2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
75 #include <opr/queue.h>
76 #include <hcrypto/rand.h>
80 #include "rx_atomic.h"
81 #include "rx_globals.h"
83 #include "rx_internal.h"
90 #include "rx_packet.h"
91 #include "rx_server.h"
93 #include <afs/rxgen_consts.h>
96 #ifdef AFS_PTHREAD_ENV
98 int (*registerProgram) (pid_t, char *) = 0;
99 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
102 int (*registerProgram) (PROCESS, char *) = 0;
103 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
107 /* Local static routines */
108 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
109 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
110 struct rx_call *, struct rx_peer *,
112 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
114 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
115 void *dummy, int dummy2);
116 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
117 void *dummy, int dummy2);
118 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
119 void *unused, int unused2);
120 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
121 void *unused2, int unused3);
122 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
123 struct rx_packet *packet,
124 int istack, int force);
125 static void rxi_AckAll(struct rx_call *call);
126 static struct rx_connection
127 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
128 u_short serviceId, afs_uint32 cid,
129 afs_uint32 epoch, int type, u_int securityIndex,
130 int *unknownService);
131 static struct rx_packet
132 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
133 int istack, osi_socket socket,
134 afs_uint32 host, u_short port, int *tnop,
135 struct rx_call **newcallp);
136 static struct rx_packet
137 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
139 static struct rx_packet
140 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
141 struct rx_packet *np, int istack);
142 static struct rx_packet
143 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
144 struct rx_packet *np, int istack);
145 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
146 int *tnop, struct rx_call **newcallp);
147 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
148 static void rxi_ClearReceiveQueue(struct rx_call *call);
149 static void rxi_ResetCall(struct rx_call *call, int newcall);
150 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
151 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
152 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
153 static void rxi_KeepAliveOn(struct rx_call *call);
154 static void rxi_GrowMTUOn(struct rx_call *call);
155 static void rxi_ChallengeOn(struct rx_connection *conn);
156 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
157 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
158 static void rxi_CancelKeepAliveEvent(struct rx_call *call);
159 static void rxi_CancelDelayedAbortEvent(struct rx_call *call);
160 static void rxi_CancelGrowMTUEvent(struct rx_call *call);
161 static void update_nextCid(void);
163 #ifdef RX_ENABLE_LOCKS
165 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
166 rx_atomic_t rxi_start_in_error;
168 #endif /* RX_ENABLE_LOCKS */
170 /* Constant delay time before sending an acknowledge of the last packet
171 * received. This is to avoid sending an extra acknowledge when the
172 * client is about to make another call, anyway, or the server is
175 * The lastAckDelay may not exceeed 400ms without causing peers to
176 * unecessarily timeout.
178 struct clock rx_lastAckDelay = {0, 400000};
180 /* Constant delay time before sending a soft ack when none was requested.
181 * This is to make sure we send soft acks before the sender times out,
182 * Normally we wait and send a hard ack when the receiver consumes the packet
184 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
185 * will require changes to the peer's RTT calculations.
187 struct clock rx_softAckDelay = {0, 100000};
190 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
191 * currently allocated within rx. This number is used to allocate the
192 * memory required to return the statistics when queried.
193 * Protected by the rx_rpc_stats mutex.
196 static unsigned int rxi_rpc_peer_stat_cnt;
199 * rxi_rpc_process_stat_cnt counts the total number of local process stat
200 * structures currently allocated within rx. The number is used to allocate
201 * the memory required to return the statistics when queried.
202 * Protected by the rx_rpc_stats mutex.
205 static unsigned int rxi_rpc_process_stat_cnt;
208 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
209 * errors should be reported to the application when a call channel appears busy
210 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
211 * and there are other call channels in the connection that are not busy.
212 * If 0, we do not return errors upon receiving busy packets; we just keep
213 * trying on the same call channel until we hit a timeout.
