struct rx_header)];
6433 taddr.sin_family = AF_INET;
6434 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6435 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6436 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6437 taddr.sin_len = sizeof(struct sockaddr_in);
6439 memset(&theader, 0, sizeof(theader));
6440 theader.epoch = htonl(999);
6442 theader.callNumber = 0;
6445 theader.type = RX_PACKET_TYPE_VERSION;
6446 theader.flags = RX_LAST_PACKET;
6447 theader.serviceId = 0;
6449 memcpy(tbuffer, &theader, sizeof(theader));
6450 memcpy(tp, &a, sizeof(a));
6451 tmpiov[0].iov_base = tbuffer;
6452 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6454 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6456 MUTEX_ENTER(&conn->conn_data_lock);
6457 MUTEX_ENTER(&rx_refcnt_mutex);
6458 /* Only reschedule ourselves if the connection would not be destroyed */
6459 if (conn->refCount <= 1) {
6460 rxevent_Put(conn->natKeepAliveEvent);
6461 conn->natKeepAliveEvent = NULL;
6462 MUTEX_EXIT(&rx_refcnt_mutex);
6463 MUTEX_EXIT(&conn->conn_data_lock);
6464 rx_DestroyConnection(conn); /* drop the reference for this */
6466 conn->refCount--; /* drop the reference for this */
6467 MUTEX_EXIT(&rx_refcnt_mutex);
6468 rxevent_Put(conn->natKeepAliveEvent);
6469 conn->natKeepAliveEvent = NULL;
6470 rxi_ScheduleNatKeepAliveEvent(conn);
6471 MUTEX_EXIT(&conn->conn_data_lock);
6476 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6478 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6479 struct clock when, now;
6480 clock_GetTime(&now);
6482 when.sec += conn->secondsUntilNatPing;
6483 MUTEX_ENTER(&rx_refcnt_mutex);
6484 conn->refCount++; /* hold a reference for this */
6485 MUTEX_EXIT(&rx_refcnt_mutex);
6486 conn->natKeepAliveEvent =
6487 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6492 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6494 MUTEX_ENTER(&conn->conn_data_lock);
6495 conn->secondsUntilNatPing = seconds;
6497 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6498 rxi_ScheduleNatKeepAliveEvent(conn);
6500 conn->flags |= RX_CONN_NAT_PING;
6502 MUTEX_EXIT(&conn->conn_data_lock);
6505 /* When a call is in progress, this routine is called occasionally to
6506 * make sure that some traffic has arrived (or been sent to) the peer.
6507 * If nothing has arrived in a reasonable amount of time, the call is
6508 * declared dead; if nothing has been sent for a while, we send a
6509 * keep-alive packet (if we're actually trying to keep the call alive)
6512 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6515 struct rx_call *call = arg1;
6516 struct rx_connection *conn;
6519 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6520 MUTEX_ENTER(&call->lock);
6522 if (event == call->keepAliveEvent) {
6523 rxevent_Put(call->keepAliveEvent);
6524 call->keepAliveEvent = NULL;
6529 if (rxi_CheckCall(call, 0)) {
6530 MUTEX_EXIT(&call->lock);
6534 /* Don't try to keep alive dallying calls */
6535 if (call->state == RX_STATE_DALLY) {
6536 MUTEX_EXIT(&call->lock);
6541 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6542 /* Don't try to send keepalives if there is unacknowledged data */
6543 /* the rexmit code should be good enough, this little hack
6544 * doesn't quite work XXX */
6545 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6547 rxi_ScheduleKeepAliveEvent(call);
6548 MUTEX_EXIT(&call->lock);
6551 /* Does what's on the nameplate. */
6553 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6555 struct rx_call *call = arg1;
6556 struct rx_connection *conn;
6558 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6559 MUTEX_ENTER(&call->lock);
6561 if (event == call->growMTUEvent) {
6562 rxevent_Put(call->growMTUEvent);
6563 call->growMTUEvent = NULL;
6566 if (rxi_CheckCall(call, 0)) {
6567 MUTEX_EXIT(&call->lock);
6571 /* Don't bother with dallying calls */
6572 if (call->state == RX_STATE_DALLY) {
6573 MUTEX_EXIT(&call->lock);
6580 * keep being scheduled, just don't do anything if we're at peak,
6581 * or we're not set up to be properly handled (idle timeout required)
6583 if ((conn->peer->maxPacketSize != 0) &&
6584 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6585 conn->idleDeadDetection)
6586 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6587 rxi_ScheduleGrowMTUEvent(call, 0);
6588 MUTEX_EXIT(&call->lock);
6592 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6594 if (!call->keepAliveEvent) {
6595 struct clock when, now;
6596 clock_GetTime(&now);
6598 when.sec += call->conn->secondsUntilPing;
6599 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6600 call->keepAliveEvent =
6601 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6606 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6608 if (!call->growMTUEvent) {
6609 struct clock when, now;
6611 clock_GetTime(&now);
6614 if (call->conn->secondsUntilPing)
6615 secs = (6*call->conn->secondsUntilPing)-1;
6617 if (call->conn->secondsUntilDead)
6618 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6622 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6623 call->growMTUEvent =
6624 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6628 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6630 rxi_KeepAliveOn(struct rx_call *call)
6632 /* Pretend last packet received was received now--i.e. if another
6633 * packet isn't received within the keep alive time, then the call
6634 * will die; Initialize last send time to the current time--even
6635 * if a packet hasn't been sent yet. This will guarantee that a
6636 * keep-alive is sent within the ping time */
6637 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6638 rxi_ScheduleKeepAliveEvent(call);
6642 * Solely in order that callers not need to include rx_call.h
6645 rx_KeepAliveOff(struct rx_call *call)
6647 MUTEX_ENTER(&call->lock);
6648 rxi_KeepAliveOff(call);
6649 MUTEX_EXIT(&call->lock);
6652 rx_KeepAliveOn(struct rx_call *call)
6654 MUTEX_ENTER(&call->lock);
6655 rxi_KeepAliveOn(call);
6656 MUTEX_EXIT(&call->lock);
6660 rxi_GrowMTUOn(struct rx_call *call)
6662 struct rx_connection *conn = call->conn;
6663 MUTEX_ENTER(&conn->conn_data_lock);
6664 conn->lastPingSizeSer = conn->lastPingSize = 0;
6665 MUTEX_EXIT(&conn->conn_data_lock);
6666 rxi_ScheduleGrowMTUEvent(call, 1);
6669 /* This routine is called to send connection abort messages
6670 * that have been delayed to throttle looping clients. */
6672 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6675 struct rx_connection *conn = arg1;
6678 struct rx_packet *packet;
6680 MUTEX_ENTER(&conn->conn_data_lock);
6681 rxevent_Put(conn->delayedAbortEvent);
6682 conn->delayedAbortEvent = NULL;
6683 error = htonl(conn->error);
6685 MUTEX_EXIT(&conn->conn_data_lock);
6686 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6689 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6690 RX_PACKET_TYPE_ABORT, (char *)&error,
6692 rxi_FreePacket(packet);
6696 /* This routine is called to send call abort messages
6697 * that have been delayed to throttle looping clients. */
6699 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6702 struct rx_call *call = arg1;
6705 struct rx_packet *packet;
6707 MUTEX_ENTER(&call->lock);
6708 rxevent_Put(call->delayedAbortEvent);
6709 call->delayedAbortEvent = NULL;
6710 error = htonl(call->error);
6712 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6715 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6716 (char *)&error, sizeof(error), 0);
6717 rxi_FreePacket(packet);
6719 MUTEX_EXIT(&call->lock);
6720 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6723 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6724 * seconds) to ask the client to authenticate itself. The routine
6725 * issues a challenge to the client, which is obtained from the
6726 * security object associated with the connection */
6728 rxi_ChallengeEvent(struct rxevent *event,
6729 void *arg0, void *arg1, int tries)
6731 struct rx_connection *conn = arg0;
6734 rxevent_Put(conn->challengeEvent);
6735 conn->challengeEvent = NULL;
6738 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6739 struct rx_packet *packet;
6740 struct clock when, now;
6743 /* We've failed to authenticate for too long.
6744 * Reset any calls waiting for authentication;
6745 * they are all in RX_STATE_PRECALL.