215 static afs_int32 rxi_busyChannelError = 0;
217 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
218 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
220 /* Incoming calls wait on this queue when there are no available
221 * server processes */
222 struct opr_queue rx_incomingCallQueue;
224 /* Server processes wait on this queue when there are no appropriate
225 * calls to process */
226 struct opr_queue rx_idleServerQueue;
228 #if !defined(offsetof)
229 #include <stddef.h> /* for definition of offsetof() */
232 #ifdef RX_ENABLE_LOCKS
233 afs_kmutex_t rx_atomic_mutex;
236 /* Forward prototypes */
237 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
240 putConnection (struct rx_connection *conn) {
241 MUTEX_ENTER(&rx_refcnt_mutex);
243 MUTEX_EXIT(&rx_refcnt_mutex);
246 #ifdef AFS_PTHREAD_ENV
249 * Use procedural initialization of mutexes/condition variables
253 extern afs_kmutex_t rx_quota_mutex;
254 extern afs_kmutex_t rx_pthread_mutex;
255 extern afs_kmutex_t rx_packets_mutex;
256 extern afs_kmutex_t rx_refcnt_mutex;
257 extern afs_kmutex_t des_init_mutex;
258 extern afs_kmutex_t des_random_mutex;
260 extern afs_kmutex_t rx_clock_mutex;
261 extern afs_kmutex_t rxi_connCacheMutex;
262 extern afs_kmutex_t event_handler_mutex;
263 extern afs_kmutex_t listener_mutex;
264 extern afs_kmutex_t rx_if_init_mutex;
265 extern afs_kmutex_t rx_if_mutex;
267 extern afs_kcondvar_t rx_event_handler_cond;
268 extern afs_kcondvar_t rx_listener_cond;
271 static afs_kmutex_t epoch_mutex;
272 static afs_kmutex_t rx_init_mutex;
273 static afs_kmutex_t rx_debug_mutex;
274 static afs_kmutex_t rx_rpc_stats;
277 rxi_InitPthread(void)
279 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
286 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
287 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
288 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
289 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
291 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
292 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
293 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
294 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
295 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
298 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
299 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
302 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
303 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
305 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
306 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
307 #ifdef RX_ENABLE_LOCKS
310 #endif /* RX_LOCKS_DB */
311 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
312 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
314 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
316 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
318 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
320 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
322 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
324 #endif /* RX_ENABLE_LOCKS */
327 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
328 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
330 * The rx_stats_mutex mutex protects the following global variables:
331 * rxi_lowConnRefCount
332 * rxi_lowPeerRefCount
341 * The rx_quota_mutex mutex protects the following global variables:
349 * The rx_freePktQ_lock protects the following global variables:
354 * The rx_packets_mutex mutex protects the following global variables:
362 * The rx_pthread_mutex mutex protects the following global variables:
363 * rxi_fcfs_thread_num
366 #define INIT_PTHREAD_LOCKS
370 /* Variables for handling the minProcs implementation. availProcs gives the
371 * number of threads available in the pool at this moment (not counting dudes
372 * executing right now). totalMin gives the total number of procs required
373 * for handling all minProcs requests. minDeficit is a dynamic variable
374 * tracking the # of procs required to satisfy all of the remaining minProcs
376 * For fine grain locking to work, the quota check and the reservation of
377 * a server thread has to come while rxi_availProcs and rxi_minDeficit
378 * are locked. To this end, the code has been modified under #ifdef
379 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
380 * same time. A new function, ReturnToServerPool() returns the allocation.
382 * A call can be on several queue's (but only one at a time). When
383 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
384 * that no one else is touching the queue. To this end, we store the address
385 * of the queue lock in the call structure (under the call lock) when we
386 * put the call on a queue, and we clear the call_queue_lock when the
387 * call is removed from a queue (once the call lock has been obtained).
388 * This allows rxi_ResetCall to safely synchronize with others wishing
389 * to manipulate the queue.
392 #if defined(RX_ENABLE_LOCKS)
393 static afs_kmutex_t rx_rpc_stats;
396 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
397 ** pretty good that the next packet coming in is from the same connection
398 ** as the last packet, since we're send multiple packets in a transmit window.
400 struct rx_connection *rxLastConn = 0;
402 #ifdef RX_ENABLE_LOCKS
403 /* The locking hierarchy for rx fine grain locking is composed of these
406 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
407 * also protects updates to rx_nextCid
408 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
409 * call->lock - locks call data fields.
410 * These are independent of each other:
411 * rx_freeCallQueue_lock
416 * serverQueueEntry->lock
417 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
419 * peer->lock - locks peer data fields.
420 * conn_data_lock - that more than one thread is not updating a conn data
421 * field at the same time.
432 * Do we need a lock to protect the peer field in the conn structure?
433 * conn->peer was previously a constant for all intents and so has no
434 * lock protecting this field. The multihomed client delta introduced
435 * a RX code change : change the peer field in the connection structure
436 * to that remote interface from which the last packet for this
437 * connection was sent out. This may become an issue if further changes
440 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
441 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
443 /* rxdb_fileID is used to identify the lock location, along with line#. */
444 static int rxdb_fileID = RXDB_FILE_RX;
445 #endif /* RX_LOCKS_DB */
446 #else /* RX_ENABLE_LOCKS */
447 #define SET_CALL_QUEUE_LOCK(C, L)
448 #define CLEAR_CALL_QUEUE_LOCK(C)
449 #endif /* RX_ENABLE_LOCKS */
450 struct rx_serverQueueEntry *rx_waitForPacket = 0;
452 /* ------------Exported Interfaces------------- */
454 /* Initialize rx. A port number may be mentioned, in which case this
455 * becomes the default port number for any service installed later.