6749 MUTEX_ENTER(&conn->conn_call_lock);
6750 for (i = 0; i < RX_MAXCALLS; i++) {
6751 struct rx_call *call = conn->call[i];
6753 MUTEX_ENTER(&call->lock);
6754 if (call->state == RX_STATE_PRECALL) {
6755 rxi_CallError(call, RX_CALL_DEAD);
6756 rxi_SendCallAbort(call, NULL, 0, 0);
6758 MUTEX_EXIT(&call->lock);
6761 MUTEX_EXIT(&conn->conn_call_lock);
6765 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6767 /* If there's no packet available, do this later. */
6768 RXS_GetChallenge(conn->securityObject, conn, packet);
6769 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6770 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6771 rxi_FreePacket(packet);
6773 clock_GetTime(&now);
6775 when.sec += RX_CHALLENGE_TIMEOUT;
6776 conn->challengeEvent =
6777 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6782 /* Call this routine to start requesting the client to authenticate
6783 * itself. This will continue until authentication is established,
6784 * the call times out, or an invalid response is returned. The
6785 * security object associated with the connection is asked to create
6786 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6787 * defined earlier. */
6789 rxi_ChallengeOn(struct rx_connection *conn)
6791 if (!conn->challengeEvent) {
6792 RXS_CreateChallenge(conn->securityObject, conn);
6793 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6798 /* rxi_ComputeRoundTripTime is called with peer locked. */
6799 /* peer may be null */
6801 rxi_ComputeRoundTripTime(struct rx_packet *p,
6802 struct rx_ackPacket *ack,
6803 struct rx_call *call,
6804 struct rx_peer *peer,
6807 struct clock thisRtt, *sentp;
6811 /* If the ACK is delayed, then do nothing */
6812 if (ack->reason == RX_ACK_DELAY)
6815 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6816 * their RTT multiple times, so only include the RTT of the last packet
6818 if (p->flags & RX_JUMBO_PACKET)
6821 /* Use the serial number to determine which transmission the ACK is for,
6822 * and set the sent time to match this. If we have no serial number, then
6823 * only use the ACK for RTT calculations if the packet has not been
6827 serial = ntohl(ack->serial);
6829 if (serial == p->header.serial) {
6830 sentp = &p->timeSent;
6831 } else if (serial == p->firstSerial) {
6832 sentp = &p->firstSent;
6833 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6834 sentp = &p->firstSent;
6838 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6839 sentp = &p->firstSent;
6846 if (clock_Lt(&thisRtt, sentp))
6847 return; /* somebody set the clock back, don't count this time. */
6849 clock_Sub(&thisRtt, sentp);
6850 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6851 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6853 if (clock_IsZero(&thisRtt)) {
6855 * The actual round trip time is shorter than the
6856 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6857 * Since we can't tell which at the moment we will assume 1ms.
6859 thisRtt.usec = 1000;
6862 if (rx_stats_active) {
6863 MUTEX_ENTER(&rx_stats_mutex);
6864 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6865 rx_stats.minRtt = thisRtt;
6866 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6867 if (thisRtt.sec > 60) {
6868 MUTEX_EXIT(&rx_stats_mutex);
6869 return; /* somebody set the clock ahead */
6871 rx_stats.maxRtt = thisRtt;
6873 clock_Add(&rx_stats.totalRtt, &thisRtt);
6874 rx_atomic_inc(&rx_stats.nRttSamples);
6875 MUTEX_EXIT(&rx_stats_mutex);
6878 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6880 /* Apply VanJacobson round-trip estimations */
6885 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6886 * srtt is stored as fixed point with 3 bits after the binary
6887 * point (i.e., scaled by 8). The following magic is
6888 * equivalent to the smoothing algorithm in rfc793 with an
6889 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6890 * srtt'*8 = rtt + srtt*7
6891 * srtt'*8 = srtt*8 + rtt - srtt
6892 * srtt' = srtt + rtt/8 - srtt/8
6893 * srtt' = srtt + (rtt - srtt)/8
6896 delta = _8THMSEC(&thisRtt) - call->rtt;
6897 call->rtt += (delta >> 3);
6900 * We accumulate a smoothed rtt variance (actually, a smoothed
6901 * mean difference), then set the retransmit timer to smoothed
6902 * rtt + 4 times the smoothed variance (was 2x in van's original
6903 * paper, but 4x works better for me, and apparently for him as
6905 * rttvar is stored as
6906 * fixed point with 2 bits after the binary point (scaled by
6907 * 4). The following is equivalent to rfc793 smoothing with
6908 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6909 * rttvar'*4 = rttvar*3 + |delta|
6910 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6911 * rttvar' = rttvar + |delta|/4 - rttvar/4
6912 * rttvar' = rttvar + (|delta| - rttvar)/4
6913 * This replaces rfc793's wired-in beta.
6914 * dev*4 = dev*4 + (|actual - expected| - dev)
6920 delta -= (call->rtt_dev << 1);
6921 call->rtt_dev += (delta >> 3);
6923 /* I don't have a stored RTT so I start with this value. Since I'm
6924 * probably just starting a call, and will be pushing more data down
6925 * this, I expect congestion to increase rapidly. So I fudge a
6926 * little, and I set deviance to half the rtt. In practice,
6927 * deviance tends to approach something a little less than
6928 * half the smoothed rtt. */
6929 call->rtt = _8THMSEC(&thisRtt) + 8;
6930 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6932 /* the smoothed RTT time is RTT + 4*MDEV
6934 * We allow a user specified minimum to be set for this, to allow clamping
6935 * at a minimum value in the same way as TCP. In addition, we have to allow
6936 * for the possibility that this packet is answered by a delayed ACK, so we
6937 * add on a fixed 200ms to account for that timer expiring.
6940 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6941 rx_minPeerTimeout) + 200;
6942 clock_Zero(&call->rto);
6943 clock_Addmsec(&call->rto, rtt_timeout);
6945 /* Update the peer, so any new calls start with our values */
6946 peer->rtt_dev = call->rtt_dev;
6947 peer->rtt = call->rtt;
6949 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6950 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6954 /* Find all server connections that have not been active for a long time, and
6957 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6960 struct clock now, when;
6961 clock_GetTime(&now);
6963 /* Find server connection structures that haven't been used for
6964 * greater than rx_idleConnectionTime */
6966 struct rx_connection **conn_ptr, **conn_end;
6967 int i, havecalls = 0;
6968 MUTEX_ENTER(&rx_connHashTable_lock);
6969 for (conn_ptr = &rx_connHashTable[0], conn_end =
6970 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6972 struct rx_connection *conn, *next;
6973 struct rx_call *call;
6977 for (conn = *conn_ptr; conn; conn = next) {
6978 /* XXX -- Shouldn't the connection be locked? */
6981 for (i = 0; i < RX_MAXCALLS; i++) {
6982 call = conn->call[i];
6986 code = MUTEX_TRYENTER(&call->lock);
6989 result = rxi_CheckCall(call, 1);
6990 MUTEX_EXIT(&call->lock);
6992 /* If CheckCall freed the call, it might
6993 * have destroyed the connection as well,
6994 * which screws up the linked lists.
7000 if (conn->type == RX_SERVER_CONNECTION) {
7001 /* This only actually destroys the connection if
7002 * there are no outstanding calls */
7003 MUTEX_ENTER(&conn->conn_data_lock);
7004 MUTEX_ENTER(&rx_refcnt_mutex);
7005 if (!havecalls && !conn->refCount
7006 && ((conn->lastSendTime + rx_idleConnectionTime) <
7008 conn->refCount++; /* it will be decr in rx_DestroyConn */
7009 MUTEX_EXIT(&rx_refcnt_mutex);
7010 MUTEX_EXIT(&conn->conn_data_lock);
7011 #ifdef RX_ENABLE_LOCKS
7012 rxi_DestroyConnectionNoLock(conn);
7013 #else /* RX_ENABLE_LOCKS */
7014 rxi_DestroyConnection(conn);
7015 #endif /* RX_ENABLE_LOCKS */
7017 #ifdef RX_ENABLE_LOCKS
7019 MUTEX_EXIT(&rx_refcnt_mutex);
7020 MUTEX_EXIT(&conn->conn_data_lock);
7022 #endif /* RX_ENABLE_LOCKS */
7026 #ifdef RX_ENABLE_LOCKS
7027 while (rx_connCleanup_list) {
7028 struct rx_connection *conn;
7029 conn = rx_connCleanup_list;
7030 rx_connCleanup_list = rx_connCleanup_list->next;
7031 MUTEX_EXIT(&rx_connHashTable_lock);
7032 rxi_CleanupConnection(conn);
7033 MUTEX_ENTER(&rx_connHashTable_lock);
7035 MUTEX_EXIT(&rx_connHashTable_lock);
7036 #endif /* RX_ENABLE_LOCKS */
7039 /* Find any peer structures that haven't been used (haven't had an
7040 * associated connection) for greater than rx_idlePeerTime */
7042 struct rx_peer **peer_ptr, **peer_end;
7046 * Why do we need to hold the rx_peerHashTable_lock across
7047 * the incrementing of peer_ptr since the rx_peerHashTable
7048 * array is not changing? We don't.
7050 * By dropping the lock periodically we can permit other
7051 * activities to be performed while a rxi_ReapConnections
7052 * call is in progress. The goal of reap connections
7053 * is to clean up quickly without causing large amounts
7054 * of contention. Therefore, it is important that global
7055 * mutexes not be held for extended periods of time.
7057 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7058 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7060 struct rx_peer *peer, *next, *prev;
7062 MUTEX_ENTER(&rx_peerHashTable_lock);
7063 for (prev = peer = *peer_ptr; peer; peer = next) {
7065 code = MUTEX_TRYENTER(&peer->peer_lock);
7066 if ((code) && (peer->refCount == 0)
7067 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7068 struct opr_queue *cursor, *store;
7072 * now know that this peer object is one to be
7073 * removed from the hash table. Once it is removed
7074 * it can't be referenced by other threads.
7075 * Lets remove it first and decrement the struct
7076 * nPeerStructs count.
7078 if (peer == *peer_ptr) {
7084 if (rx_stats_active)
7085 rx_atomic_dec(&rx_stats.nPeerStructs);
7088 * Now if we hold references on 'prev' and 'next'
7089 * we can safely drop the rx_peerHashTable_lock
7090 * while we destroy this 'peer' object.
7096 MUTEX_EXIT(&rx_peerHashTable_lock);
7098 MUTEX_EXIT(&peer->peer_lock);
7099 MUTEX_DESTROY(&peer->peer_lock);
7101 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7102 unsigned int num_funcs;
7103 struct rx_interface_stat *rpc_stat
7104 = opr_queue_Entry(cursor, struct rx_interface_stat,
7109 opr_queue_Remove(&rpc_stat->entry);
7110 opr_queue_Remove(&rpc_stat->entryPeers);
7112 num_funcs = rpc_stat->stats[0].func_total;
7114 sizeof(rx_interface_stat_t) +
7115 rpc_stat->stats[0].func_total *
7116 sizeof(rx_function_entry_v1_t);
7118 rxi_Free(rpc_stat, space);
7120 MUTEX_ENTER(&rx_rpc_stats);
7121 rxi_rpc_peer_stat_cnt -= num_funcs;
7122 MUTEX_EXIT(&rx_rpc_stats);
7127 * Regain the rx_peerHashTable_lock and
7128 * decrement the reference count on 'prev'
7131 MUTEX_ENTER(&rx_peerHashTable_lock);
7138 MUTEX_EXIT(&peer->peer_lock);
7143 MUTEX_EXIT(&rx_peerHashTable_lock);
7147 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7148 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7149 * GC, just below. Really, we shouldn't have to keep moving packets from
7150 * one place to another, but instead ought to always know if we can
7151 * afford to hold onto a packet in its particular use. */
7152 MUTEX_ENTER(&rx_freePktQ_lock);
7153 if (rx_waitingForPackets) {
7154 rx_waitingForPackets = 0;
7155 #ifdef RX_ENABLE_LOCKS
7156 CV_BROADCAST(&rx_waitingForPackets_cv);
7158 osi_rxWakeup(&rx_waitingForPackets);
7161 MUTEX_EXIT(&rx_freePktQ_lock);
7164 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7165 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7169 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7170 * rx.h is sort of strange this is better. This is called with a security
7171 * object before it is discarded. Each connection using a security object has
7172 * its own refcount to the object so it won't actually be freed until the last
7173 * connection is destroyed.
7175 * This is the only rxs module call. A hold could also be written but no one
7179 rxs_Release(struct rx_securityClass *aobj)
7181 return RXS_Close(aobj);
7189 #define TRACE_OPTION_RX_DEBUG 16
7197 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7198 0, KEY_QUERY_VALUE, &parmKey);
7199 if (code != ERROR_SUCCESS)
7202 dummyLen = sizeof(TraceOption);
7203 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7204 (BYTE *) &TraceOption, &dummyLen);
7205 if (code == ERROR_SUCCESS) {
7206 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7208 RegCloseKey (parmKey);
7209 #endif /* AFS_NT40_ENV */
7214 rx_DebugOnOff(int on)
7218 rxdebug_active = on;
7224 rx_StatsOnOff(int on)
7226 rx_stats_active = on;
7230 /* Don't call this debugging routine directly; use dpf */
7232 rxi_DebugPrint(char *format, ...)
7241 va_start(ap, format);
7243 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7246 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7248 OutputDebugString(msg);
7254 va_start(ap, format);
7256 clock_GetTime(&now);
7257 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7258 (unsigned int)now.usec);
7259 vfprintf(rx_Log, format, ap);
7267 * This function is used to process the rx_stats structure that is local
7268 * to a process as well as an rx_stats structure received from a remote
7269 * process (via rxdebug). Therefore, it needs to do minimal version
7273 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7274 afs_int32 freePackets, char version)
7278 if (size != sizeof(struct rx_statistics)) {
7280 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7281 size, sizeof(struct rx_statistics));
7284 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7287 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7288 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7289 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7290 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7291 s->specialPktAllocFailures);
7293 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7294 s->receivePktAllocFailures, s->sendPktAllocFailures,
7295 s->specialPktAllocFailures);
7299 " greedy %u, " "bogusReads %u (last from host %x), "
7300 "noPackets %u, " "noBuffers %u, " "selects %u, "
7301 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7302 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7303 s->selects, s->sendSelects);
7305 fprintf(file, " packets read: ");
7306 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7307 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7309 fprintf(file, "\n");
7312 " other read counters: data %u, " "ack %u, " "dup %u "
7313 "spurious %u " "dally %u\n", s->dataPacketsRead,
7314 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7315 s->ignorePacketDally);
7317 fprintf(file, " packets sent: ");
7318 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7319 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7321 fprintf(file, "\n");
7324 " other send counters: ack %u, " "data %u (not resends), "
7325 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7326 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7327 s->dataPacketsPushed, s->ignoreAckedPacket);
7330 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7331 s->netSendFailures, (int)s->fatalErrors);
7333 if (s->nRttSamples) {
7334 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7335 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7337 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7338 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7342 " %d server connections, " "%d client connections, "
7343 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7344 s->nServerConns, s->nClientConns, s->nPeerStructs,
7345 s->nCallStructs, s->nFreeCallStructs);
7347 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7348 fprintf(file, " %d clock updates\n", clock_nUpdates);
7352 /* for backward compatibility */
7354 rx_PrintStats(FILE * file)
7356 MUTEX_ENTER(&rx_stats_mutex);
7357 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7358 sizeof(rx_stats), rx_nFreePackets,
7360 MUTEX_EXIT(&rx_stats_mutex);
7364 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7366 fprintf(file, "Peer %x.%d.\n",
7367 ntohl(peer->host), (int)ntohs(peer->port));
7370 " Rtt %d, " "total sent %d, " "resent %d\n",
7371 peer->rtt, peer->nSent, peer->reSends);
7373 fprintf(file, " Packet size %d\n", peer->ifMTU);
7377 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7379 * This mutex protects the following static variables:
7383 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7384 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7386 #define LOCK_RX_DEBUG
7387 #define UNLOCK_RX_DEBUG
7388 #endif /* AFS_PTHREAD_ENV */
7390 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7392 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7393 u_char type, void *inputData, size_t inputLength,