456 * If 0 is provided for the port number, a random port will be chosen
457 * by the kernel. Whether this will ever overlap anything in
458 * /etc/services is anybody's guess... Returns 0 on success, -1 on
463 rx_atomic_t rxinit_status = RX_ATOMIC_INIT(1);
466 rx_InitHost(u_int host, u_int port)
473 char *htable, *ptable;
478 if (!rx_atomic_test_and_clear_bit(&rxinit_status, 0))
479 return 0; /* already started */
485 if (afs_winsockInit() < 0)
491 * Initialize anything necessary to provide a non-premptive threading
494 rxi_InitializeThreadSupport();
497 /* Allocate and initialize a socket for client and perhaps server
500 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
501 if (rx_socket == OSI_NULLSOCKET) {
504 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
507 #endif /* RX_LOCKS_DB */
508 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
509 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
510 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
511 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
512 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
513 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
514 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
515 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
516 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
517 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
519 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
521 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
523 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
525 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
526 #if defined(AFS_HPUX110_ENV)
528 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
529 #endif /* AFS_HPUX110_ENV */
530 #endif /* RX_ENABLE_LOCKS && KERNEL */
533 rx_connDeadTime = 12;
534 rx_tranquil = 0; /* reset flag */
535 rxi_ResetStatistics();
536 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
537 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
538 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
539 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
540 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
541 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
543 /* Malloc up a bunch of packets & buffers */
545 opr_queue_Init(&rx_freePacketQueue);
546 rxi_NeedMorePackets = FALSE;
547 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
549 /* enforce a minimum number of allocated packets */
550 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
551 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
553 /* allocate the initial free packet pool */
554 #ifdef RX_ENABLE_TSFPQ
555 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
556 #else /* RX_ENABLE_TSFPQ */
557 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
558 #endif /* RX_ENABLE_TSFPQ */
565 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
566 tv.tv_sec = clock_now.sec;
567 tv.tv_usec = clock_now.usec;
568 srand((unsigned int)tv.tv_usec);
575 #if defined(KERNEL) && !defined(UKERNEL)
576 /* Really, this should never happen in a real kernel */
579 struct sockaddr_in addr;
581 int addrlen = sizeof(addr);
583 socklen_t addrlen = sizeof(addr);
585 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
587 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
590 rx_port = addr.sin_port;
593 rx_stats.minRtt.sec = 9999999;
594 if (RAND_bytes(&rx_epoch, sizeof(rx_epoch)) != 1)
596 rx_epoch = (rx_epoch & ~0x40000000) | 0x80000000;
597 if (RAND_bytes(&rx_nextCid, sizeof(rx_nextCid)) != 1)
599 rx_nextCid &= RX_CIDMASK;
600 MUTEX_ENTER(&rx_quota_mutex);
601 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
602 MUTEX_EXIT(&rx_quota_mutex);
603 /* *Slightly* random start time for the cid. This is just to help
604 * out with the hashing function at the peer */
605 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
606 rx_connHashTable = (struct rx_connection **)htable;
607 rx_peerHashTable = (struct rx_peer **)ptable;
609 rx_hardAckDelay.sec = 0;
610 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
612 rxevent_Init(20, rxi_ReScheduleEvents);
614 /* Initialize various global queues */
615 opr_queue_Init(&rx_idleServerQueue);
616 opr_queue_Init(&rx_incomingCallQueue);
617 opr_queue_Init(&rx_freeCallQueue);
619 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
620 /* Initialize our list of usable IP addresses. */
624 /* Start listener process (exact function is dependent on the
625 * implementation environment--kernel or user space) */
629 rx_atomic_clear_bit(&rxinit_status, 0);
636 return rx_InitHost(htonl(INADDR_ANY), port);
642 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
643 * maintaing the round trip timer.
648 * Start a new RTT timer for a given call and packet.
650 * There must be no resendEvent already listed for this call, otherwise this
651 * will leak events - intended for internal use within the RTO code only
654 * the RX call to start the timer for
655 * @param[in] lastPacket
656 * a flag indicating whether the last packet has been sent or not
658 * @pre call must be locked before calling this function
662 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
664 struct clock now, retryTime;
669 clock_Add(&retryTime, &call->rto);
671 /* If we're sending the last packet, and we're the client, then the server
672 * may wait for an additional 400ms before returning the ACK, wait for it
673 * rather than hitting a timeout */
674 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
675 clock_Addmsec(&retryTime, 400);
677 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
678 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
683 * Cancel an RTT timer for a given call.
687 * the RX call to cancel the timer for
689 * @pre call must be locked before calling this function
694 rxi_rto_cancel(struct rx_call *call)
696 if (call->resendEvent != NULL) {
697 rxevent_Cancel(&call->resendEvent);
698 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
703 * Tell the RTO timer that we have sent a packet.
705 * If the timer isn't already running, then start it. If the timer is running,
709 * the RX call that the packet has been sent on
710 * @param[in] lastPacket
711 * A flag which is true if this is the last packet for the call
713 * @pre The call must be locked before calling this function
718 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
720 if (call->resendEvent)
723 rxi_rto_startTimer(call, lastPacket, istack);
727 * Tell the RTO timer that we have received an new ACK message
729 * This function should be called whenever a call receives an ACK that
730 * acknowledges new packets. Whatever happens, we stop the current timer.