7394 void *outputData, size_t outputLength)
7396 static afs_int32 counter = 100;
7397 time_t waitTime, waitCount;
7398 struct rx_header theader;
7401 struct timeval tv_now, tv_wake, tv_delta;
7402 struct sockaddr_in taddr, faddr;
7416 tp = &tbuffer[sizeof(struct rx_header)];
7417 taddr.sin_family = AF_INET;
7418 taddr.sin_port = remotePort;
7419 taddr.sin_addr.s_addr = remoteAddr;
7420 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7421 taddr.sin_len = sizeof(struct sockaddr_in);
7424 memset(&theader, 0, sizeof(theader));
7425 theader.epoch = htonl(999);
7427 theader.callNumber = htonl(counter);
7430 theader.type = type;
7431 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7432 theader.serviceId = 0;
7434 memcpy(tbuffer, &theader, sizeof(theader));
7435 memcpy(tp, inputData, inputLength);
7437 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7438 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7440 /* see if there's a packet available */
7441 gettimeofday(&tv_wake, NULL);
7442 tv_wake.tv_sec += waitTime;
7445 FD_SET(socket, &imask);
7446 tv_delta.tv_sec = tv_wake.tv_sec;
7447 tv_delta.tv_usec = tv_wake.tv_usec;
7448 gettimeofday(&tv_now, NULL);
7450 if (tv_delta.tv_usec < tv_now.tv_usec) {
7452 tv_delta.tv_usec += 1000000;
7455 tv_delta.tv_usec -= tv_now.tv_usec;
7457 if (tv_delta.tv_sec < tv_now.tv_sec) {
7461 tv_delta.tv_sec -= tv_now.tv_sec;
7464 code = select(0, &imask, 0, 0, &tv_delta);
7465 #else /* AFS_NT40_ENV */
7466 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7467 #endif /* AFS_NT40_ENV */
7468 if (code == 1 && FD_ISSET(socket, &imask)) {
7469 /* now receive a packet */
7470 faddrLen = sizeof(struct sockaddr_in);
7472 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7473 (struct sockaddr *)&faddr, &faddrLen);
7476 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7477 if (counter == ntohl(theader.callNumber))
7485 /* see if we've timed out */
7493 code -= sizeof(struct rx_header);
7494 if (code > outputLength)
7495 code = outputLength;
7496 memcpy(outputData, tp, code);
7499 #endif /* RXDEBUG */
7502 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7503 afs_uint16 remotePort, struct rx_debugStats * stat,
7504 afs_uint32 * supportedValues)
7506 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7508 struct rx_debugIn in;
7510 *supportedValues = 0;
7511 in.type = htonl(RX_DEBUGI_GETSTATS);
7514 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7515 &in, sizeof(in), stat, sizeof(*stat));
7518 * If the call was successful, fixup the version and indicate
7519 * what contents of the stat structure are valid.
7520 * Also do net to host conversion of fields here.
7524 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7525 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7527 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7528 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7530 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7531 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7533 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7534 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7536 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7537 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7539 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7540 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7542 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7543 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7545 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7546 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7548 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7549 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7551 stat->nFreePackets = ntohl(stat->nFreePackets);
7552 stat->packetReclaims = ntohl(stat->packetReclaims);
7553 stat->callsExecuted = ntohl(stat->callsExecuted);
7554 stat->nWaiting = ntohl(stat->nWaiting);
7555 stat->idleThreads = ntohl(stat->idleThreads);
7556 stat->nWaited = ntohl(stat->nWaited);
7557 stat->nPackets = ntohl(stat->nPackets);
7566 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7567 afs_uint16 remotePort, struct rx_statistics * stat,
7568 afs_uint32 * supportedValues)
7570 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7572 struct rx_debugIn in;
7573 afs_int32 *lp = (afs_int32 *) stat;
7577 * supportedValues is currently unused, but added to allow future
7578 * versioning of this function.
7581 *supportedValues = 0;
7582 in.type = htonl(RX_DEBUGI_RXSTATS);
7584 memset(stat, 0, sizeof(*stat));
7586 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7587 &in, sizeof(in), stat, sizeof(*stat));
7592 * Do net to host conversion here
7595 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7606 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7607 afs_uint16 remotePort, size_t version_length,
7610 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7612 return MakeDebugCall(socket, remoteAddr, remotePort,
7613 RX_PACKET_TYPE_VERSION, a, 1, version,
7621 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7622 afs_uint16 remotePort, afs_int32 * nextConnection,
7623 int allConnections, afs_uint32 debugSupportedValues,
7624 struct rx_debugConn * conn,
7625 afs_uint32 * supportedValues)
7627 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7629 struct rx_debugIn in;
7633 * supportedValues is currently unused, but added to allow future
7634 * versioning of this function.
7637 *supportedValues = 0;
7638 if (allConnections) {
7639 in.type = htonl(RX_DEBUGI_GETALLCONN);
7641 in.type = htonl(RX_DEBUGI_GETCONN);
7643 in.index = htonl(*nextConnection);
7644 memset(conn, 0, sizeof(*conn));
7646 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7647 &in, sizeof(in), conn, sizeof(*conn));
7650 *nextConnection += 1;
7653 * Convert old connection format to new structure.
7656 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7657 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7658 #define MOVEvL(a) (conn->a = vL->a)
7660 /* any old or unrecognized version... */
7661 for (i = 0; i < RX_MAXCALLS; i++) {
7662 MOVEvL(callState[i]);
7663 MOVEvL(callMode[i]);
7664 MOVEvL(callFlags[i]);
7665 MOVEvL(callOther[i]);
7667 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7668 MOVEvL(secStats.type);
7669 MOVEvL(secStats.level);
7670 MOVEvL(secStats.flags);
7671 MOVEvL(secStats.expires);
7672 MOVEvL(secStats.packetsReceived);
7673 MOVEvL(secStats.packetsSent);
7674 MOVEvL(secStats.bytesReceived);
7675 MOVEvL(secStats.bytesSent);
7680 * Do net to host conversion here
7682 * I don't convert host or port since we are most likely
7683 * going to want these in NBO.
7685 conn->cid = ntohl(conn->cid);
7686 conn->serial = ntohl(conn->serial);
7687 for (i = 0; i < RX_MAXCALLS; i++) {
7688 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7690 conn->error = ntohl(conn->error);
7691 conn->secStats.flags = ntohl(conn->secStats.flags);
7692 conn->secStats.expires = ntohl(conn->secStats.expires);
7693 conn->secStats.packetsReceived =
7694 ntohl(conn->secStats.packetsReceived);
7695 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7696 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7697 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7698 conn->epoch = ntohl(conn->epoch);
7699 conn->natMTU = ntohl(conn->natMTU);
7708 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7709 afs_uint16 remotePort, afs_int32 * nextPeer,
7710 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7711 afs_uint32 * supportedValues)
7713 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7715 struct rx_debugIn in;
7718 * supportedValues is currently unused, but added to allow future
7719 * versioning of this function.
7722 *supportedValues = 0;
7723 in.type = htonl(RX_DEBUGI_GETPEER);
7724 in.index = htonl(*nextPeer);
7725 memset(peer, 0, sizeof(*peer));
7727 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7728 &in, sizeof(in), peer, sizeof(*peer));
7734 * Do net to host conversion here
7736 * I don't convert host or port since we are most likely
7737 * going to want these in NBO.