731 * If there are unacked packets in the queue which have been sent, then
732 * we restart the timer from now. Otherwise, we leave it stopped.
735 * the RX call that the ACK has been received on
739 rxi_rto_packet_acked(struct rx_call *call, int istack)
741 struct opr_queue *cursor;
743 rxi_rto_cancel(call);
745 if (opr_queue_IsEmpty(&call->tq))
748 for (opr_queue_Scan(&call->tq, cursor)) {
749 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
750 if (p->header.seq > call->tfirst + call->twind)
753 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
754 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
762 * Set an initial round trip timeout for a peer connection
764 * @param[in] secs The timeout to set in seconds
768 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
769 peer->rtt = secs * 8000;
773 * Enables or disables the busy call channel error (RX_CALL_BUSY).
775 * @param[in] onoff Non-zero to enable busy call channel errors.
777 * @pre Neither rx_Init nor rx_InitHost have been called yet
780 rx_SetBusyChannelError(afs_int32 onoff)
782 osi_Assert(rx_atomic_test_bit(&rxinit_status, 0));
783 rxi_busyChannelError = onoff ? 1 : 0;
787 * Set a delayed ack event on the specified call for the given time
789 * @param[in] call - the call on which to set the event
790 * @param[in] offset - the delay from now after which the event fires
793 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
795 struct clock now, when;
799 clock_Add(&when, offset);
801 if (call->delayedAckEvent && clock_Gt(&call->delayedAckTime, &when)) {
802 /* The event we're cancelling already has a reference, so we don't
804 rxevent_Cancel(&call->delayedAckEvent);
805 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
808 call->delayedAckTime = when;
809 } else if (!call->delayedAckEvent) {
810 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
811 call->delayedAckEvent = rxevent_Post(&when, &now,
814 call->delayedAckTime = when;
819 rxi_CancelDelayedAckEvent(struct rx_call *call)
821 if (call->delayedAckEvent) {
822 rxevent_Cancel(&call->delayedAckEvent);
823 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
827 /* called with unincremented nRequestsRunning to see if it is OK to start
828 * a new thread in this service. Could be "no" for two reasons: over the
829 * max quota, or would prevent others from reaching their min quota.
831 #ifdef RX_ENABLE_LOCKS
832 /* This verion of QuotaOK reserves quota if it's ok while the
833 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
836 QuotaOK(struct rx_service *aservice)
838 /* check if over max quota */
839 if (aservice->nRequestsRunning >= aservice->maxProcs) {
843 /* under min quota, we're OK */
844 /* otherwise, can use only if there are enough to allow everyone
845 * to go to their min quota after this guy starts.
848 MUTEX_ENTER(&rx_quota_mutex);
849 if ((aservice->nRequestsRunning < aservice->minProcs)
850 || (rxi_availProcs > rxi_minDeficit)) {
851 aservice->nRequestsRunning++;
852 /* just started call in minProcs pool, need fewer to maintain
854 if (aservice->nRequestsRunning <= aservice->minProcs)
857 MUTEX_EXIT(&rx_quota_mutex);
860 MUTEX_EXIT(&rx_quota_mutex);
866 ReturnToServerPool(struct rx_service *aservice)
868 aservice->nRequestsRunning--;
869 MUTEX_ENTER(&rx_quota_mutex);
870 if (aservice->nRequestsRunning < aservice->minProcs)
873 MUTEX_EXIT(&rx_quota_mutex);
876 #else /* RX_ENABLE_LOCKS */
878 QuotaOK(struct rx_service *aservice)
881 /* under min quota, we're OK */
882 if (aservice->nRequestsRunning < aservice->minProcs)
885 /* check if over max quota */
886 if (aservice->nRequestsRunning >= aservice->maxProcs)
889 /* otherwise, can use only if there are enough to allow everyone
890 * to go to their min quota after this guy starts.
892 MUTEX_ENTER(&rx_quota_mutex);
893 if (rxi_availProcs > rxi_minDeficit)
895 MUTEX_EXIT(&rx_quota_mutex);
898 #endif /* RX_ENABLE_LOCKS */
901 /* Called by rx_StartServer to start up lwp's to service calls.
902 NExistingProcs gives the number of procs already existing, and which
903 therefore needn't be created. */
905 rxi_StartServerProcs(int nExistingProcs)
907 struct rx_service *service;
912 /* For each service, reserve N processes, where N is the "minimum"
913 * number of processes that MUST be able to execute a request in parallel,
914 * at any time, for that process. Also compute the maximum difference
915 * between any service's maximum number of processes that can run
916 * (i.e. the maximum number that ever will be run, and a guarantee
917 * that this number will run if other services aren't running), and its
918 * minimum number. The result is the extra number of processes that
919 * we need in order to provide the latter guarantee */
920 for (i = 0; i < RX_MAX_SERVICES; i++) {
922 service = rx_services[i];
923 if (service == (struct rx_service *)0)
925 nProcs += service->minProcs;
926 diff = service->maxProcs - service->minProcs;
930 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
931 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
932 for (i = 0; i < nProcs; i++) {
933 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
939 /* This routine is only required on Windows */
941 rx_StartClientThread(void)
943 #ifdef AFS_PTHREAD_ENV
945 pid = pthread_self();
946 #endif /* AFS_PTHREAD_ENV */
948 #endif /* AFS_NT40_ENV */
950 /* This routine must be called if any services are exported. If the
951 * donateMe flag is set, the calling process is donated to the server
954 rx_StartServer(int donateMe)
956 struct rx_service *service;
962 /* Start server processes, if necessary (exact function is dependent
963 * on the implementation environment--kernel or user space). DonateMe
964 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
965 * case, one less new proc will be created rx_StartServerProcs.