7739 peer->ifMTU = ntohs(peer->ifMTU);
7740 peer->idleWhen = ntohl(peer->idleWhen);
7741 peer->refCount = ntohs(peer->refCount);
7742 peer->rtt = ntohl(peer->rtt);
7743 peer->rtt_dev = ntohl(peer->rtt_dev);
7744 peer->timeout.sec = 0;
7745 peer->timeout.usec = 0;
7746 peer->nSent = ntohl(peer->nSent);
7747 peer->reSends = ntohl(peer->reSends);
7748 peer->natMTU = ntohs(peer->natMTU);
7749 peer->maxMTU = ntohs(peer->maxMTU);
7750 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7751 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7752 peer->MTU = ntohs(peer->MTU);
7753 peer->cwind = ntohs(peer->cwind);
7754 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7755 peer->congestSeq = ntohs(peer->congestSeq);
7756 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7757 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7758 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7759 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7768 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7769 struct rx_debugPeer * peerStats)
7772 afs_int32 error = 1; /* default to "did not succeed" */
7773 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7775 MUTEX_ENTER(&rx_peerHashTable_lock);
7776 for(tp = rx_peerHashTable[hashValue];
7777 tp != NULL; tp = tp->next) {
7778 if (tp->host == peerHost)
7784 MUTEX_EXIT(&rx_peerHashTable_lock);
7788 MUTEX_ENTER(&tp->peer_lock);
7789 peerStats->host = tp->host;
7790 peerStats->port = tp->port;
7791 peerStats->ifMTU = tp->ifMTU;
7792 peerStats->idleWhen = tp->idleWhen;
7793 peerStats->refCount = tp->refCount;
7794 peerStats->burstSize = 0;
7795 peerStats->burst = 0;
7796 peerStats->burstWait.sec = 0;
7797 peerStats->burstWait.usec = 0;
7798 peerStats->rtt = tp->rtt;
7799 peerStats->rtt_dev = tp->rtt_dev;
7800 peerStats->timeout.sec = 0;
7801 peerStats->timeout.usec = 0;
7802 peerStats->nSent = tp->nSent;
7803 peerStats->reSends = tp->reSends;
7804 peerStats->natMTU = tp->natMTU;
7805 peerStats->maxMTU = tp->maxMTU;
7806 peerStats->maxDgramPackets = tp->maxDgramPackets;
7807 peerStats->ifDgramPackets = tp->ifDgramPackets;
7808 peerStats->MTU = tp->MTU;
7809 peerStats->cwind = tp->cwind;
7810 peerStats->nDgramPackets = tp->nDgramPackets;
7811 peerStats->congestSeq = tp->congestSeq;
7812 peerStats->bytesSent.high = tp->bytesSent >> 32;
7813 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7814 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7815 peerStats->bytesReceived.low
7816 = tp->bytesReceived & MAX_AFS_UINT32;
7817 MUTEX_EXIT(&tp->peer_lock);
7819 MUTEX_ENTER(&rx_peerHashTable_lock);
7822 MUTEX_EXIT(&rx_peerHashTable_lock);
7830 struct rx_serverQueueEntry *np;
7833 struct rx_call *call;
7834 struct rx_serverQueueEntry *sq;
7838 if (rxinit_status == 1) {
7840 return; /* Already shutdown. */
7844 #ifndef AFS_PTHREAD_ENV
7845 FD_ZERO(&rx_selectMask);
7846 #endif /* AFS_PTHREAD_ENV */
7847 rxi_dataQuota = RX_MAX_QUOTA;
7848 #ifndef AFS_PTHREAD_ENV
7850 #endif /* AFS_PTHREAD_ENV */
7853 #ifndef AFS_PTHREAD_ENV
7854 #ifndef AFS_USE_GETTIMEOFDAY
7856 #endif /* AFS_USE_GETTIMEOFDAY */
7857 #endif /* AFS_PTHREAD_ENV */
7859 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7860 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7861 opr_queue_Remove(&call->entry);
7862 rxi_Free(call, sizeof(struct rx_call));
7865 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7866 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7868 opr_queue_Remove(&sq->entry);
7873 struct rx_peer **peer_ptr, **peer_end;
7874 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7875 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7877 struct rx_peer *peer, *next;
7879 MUTEX_ENTER(&rx_peerHashTable_lock);
7880 for (peer = *peer_ptr; peer; peer = next) {
7881 struct opr_queue *cursor, *store;
7884 MUTEX_ENTER(&rx_rpc_stats);
7885 MUTEX_ENTER(&peer->peer_lock);
7886 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7887 unsigned int num_funcs;
7888 struct rx_interface_stat *rpc_stat
7889 = opr_queue_Entry(cursor, struct rx_interface_stat,
7893 opr_queue_Remove(&rpc_stat->entry);
7894 opr_queue_Remove(&rpc_stat->entryPeers);
7895 num_funcs = rpc_stat->stats[0].func_total;
7897 sizeof(rx_interface_stat_t) +
7898 rpc_stat->stats[0].func_total *
7899 sizeof(rx_function_entry_v1_t);
7901 rxi_Free(rpc_stat, space);
7903 /* rx_rpc_stats must be held */
7904 rxi_rpc_peer_stat_cnt -= num_funcs;
7906 MUTEX_EXIT(&peer->peer_lock);
7907 MUTEX_EXIT(&rx_rpc_stats);
7911 if (rx_stats_active)
7912 rx_atomic_dec(&rx_stats.nPeerStructs);
7914 MUTEX_EXIT(&rx_peerHashTable_lock);
7917 for (i = 0; i < RX_MAX_SERVICES; i++) {
7919 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7921 for (i = 0; i < rx_hashTableSize; i++) {
7922 struct rx_connection *tc, *ntc;
7923 MUTEX_ENTER(&rx_connHashTable_lock);
7924 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7926 for (j = 0; j < RX_MAXCALLS; j++) {
7928 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7931 rxi_Free(tc, sizeof(*tc));
7933 MUTEX_EXIT(&rx_connHashTable_lock);
7936 MUTEX_ENTER(&freeSQEList_lock);
7938 while ((np = rx_FreeSQEList)) {
7939 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7940 MUTEX_DESTROY(&np->lock);
7941 rxi_Free(np, sizeof(*np));
7944 MUTEX_EXIT(&freeSQEList_lock);
7945 MUTEX_DESTROY(&freeSQEList_lock);
7946 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7947 MUTEX_DESTROY(&rx_connHashTable_lock);
7948 MUTEX_DESTROY(&rx_peerHashTable_lock);
7949 MUTEX_DESTROY(&rx_serverPool_lock);
7951 osi_Free(rx_connHashTable,
7952 rx_hashTableSize * sizeof(struct rx_connection *));
7953 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7955 UNPIN(rx_connHashTable,
7956 rx_hashTableSize * sizeof(struct rx_connection *));
7957 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7959 rxi_FreeAllPackets();
7961 MUTEX_ENTER(&rx_quota_mutex);
7962 rxi_dataQuota = RX_MAX_QUOTA;
7963 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7964 MUTEX_EXIT(&rx_quota_mutex);
7972 * Routines to implement connection specific data.
7976 rx_KeyCreate(rx_destructor_t rtn)
7979 MUTEX_ENTER(&rxi_keyCreate_lock);
7980 key = rxi_keyCreate_counter++;
7981 rxi_keyCreate_destructor = (rx_destructor_t *)
7982 realloc((void *)rxi_keyCreate_destructor,
7983 (key + 1) * sizeof(rx_destructor_t));
7984 rxi_keyCreate_destructor[key] = rtn;
7985 MUTEX_EXIT(&rxi_keyCreate_lock);
7990 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7993 MUTEX_ENTER(&conn->conn_data_lock);
7994 if (!conn->specific) {
7995 conn->specific = malloc((key + 1) * sizeof(void *));
7996 for (i = 0; i < key; i++)
7997 conn->specific[i] = NULL;
7998 conn->nSpecific = key + 1;
7999 conn->specific[key] = ptr;
8000 } else if (key >= conn->nSpecific) {
8001 conn->specific = (void **)
8002 realloc(conn->specific, (key + 1) * sizeof(void *));
8003 for (i = conn->nSpecific; i < key; i++)
8004 conn->specific[i] = NULL;
8005 conn->nSpecific = key + 1;
8006 conn->specific[key] = ptr;
8008 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8009 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8010 conn->specific[key] = ptr;
8012 MUTEX_EXIT(&conn->conn_data_lock);
8016 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8019 MUTEX_ENTER(&svc->svc_data_lock);
8020 if (!svc->specific) {
8021 svc->specific = malloc((key + 1) * sizeof(void *));
8022 for (i = 0; i < key; i++)
8023 svc->specific[i] = NULL;
8024 svc->nSpecific = key + 1;
8025 svc->specific[key] = ptr;
8026 } else if (key >= svc->nSpecific) {
8027 svc->specific = (void **)
8028 realloc(svc->specific, (key + 1) * sizeof(void *));
8029 for (i = svc->nSpecific; i < key; i++)
8030 svc->specific[i] = NULL;
8031 svc->nSpecific = key + 1;
8032 svc->specific[key] = ptr;
8034 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8035 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8036 svc->specific[key] = ptr;
8038 MUTEX_EXIT(&svc->svc_data_lock);
8042 rx_GetSpecific(struct rx_connection *conn, int key)
8045 MUTEX_ENTER(&conn->conn_data_lock);
8046 if (key >= conn->nSpecific)
8049 ptr = conn->specific[key];
8050 MUTEX_EXIT(&conn->conn_data_lock);
8055 rx_GetServiceSpecific(struct rx_service *svc, int key)
8058 MUTEX_ENTER(&svc->svc_data_lock);
8059 if (key >= svc->nSpecific)
8062 ptr = svc->specific[key];
8063 MUTEX_EXIT(&svc->svc_data_lock);
8068 #endif /* !KERNEL */
8071 * processStats is a queue used to store the statistics for the local
8072 * process. Its contents are similar to the contents of the rpcStats
8073 * queue on a rx_peer structure, but the actual data stored within
8074 * this queue contains totals across the lifetime of the process (assuming
8075 * the stats have not been reset) - unlike the per peer structures
8076 * which can come and go based upon the peer lifetime.
8079 static struct opr_queue processStats = { &processStats, &processStats };
8082 * peerStats is a queue used to store the statistics for all peer structs.
8083 * Its contents are the union of all the peer rpcStats queues.
8086 static struct opr_queue peerStats = { &peerStats, &peerStats };
8089 * rxi_monitor_processStats is used to turn process wide stat collection
8093 static int rxi_monitor_processStats = 0;
8096 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8099 static int rxi_monitor_peerStats = 0;
8103 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8105 rpc_stat->invocations = 0;
8106 rpc_stat->bytes_sent = 0;
8107 rpc_stat->bytes_rcvd = 0;
8108 rpc_stat->queue_time_sum.sec = 0;
8109 rpc_stat->queue_time_sum.usec = 0;
8110 rpc_stat->queue_time_sum_sqr.sec = 0;
8111 rpc_stat->queue_time_sum_sqr.usec = 0;
8112 rpc_stat->queue_time_min.sec = 9999999;
8113 rpc_stat->queue_time_min.usec = 9999999;
8114 rpc_stat->queue_time_max.sec = 0;
8115 rpc_stat->queue_time_max.usec = 0;
8116 rpc_stat->execution_time_sum.sec = 0;
8117 rpc_stat->execution_time_sum.usec = 0;
8118 rpc_stat->execution_time_sum_sqr.sec = 0;
8119 rpc_stat->execution_time_sum_sqr.usec = 0;
8120 rpc_stat->execution_time_min.sec = 9999999;
8121 rpc_stat->execution_time_min.usec = 9999999;
8122 rpc_stat->execution_time_max.sec = 0;
8123 rpc_stat->execution_time_max.usec = 0;
8127 * Given all of the information for a particular rpc
8128 * call, find or create (if requested) the stat structure for the rpc.
8131 * the queue of stats that will be updated with the new value
8133 * @param rxInterface
8134 * a unique number that identifies the rpc interface
8137 * the total number of functions in this interface. this is only
8138 * required if create is true
8141 * if true, this invocation was made to a server
8144 * the ip address of the remote host. this is only required if create
8145 * and addToPeerList are true
8148 * the port of the remote host. this is only required if create
8149 * and addToPeerList are true
8151 * @param addToPeerList
8152 * if != 0, add newly created stat to the global peer list
8155 * if a new stats structure is allocated, the counter will
8156 * be updated with the new number of allocated stat structures.
8157 * only required if create is true
8160 * if no stats structure exists, allocate one
8164 static rx_interface_stat_p
8165 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8166 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8167 afs_uint32 remotePort, int addToPeerList,
8168 unsigned int *counter, int create)
8170 rx_interface_stat_p rpc_stat = NULL;
8171 struct opr_queue *cursor;
8174 * See if there's already a structure for this interface
8177 for (opr_queue_Scan(stats, cursor)) {
8178 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8180 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8181 && (rpc_stat->stats[0].remote_is_server == isServer))
8185 /* if they didn't ask us to create, we're done */
8187 if (opr_queue_IsEnd(stats, cursor))
8193 /* can't proceed without these */
8194 if (!totalFunc || !counter)
8198 * Didn't find a match so allocate a new structure and add it to the
8202 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8203 || (rpc_stat->stats[0].interfaceId != rxInterface)
8204 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8209 sizeof(rx_interface_stat_t) +
8210 totalFunc * sizeof(rx_function_entry_v1_t);
8212 rpc_stat = rxi_Alloc(space);
8213 if (rpc_stat == NULL)
8216 *counter += totalFunc;
8217 for (i = 0; i < totalFunc; i++) {
8218 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8219 rpc_stat->stats[i].remote_peer = remoteHost;
8220 rpc_stat->stats[i].remote_port = remotePort;
8221 rpc_stat->stats[i].remote_is_server = isServer;
8222 rpc_stat->stats[i].interfaceId = rxInterface;
8223 rpc_stat->stats[i].func_total = totalFunc;
8224 rpc_stat->stats[i].func_index = i;
8226 opr_queue_Prepend(stats, &rpc_stat->entry);
8227 if (addToPeerList) {
8228 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8235 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8237 rx_interface_stat_p rpc_stat;
8240 if (rxInterface == -1)
8243 MUTEX_ENTER(&rx_rpc_stats);
8244 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8247 totalFunc = rpc_stat->stats[0].func_total;
8248 for (i = 0; i < totalFunc; i++)
8249 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8251 MUTEX_EXIT(&rx_rpc_stats);
8256 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8258 rx_interface_stat_p rpc_stat;
8260 struct rx_peer * peer;
8262 if (rxInterface == -1)
8265 peer = rxi_FindPeer(peerHost, peerPort, 0);
8269 MUTEX_ENTER(&rx_rpc_stats);
8270 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8273 totalFunc = rpc_stat->stats[0].func_total;
8274 for (i = 0; i < totalFunc; i++)
8275 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8277 MUTEX_EXIT(&rx_rpc_stats);
8282 rx_CopyProcessRPCStats(afs_uint64 op)
8284 rx_interface_stat_p rpc_stat;
8285 rx_function_entry_v1_p rpcop_stat =
8286 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8287 int currentFunc = (op & MAX_AFS_UINT32);
8288 afs_int32 rxInterface = (op >> 32);
8290 if (!rxi_monitor_processStats)
8293 if (rxInterface == -1)
8296 if (rpcop_stat == NULL)
8299 MUTEX_ENTER(&rx_rpc_stats);
8300 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8303 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8304 sizeof(rx_function_entry_v1_t));
8305 MUTEX_EXIT(&rx_rpc_stats);
8307 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8314 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8316 rx_interface_stat_p rpc_stat;
8317 rx_function_entry_v1_p rpcop_stat =
8318 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8319 int currentFunc = (op & MAX_AFS_UINT32);
8320 afs_int32 rxInterface = (op >> 32);
8321 struct rx_peer *peer;
8323 if (!rxi_monitor_peerStats)
8326 if (rxInterface == -1)
8329 if (rpcop_stat == NULL)
8332 peer = rxi_FindPeer(peerHost, peerPort, 0);
8336 MUTEX_ENTER(&rx_rpc_stats);
8337 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8340 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8341 sizeof(rx_function_entry_v1_t));
8342 MUTEX_EXIT(&rx_rpc_stats);
8344 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8351 rx_ReleaseRPCStats(void *stats)
8354 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8358 * Given all of the information for a particular rpc
8359 * call, create (if needed) and update the stat totals for the rpc.
8362 * the queue of stats that will be updated with the new value
8364 * @param rxInterface
8365 * a unique number that identifies the rpc interface
8367 * @param currentFunc
8368 * the index of the function being invoked
8371 * the total number of functions in this interface
8374 * the amount of time this function waited for a thread
8377 * the amount of time this function invocation took to execute
8380 * the number bytes sent by this invocation
8383 * the number bytes received by this invocation
8386 * if true, this invocation was made to a server
8389 * the ip address of the remote host
8392 * the port of the remote host
8394 * @param addToPeerList
8395 * if != 0, add newly created stat to the global peer list
8398 * if a new stats structure is allocated, the counter will
8399 * be updated with the new number of allocated stat structures
8404 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8405 afs_uint32 currentFunc, afs_uint32 totalFunc,
8406 struct clock *queueTime, struct clock *execTime,
8407 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8408 afs_uint32 remoteHost, afs_uint32 remotePort,
8409 int addToPeerList, unsigned int *counter)
8412 rx_interface_stat_p rpc_stat;
8414 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8415 remoteHost, remotePort, addToPeerList, counter,
8423 * Increment the stats for this function
8426 rpc_stat->stats[currentFunc].invocations++;
8427 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8428 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8429 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8430 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8431 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8432 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8434 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8435 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8437 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8438 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8440 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8441 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8443 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8444 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8452 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8453 afs_uint32 currentFunc, afs_uint32 totalFunc,
8454 struct clock *queueTime, struct clock *execTime,
8455 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8459 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8462 MUTEX_ENTER(&rx_rpc_stats);
8464 if (rxi_monitor_peerStats) {
8465 MUTEX_ENTER(&peer->peer_lock);
8466 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8467 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8468 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8469 MUTEX_EXIT(&peer->peer_lock);
8472 if (rxi_monitor_processStats) {
8473 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8474 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8475 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8478 MUTEX_EXIT(&rx_rpc_stats);
8482 * Increment the times and count for a particular rpc function.