967 rxi_StartServerProcs(donateMe);
969 /* count up the # of threads in minProcs, and add set the min deficit to
970 * be that value, too.
972 for (i = 0; i < RX_MAX_SERVICES; i++) {
973 service = rx_services[i];
974 if (service == (struct rx_service *)0)
976 MUTEX_ENTER(&rx_quota_mutex);
977 rxi_totalMin += service->minProcs;
978 /* below works even if a thread is running, since minDeficit would
979 * still have been decremented and later re-incremented.
981 rxi_minDeficit += service->minProcs;
982 MUTEX_EXIT(&rx_quota_mutex);
985 /* Turn on reaping of idle server connections */
986 rxi_ReapConnections(NULL, NULL, NULL, 0);
995 #ifdef AFS_PTHREAD_ENV
997 pid = afs_pointer_to_int(pthread_self());
998 #else /* AFS_PTHREAD_ENV */
1000 LWP_CurrentProcess(&pid);
1001 #endif /* AFS_PTHREAD_ENV */
1003 sprintf(name, "srv_%d", ++nProcs);
1004 if (registerProgram)
1005 (*registerProgram) (pid, name);
1007 #endif /* AFS_NT40_ENV */
1008 rx_ServerProc(NULL); /* Never returns */
1010 #ifdef RX_ENABLE_TSFPQ
1011 /* no use leaving packets around in this thread's local queue if
1012 * it isn't getting donated to the server thread pool.
1014 rxi_FlushLocalPacketsTSFPQ();
1015 #endif /* RX_ENABLE_TSFPQ */
1019 /* Create a new client connection to the specified service, using the
1020 * specified security object to implement the security model for this
1022 struct rx_connection *
1023 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1024 struct rx_securityClass *securityObject,
1025 int serviceSecurityIndex)
1028 struct rx_connection *conn;
1033 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1034 "serviceSecurityIndex %d)\n",
1035 ntohl(shost), ntohs(sport), sservice, securityObject,
1036 serviceSecurityIndex));
1038 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1039 * the case of kmem_alloc? */
1040 conn = rxi_AllocConnection();
1041 #ifdef RX_ENABLE_LOCKS
1042 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1043 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1044 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1047 MUTEX_ENTER(&rx_connHashTable_lock);
1048 conn->type = RX_CLIENT_CONNECTION;
1049 conn->epoch = rx_epoch;
1050 conn->cid = rx_nextCid;
1052 conn->peer = rxi_FindPeer(shost, sport, 1);
1053 conn->serviceId = sservice;
1054 conn->securityObject = securityObject;
1055 conn->securityData = (void *) 0;
1056 conn->securityIndex = serviceSecurityIndex;
1057 rx_SetConnDeadTime(conn, rx_connDeadTime);
1058 rx_SetConnSecondsUntilNatPing(conn, 0);
1059 conn->ackRate = RX_FAST_ACK_RATE;
1060 conn->nSpecific = 0;
1061 conn->specific = NULL;
1062 conn->challengeEvent = NULL;
1063 conn->delayedAbortEvent = NULL;
1064 conn->abortCount = 0;
1066 for (i = 0; i < RX_MAXCALLS; i++) {
1067 conn->twind[i] = rx_initSendWindow;
1068 conn->rwind[i] = rx_initReceiveWindow;
1069 conn->lastBusy[i] = 0;
1072 RXS_NewConnection(securityObject, conn);
1074 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1076 conn->refCount++; /* no lock required since only this thread knows... */
1077 conn->next = rx_connHashTable[hashindex];
1078 rx_connHashTable[hashindex] = conn;
1079 if (rx_stats_active)
1080 rx_atomic_inc(&rx_stats.nClientConns);
1081 MUTEX_EXIT(&rx_connHashTable_lock);
1087 * Ensure a connection's timeout values are valid.
1089 * @param[in] conn The connection to check
1091 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1092 * unless idleDeadTime and/or hardDeadTime are not set
1096 rxi_CheckConnTimeouts(struct rx_connection *conn)
1098 /* a connection's timeouts must have the relationship
1099 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1100 * total loss of network to a peer may cause an idle timeout instead of a
1101 * dead timeout, simply because the idle timeout gets hit first. Also set
1102 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1103 /* this logic is slightly complicated by the fact that
1104 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1106 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1107 if (conn->idleDeadTime) {
1108 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1110 if (conn->hardDeadTime) {
1111 if (conn->idleDeadTime) {
1112 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1114 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1120 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1122 /* The idea is to set the dead time to a value that allows several
1123 * keepalives to be dropped without timing out the connection. */
1124 conn->secondsUntilDead = seconds;
1125 rxi_CheckConnTimeouts(conn);
1126 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1130 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1132 conn->hardDeadTime = seconds;
1133 rxi_CheckConnTimeouts(conn);
1137 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1139 conn->idleDeadTime = seconds;
1140 conn->idleDeadDetection = (seconds ? 1 : 0);
1141 rxi_CheckConnTimeouts(conn);
1144 int rxi_lowPeerRefCount = 0;
1145 int rxi_lowConnRefCount = 0;
1148 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1149 * NOTE: must not be called with rx_connHashTable_lock held.
1152 rxi_CleanupConnection(struct rx_connection *conn)
1154 /* Notify the service exporter, if requested, that this connection
1155 * is being destroyed */
1156 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1157 (*conn->service->destroyConnProc) (conn);
1159 /* Notify the security module that this connection is being destroyed */
1160 RXS_DestroyConnection(conn->securityObject, conn);
1162 /* If this is the last connection using the rx_peer struct, set its
1163 * idle time to now. rxi_ReapConnections will reap it if it's still
1164 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1166 MUTEX_ENTER(&rx_peerHashTable_lock);
1167 if (conn->peer->refCount < 2) {
1168 conn->peer->idleWhen = clock_Sec();
1169 if (conn->peer->refCount < 1) {
1170 conn->peer->refCount = 1;
1171 if (rx_stats_active) {
1172 MUTEX_ENTER(&rx_stats_mutex);
1173 rxi_lowPeerRefCount++;
1174 MUTEX_EXIT(&rx_stats_mutex);
1178 conn->peer->refCount--;
1179 MUTEX_EXIT(&rx_peerHashTable_lock);
1181 if (rx_stats_active)
1183 if (conn->type == RX_SERVER_CONNECTION)
1184 rx_atomic_dec(&rx_stats.nServerConns);
1186 rx_atomic_dec(&rx_stats.nClientConns);
1189 if (conn->specific) {
1191 for (i = 0; i < conn->nSpecific; i++) {
1192 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1193 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1194 conn->specific[i] = NULL;
1196 free(conn->specific);
1198 conn->specific = NULL;
1199 conn->nSpecific = 0;
1200 #endif /* !KERNEL */
1202 MUTEX_DESTROY(&conn->conn_call_lock);
1203 MUTEX_DESTROY(&conn->conn_data_lock);
1204 CV_DESTROY(&conn->conn_call_cv);
1206 rxi_FreeConnection(conn);
1209 /* Destroy the specified connection */
1211 rxi_DestroyConnection(struct rx_connection *conn)
1213 MUTEX_ENTER(&rx_connHashTable_lock);
1214 rxi_DestroyConnectionNoLock(conn);
1215 /* conn should be at the head of the cleanup list */
1216 if (conn == rx_connCleanup_list) {
1217 rx_connCleanup_list = rx_connCleanup_list->next;
1218 MUTEX_EXIT(&rx_connHashTable_lock);
1219 rxi_CleanupConnection(conn);
1221 #ifdef RX_ENABLE_LOCKS
1223 MUTEX_EXIT(&rx_connHashTable_lock);
1225 #endif /* RX_ENABLE_LOCKS */
1229 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1231 struct rx_connection **conn_ptr;
1233 struct rx_packet *packet;
1240 MUTEX_ENTER(&conn->conn_data_lock);
1241 MUTEX_ENTER(&rx_refcnt_mutex);
1242 if (conn->refCount > 0)
1245 if (rx_stats_active) {
1246 MUTEX_ENTER(&rx_stats_mutex);
1247 rxi_lowConnRefCount++;
1248 MUTEX_EXIT(&rx_stats_mutex);
1252 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1253 /* Busy; wait till the last guy before proceeding */
1254 MUTEX_EXIT(&rx_refcnt_mutex);
1255 MUTEX_EXIT(&conn->conn_data_lock);
1260 /* If the client previously called rx_NewCall, but it is still
1261 * waiting, treat this as a running call, and wait to destroy the
1262 * connection later when the call completes. */