8484 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8485 * call rx_RecordCallStatistics instead, so the public version of this
8486 * function is left purely for legacy callers.
8489 * The peer who invoked the rpc
8491 * @param rxInterface
8492 * A unique number that identifies the rpc interface
8494 * @param currentFunc
8495 * The index of the function being invoked
8498 * The total number of functions in this interface
8501 * The amount of time this function waited for a thread
8504 * The amount of time this function invocation took to execute
8507 * The number bytes sent by this invocation
8510 * The number bytes received by this invocation
8513 * If true, this invocation was made to a server
8517 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8518 afs_uint32 currentFunc, afs_uint32 totalFunc,
8519 struct clock *queueTime, struct clock *execTime,
8520 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8526 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8527 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8529 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8530 queueTime, execTime, sent64, rcvd64,
8537 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8541 * IN callerVersion - the rpc stat version of the caller.
8543 * IN count - the number of entries to marshall.
8545 * IN stats - pointer to stats to be marshalled.
8547 * OUT ptr - Where to store the marshalled data.
8554 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8555 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8561 * We only support the first version
8563 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8564 *(ptr++) = stats->remote_peer;
8565 *(ptr++) = stats->remote_port;
8566 *(ptr++) = stats->remote_is_server;
8567 *(ptr++) = stats->interfaceId;
8568 *(ptr++) = stats->func_total;
8569 *(ptr++) = stats->func_index;
8570 *(ptr++) = stats->invocations >> 32;
8571 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8572 *(ptr++) = stats->bytes_sent >> 32;
8573 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8574 *(ptr++) = stats->bytes_rcvd >> 32;
8575 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8576 *(ptr++) = stats->queue_time_sum.sec;
8577 *(ptr++) = stats->queue_time_sum.usec;
8578 *(ptr++) = stats->queue_time_sum_sqr.sec;
8579 *(ptr++) = stats->queue_time_sum_sqr.usec;
8580 *(ptr++) = stats->queue_time_min.sec;
8581 *(ptr++) = stats->queue_time_min.usec;
8582 *(ptr++) = stats->queue_time_max.sec;
8583 *(ptr++) = stats->queue_time_max.usec;
8584 *(ptr++) = stats->execution_time_sum.sec;
8585 *(ptr++) = stats->execution_time_sum.usec;
8586 *(ptr++) = stats->execution_time_sum_sqr.sec;
8587 *(ptr++) = stats->execution_time_sum_sqr.usec;
8588 *(ptr++) = stats->execution_time_min.sec;
8589 *(ptr++) = stats->execution_time_min.usec;
8590 *(ptr++) = stats->execution_time_max.sec;
8591 *(ptr++) = stats->execution_time_max.usec;
8597 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8602 * IN callerVersion - the rpc stat version of the caller
8604 * OUT myVersion - the rpc stat version of this function
8606 * OUT clock_sec - local time seconds
8608 * OUT clock_usec - local time microseconds
8610 * OUT allocSize - the number of bytes allocated to contain stats
8612 * OUT statCount - the number stats retrieved from this process.
8614 * OUT stats - the actual stats retrieved from this process.
8618 * Returns void. If successful, stats will != NULL.
8622 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8623 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8624 size_t * allocSize, afs_uint32 * statCount,
8625 afs_uint32 ** stats)
8635 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8638 * Check to see if stats are enabled
8641 MUTEX_ENTER(&rx_rpc_stats);
8642 if (!rxi_monitor_processStats) {
8643 MUTEX_EXIT(&rx_rpc_stats);
8647 clock_GetTime(&now);
8648 *clock_sec = now.sec;
8649 *clock_usec = now.usec;
8652 * Allocate the space based upon the caller version
8654 * If the client is at an older version than we are,
8655 * we return the statistic data in the older data format, but
8656 * we still return our version number so the client knows we
8657 * are maintaining more data than it can retrieve.
8660 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8661 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8662 *statCount = rxi_rpc_process_stat_cnt;
8665 * This can't happen yet, but in the future version changes
8666 * can be handled by adding additional code here
8670 if (space > (size_t) 0) {
8672 ptr = *stats = rxi_Alloc(space);
8675 struct opr_queue *cursor;
8677 for (opr_queue_Scan(&processStats, cursor)) {
8678 struct rx_interface_stat *rpc_stat =
8679 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8681 * Copy the data based upon the caller version
8683 rx_MarshallProcessRPCStats(callerVersion,
8684 rpc_stat->stats[0].func_total,
8685 rpc_stat->stats, &ptr);
8691 MUTEX_EXIT(&rx_rpc_stats);
8696 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8700 * IN callerVersion - the rpc stat version of the caller
8702 * OUT myVersion - the rpc stat version of this function
8704 * OUT clock_sec - local time seconds
8706 * OUT clock_usec - local time microseconds
8708 * OUT allocSize - the number of bytes allocated to contain stats
8710 * OUT statCount - the number of stats retrieved from the individual
8713 * OUT stats - the actual stats retrieved from the individual peer structures.
8717 * Returns void. If successful, stats will != NULL.
8721 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8722 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8723 size_t * allocSize, afs_uint32 * statCount,
8724 afs_uint32 ** stats)
8734 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8737 * Check to see if stats are enabled
8740 MUTEX_ENTER(&rx_rpc_stats);
8741 if (!rxi_monitor_peerStats) {
8742 MUTEX_EXIT(&rx_rpc_stats);
8746 clock_GetTime(&now);
8747 *clock_sec = now.sec;
8748 *clock_usec = now.usec;
8751 * Allocate the space based upon the caller version
8753 * If the client is at an older version than we are,
8754 * we return the statistic data in the older data format, but
8755 * we still return our version number so the client knows we
8756 * are maintaining more data than it can retrieve.
8759 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8760 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8761 *statCount = rxi_rpc_peer_stat_cnt;
8764 * This can't happen yet, but in the future version changes
8765 * can be handled by adding additional code here
8769 if (space > (size_t) 0) {
8771 ptr = *stats = rxi_Alloc(space);
8774 struct opr_queue *cursor;
8776 for (opr_queue_Scan(&peerStats, cursor)) {
8777 struct rx_interface_stat *rpc_stat
8778 = opr_queue_Entry(cursor, struct rx_interface_stat,
8782 * Copy the data based upon the caller version
8784 rx_MarshallProcessRPCStats(callerVersion,
8785 rpc_stat->stats[0].func_total,
8786 rpc_stat->stats, &ptr);
8792 MUTEX_EXIT(&rx_rpc_stats);
8797 * rx_FreeRPCStats - free memory allocated by
8798 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8802 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8803 * rx_RetrievePeerRPCStats
8805 * IN allocSize - the number of bytes in stats.
8813 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8815 rxi_Free(stats, allocSize);
8819 * rx_queryProcessRPCStats - see if process rpc stat collection is
8820 * currently enabled.