1263 if ((conn->type == RX_CLIENT_CONNECTION)
1264 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1265 conn->flags |= RX_CONN_DESTROY_ME;
1266 MUTEX_EXIT(&conn->conn_data_lock);
1270 MUTEX_EXIT(&rx_refcnt_mutex);
1271 MUTEX_EXIT(&conn->conn_data_lock);
1273 /* Check for extant references to this connection */
1274 MUTEX_ENTER(&conn->conn_call_lock);
1275 for (i = 0; i < RX_MAXCALLS; i++) {
1276 struct rx_call *call = conn->call[i];
1279 if (conn->type == RX_CLIENT_CONNECTION) {
1280 MUTEX_ENTER(&call->lock);
1281 if (call->delayedAckEvent) {
1282 /* Push the final acknowledgment out now--there
1283 * won't be a subsequent call to acknowledge the
1284 * last reply packets */
1285 rxi_CancelDelayedAckEvent(call);
1286 if (call->state == RX_STATE_PRECALL
1287 || call->state == RX_STATE_ACTIVE) {
1288 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1293 MUTEX_EXIT(&call->lock);
1297 MUTEX_EXIT(&conn->conn_call_lock);
1299 #ifdef RX_ENABLE_LOCKS
1301 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1302 MUTEX_EXIT(&conn->conn_data_lock);
1304 /* Someone is accessing a packet right now. */
1308 #endif /* RX_ENABLE_LOCKS */
1311 /* Don't destroy the connection if there are any call
1312 * structures still in use */
1313 MUTEX_ENTER(&conn->conn_data_lock);
1314 conn->flags |= RX_CONN_DESTROY_ME;
1315 MUTEX_EXIT(&conn->conn_data_lock);
1320 if (conn->natKeepAliveEvent) {
1321 rxi_NatKeepAliveOff(conn);
1324 if (conn->delayedAbortEvent) {
1325 rxevent_Cancel(&conn->delayedAbortEvent);
1326 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1328 MUTEX_ENTER(&conn->conn_data_lock);
1329 rxi_SendConnectionAbort(conn, packet, 0, 1);
1330 MUTEX_EXIT(&conn->conn_data_lock);
1331 rxi_FreePacket(packet);
1335 /* Remove from connection hash table before proceeding */
1337 &rx_connHashTable[CONN_HASH
1338 (peer->host, peer->port, conn->cid, conn->epoch,
1340 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1341 if (*conn_ptr == conn) {
1342 *conn_ptr = conn->next;
1346 /* if the conn that we are destroying was the last connection, then we
1347 * clear rxLastConn as well */
1348 if (rxLastConn == conn)
1351 /* Make sure the connection is completely reset before deleting it. */
1352 /* get rid of pending events that could zap us later */
1353 rxevent_Cancel(&conn->challengeEvent);
1354 rxevent_Cancel(&conn->checkReachEvent);
1355 rxevent_Cancel(&conn->natKeepAliveEvent);
1357 /* Add the connection to the list of destroyed connections that
1358 * need to be cleaned up. This is necessary to avoid deadlocks
1359 * in the routines we call to inform others that this connection is
1360 * being destroyed. */
1361 conn->next = rx_connCleanup_list;
1362 rx_connCleanup_list = conn;
1365 /* Externally available version */
1367 rx_DestroyConnection(struct rx_connection *conn)
1372 rxi_DestroyConnection(conn);
1377 rx_GetConnection(struct rx_connection *conn)
1382 MUTEX_ENTER(&rx_refcnt_mutex);
1384 MUTEX_EXIT(&rx_refcnt_mutex);
1388 #ifdef RX_ENABLE_LOCKS
1389 /* Wait for the transmit queue to no longer be busy.
1390 * requires the call->lock to be held */
1392 rxi_WaitforTQBusy(struct rx_call *call) {
1393 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1394 call->flags |= RX_CALL_TQ_WAIT;
1396 MUTEX_ASSERT(&call->lock);
1397 CV_WAIT(&call->cv_tq, &call->lock);
1399 if (call->tqWaiters == 0) {
1400 call->flags &= ~RX_CALL_TQ_WAIT;
1407 rxi_WakeUpTransmitQueue(struct rx_call *call)
1409 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1410 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1411 call, call->tqWaiters, call->flags));
1412 #ifdef RX_ENABLE_LOCKS
1413 MUTEX_ASSERT(&call->lock);
1414 CV_BROADCAST(&call->cv_tq);
1415 #else /* RX_ENABLE_LOCKS */
1416 osi_rxWakeup(&call->tq);
1417 #endif /* RX_ENABLE_LOCKS */
1421 /* Start a new rx remote procedure call, on the specified connection.
1422 * If wait is set to 1, wait for a free call channel; otherwise return
1423 * 0. Maxtime gives the maximum number of seconds this call may take,
1424 * after rx_NewCall returns. After this time interval, a call to any
1425 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1426 * For fine grain locking, we hold the conn_call_lock in order to
1427 * to ensure that we don't get signalle after we found a call in an active
1428 * state and before we go to sleep.
1431 rx_NewCall(struct rx_connection *conn)
1433 int i, wait, ignoreBusy = 1;
1434 struct rx_call *call;
1435 struct clock queueTime;
1436 afs_uint32 leastBusy = 0;
1440 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1443 clock_GetTime(&queueTime);
1445 * Check if there are others waiting for a new call.
1446 * If so, let them go first to avoid starving them.