8826 * Returns 0 if stats are not enabled != 0 otherwise
8830 rx_queryProcessRPCStats(void)
8833 MUTEX_ENTER(&rx_rpc_stats);
8834 rc = rxi_monitor_processStats;
8835 MUTEX_EXIT(&rx_rpc_stats);
8840 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8846 * Returns 0 if stats are not enabled != 0 otherwise
8850 rx_queryPeerRPCStats(void)
8853 MUTEX_ENTER(&rx_rpc_stats);
8854 rc = rxi_monitor_peerStats;
8855 MUTEX_EXIT(&rx_rpc_stats);
8860 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8870 rx_enableProcessRPCStats(void)
8872 MUTEX_ENTER(&rx_rpc_stats);
8873 rx_enable_stats = 1;
8874 rxi_monitor_processStats = 1;
8875 MUTEX_EXIT(&rx_rpc_stats);
8879 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8889 rx_enablePeerRPCStats(void)
8891 MUTEX_ENTER(&rx_rpc_stats);
8892 rx_enable_stats = 1;
8893 rxi_monitor_peerStats = 1;
8894 MUTEX_EXIT(&rx_rpc_stats);
8898 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8908 rx_disableProcessRPCStats(void)
8910 struct opr_queue *cursor, *store;
8913 MUTEX_ENTER(&rx_rpc_stats);
8916 * Turn off process statistics and if peer stats is also off, turn
8920 rxi_monitor_processStats = 0;
8921 if (rxi_monitor_peerStats == 0) {
8922 rx_enable_stats = 0;
8925 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8926 unsigned int num_funcs = 0;
8927 struct rx_interface_stat *rpc_stat
8928 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8930 opr_queue_Remove(&rpc_stat->entry);
8932 num_funcs = rpc_stat->stats[0].func_total;
8934 sizeof(rx_interface_stat_t) +
8935 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8937 rxi_Free(rpc_stat, space);
8938 rxi_rpc_process_stat_cnt -= num_funcs;
8940 MUTEX_EXIT(&rx_rpc_stats);
8944 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8954 rx_disablePeerRPCStats(void)
8956 struct rx_peer **peer_ptr, **peer_end;
8960 * Turn off peer statistics and if process stats is also off, turn
8964 rxi_monitor_peerStats = 0;
8965 if (rxi_monitor_processStats == 0) {
8966 rx_enable_stats = 0;
8969 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8970 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8972 struct rx_peer *peer, *next, *prev;
8974 MUTEX_ENTER(&rx_peerHashTable_lock);
8975 MUTEX_ENTER(&rx_rpc_stats);
8976 for (prev = peer = *peer_ptr; peer; peer = next) {
8978 code = MUTEX_TRYENTER(&peer->peer_lock);
8981 struct opr_queue *cursor, *store;
8983 if (prev == *peer_ptr) {
8994 MUTEX_EXIT(&rx_peerHashTable_lock);
8996 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
8997 unsigned int num_funcs = 0;
8998 struct rx_interface_stat *rpc_stat
8999 = opr_queue_Entry(cursor, struct rx_interface_stat,
9002 opr_queue_Remove(&rpc_stat->entry);
9003 opr_queue_Remove(&rpc_stat->entryPeers);
9004 num_funcs = rpc_stat->stats[0].func_total;
9006 sizeof(rx_interface_stat_t) +
9007 rpc_stat->stats[0].func_total *
9008 sizeof(rx_function_entry_v1_t);
9010 rxi_Free(rpc_stat, space);
9011 rxi_rpc_peer_stat_cnt -= num_funcs;
9013 MUTEX_EXIT(&peer->peer_lock);
9015 MUTEX_ENTER(&rx_peerHashTable_lock);
9025 MUTEX_EXIT(&rx_rpc_stats);
9026 MUTEX_EXIT(&rx_peerHashTable_lock);
9031 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9036 * IN clearFlag - flag indicating which stats to clear
9044 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9046 struct opr_queue *cursor;
9048 MUTEX_ENTER(&rx_rpc_stats);
9050 for (opr_queue_Scan(&processStats, cursor)) {
9051 unsigned int num_funcs = 0, i;
9052 struct rx_interface_stat *rpc_stat
9053 = opr_queue_Entry(rpc_stat, struct rx_interface_stat, entry);
9055 num_funcs = rpc_stat->stats[0].func_total;
9056 for (i = 0; i < num_funcs; i++) {
9057 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9058 rpc_stat->stats[i].invocations = 0;
9060 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9061 rpc_stat->stats[i].bytes_sent = 0;
9063 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9064 rpc_stat->stats[i].bytes_rcvd = 0;
9066 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9067 rpc_stat->stats[i].queue_time_sum.sec = 0;
9068 rpc_stat->stats[i].queue_time_sum.usec = 0;
9070 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9071 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9072 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9074 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9075 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9076 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9078 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9079 rpc_stat->stats[i].queue_time_max.sec = 0;
9080 rpc_stat->stats[i].queue_time_max.usec = 0;
9082 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9083 rpc_stat->stats[i].execution_time_sum.sec = 0;
9084 rpc_stat->stats[i].execution_time_sum.usec = 0;
9086 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9087 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9088 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9090 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9091 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9092 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9094 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9095 rpc_stat->stats[i].execution_time_max.sec = 0;
9096 rpc_stat->stats[i].execution_time_max.usec = 0;
9101 MUTEX_EXIT(&rx_rpc_stats);
9105 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9110 * IN clearFlag - flag indicating which stats to clear
9118 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9120 struct opr_queue *cursor;
9122 MUTEX_ENTER(&rx_rpc_stats);
9124 for (opr_queue_Scan(&peerStats, cursor)) {
9125 unsigned int num_funcs, i;
9126 struct rx_interface_stat *rpc_stat
9127 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9129 num_funcs = rpc_stat->stats[0].func_total;
9130 for (i = 0; i < num_funcs; i++) {
9131 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9132 rpc_stat->stats[i].invocations = 0;
9134 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9135 rpc_stat->stats[i].bytes_sent = 0;
9137 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9138 rpc_stat->stats[i].bytes_rcvd = 0;
9140 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9141 rpc_stat->stats[i].queue_time_sum.sec = 0;
9142 rpc_stat->stats[i].queue_time_sum.usec = 0;
9144 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9145 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9146 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9148 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9149 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9150 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9152 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9153 rpc_stat->stats[i].queue_time_max.sec = 0;
9154 rpc_stat->stats[i].queue_time_max.usec = 0;
9156 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9157 rpc_stat->stats[i].execution_time_sum.sec = 0;
9158 rpc_stat->stats[i].execution_time_sum.usec = 0;
9160 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9161 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9162 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9164 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9165 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9166 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9168 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9169 rpc_stat->stats[i].execution_time_max.sec = 0;
9170 rpc_stat->stats[i].execution_time_max.usec = 0;
9175 MUTEX_EXIT(&rx_rpc_stats);
9179 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9180 * is authorized to enable/disable/clear RX statistics.
9182 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9185 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9187 rxi_rxstat_userok = proc;
9191 rx_RxStatUserOk(struct rx_call *call)
9193 if (!rxi_rxstat_userok)
9195 return rxi_rxstat_userok(call);
9200 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9201 * function in the MSVC runtime DLL (msvcrt.dll).
9203 * Note: the system serializes calls to this function.
9206 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9207 DWORD reason, /* reason function is being called */
9208 LPVOID reserved) /* reserved for future use */
9211 case DLL_PROCESS_ATTACH:
9212 /* library is being attached to a process */
9216 case DLL_PROCESS_DETACH:
9223 #endif /* AFS_NT40_ENV */
9226 int rx_DumpCalls(FILE *outputFile, char *cookie)
9228 #ifdef RXDEBUG_PACKET
9229 #ifdef KDUMP_RX_LOCK
9230 struct rx_call_rx_lock *c;
9237 #define RXDPRINTF sprintf
9238 #define RXDPRINTOUT output
9240 #define RXDPRINTF fprintf
9241 #define RXDPRINTOUT outputFile
9244 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9246 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9249 for (c = rx_allCallsp; c; c = c->allNextp) {
9250 u_short rqc, tqc, iovqc;
9252 MUTEX_ENTER(&c->lock);
9253 rqc = opr_queue_Count(&c->rq);
9254 tqc = opr_queue_Count(&c->tq);
9255 iovqc = opr_queue_Count(&c->app.iovq);
9257 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, "
9258 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9259 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9260 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9261 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9262 #ifdef RX_ENABLE_LOCKS
9265 #ifdef RX_REFCOUNT_CHECK
9266 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9267 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9270 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->app.mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
9271 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9272 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9273 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9274 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9275 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9276 #ifdef RX_ENABLE_LOCKS
9277 , (afs_uint32)c->refCount
9279 #ifdef RX_REFCOUNT_CHECK
9280 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9283 MUTEX_EXIT(&c->lock);
9286 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9289 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9291 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9293 #endif /* RXDEBUG_PACKET */