1447 * This is a fairly simple scheme, and might not be
1448 * a complete solution for large numbers of waiters.
1450 * makeCallWaiters keeps track of the number of
1451 * threads waiting to make calls and the
1452 * RX_CONN_MAKECALL_WAITING flag bit is used to
1453 * indicate that there are indeed calls waiting.
1454 * The flag is set when the waiter is incremented.
1455 * It is only cleared when makeCallWaiters is 0.
1456 * This prevents us from accidently destroying the
1457 * connection while it is potentially about to be used.
1459 MUTEX_ENTER(&conn->conn_call_lock);
1460 MUTEX_ENTER(&conn->conn_data_lock);
1461 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1462 conn->flags |= RX_CONN_MAKECALL_WAITING;
1463 conn->makeCallWaiters++;
1464 MUTEX_EXIT(&conn->conn_data_lock);
1466 #ifdef RX_ENABLE_LOCKS
1467 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1471 MUTEX_ENTER(&conn->conn_data_lock);
1472 conn->makeCallWaiters--;
1473 if (conn->makeCallWaiters == 0)
1474 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1477 /* We are now the active thread in rx_NewCall */
1478 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1479 MUTEX_EXIT(&conn->conn_data_lock);
1484 for (i = 0; i < RX_MAXCALLS; i++) {
1485 call = conn->call[i];
1487 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1488 /* we're not ignoring busy call slots; only look at the
1489 * call slot that is the "least" busy */
1493 if (call->state == RX_STATE_DALLY) {
1494 MUTEX_ENTER(&call->lock);
1495 if (call->state == RX_STATE_DALLY) {
1496 if (ignoreBusy && conn->lastBusy[i]) {
1497 /* if we're ignoring busy call slots, skip any ones that
1498 * have lastBusy set */
1499 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1500 leastBusy = conn->lastBusy[i];
1502 MUTEX_EXIT(&call->lock);
1507 * We are setting the state to RX_STATE_RESET to
1508 * ensure that no one else will attempt to use this
1509 * call once we drop the conn->conn_call_lock and
1510 * call->lock. We must drop the conn->conn_call_lock
1511 * before calling rxi_ResetCall because the process
1512 * of clearing the transmit queue can block for an
1513 * extended period of time. If we block while holding
1514 * the conn->conn_call_lock, then all rx_EndCall
1515 * processing will block as well. This has a detrimental
1516 * effect on overall system performance.
1518 call->state = RX_STATE_RESET;
1519 (*call->callNumber)++;
1520 MUTEX_EXIT(&conn->conn_call_lock);
1521 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1522 rxi_ResetCall(call, 0);
1523 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1527 * If we failed to be able to safely obtain the
1528 * conn->conn_call_lock we will have to drop the
1529 * call->lock to avoid a deadlock. When the call->lock
1530 * is released the state of the call can change. If it
1531 * is no longer RX_STATE_RESET then some other thread is
1534 MUTEX_EXIT(&call->lock);
1535 MUTEX_ENTER(&conn->conn_call_lock);
1536 MUTEX_ENTER(&call->lock);
1538 if (call->state == RX_STATE_RESET)
1542 * If we get here it means that after dropping
1543 * the conn->conn_call_lock and call->lock that
1544 * the call is no longer ours. If we can't find
1545 * a free call in the remaining slots we should
1546 * not go immediately to RX_CONN_MAKECALL_WAITING
1547 * because by dropping the conn->conn_call_lock
1548 * we have given up synchronization with rx_EndCall.
1549 * Instead, cycle through one more time to see if
1550 * we can find a call that can call our own.
1552 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1555 MUTEX_EXIT(&call->lock);
1558 if (ignoreBusy && conn->lastBusy[i]) {
1559 /* if we're ignoring busy call slots, skip any ones that
1560 * have lastBusy set */
1561 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1562 leastBusy = conn->lastBusy[i];
1567 /* rxi_NewCall returns with mutex locked */
1568 call = rxi_NewCall(conn, i);
1569 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1573 if (i < RX_MAXCALLS) {
1574 conn->lastBusy[i] = 0;
1575 call->flags &= ~RX_CALL_PEER_BUSY;
1580 if (leastBusy && ignoreBusy) {
1581 /* we didn't find a useable call slot, but we did see at least one
1582 * 'busy' slot; look again and only use a slot with the 'least
1588 MUTEX_ENTER(&conn->conn_data_lock);
1589 conn->flags |= RX_CONN_MAKECALL_WAITING;
1590 conn->makeCallWaiters++;
1591 MUTEX_EXIT(&conn->conn_data_lock);
1593 #ifdef RX_ENABLE_LOCKS
1594 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1598 MUTEX_ENTER(&conn->conn_data_lock);
1599 conn->makeCallWaiters--;
1600 if (conn->makeCallWaiters == 0)
1601 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1602 MUTEX_EXIT(&conn->conn_data_lock);
1604 /* Client is initially in send mode */
1605 call->state = RX_STATE_ACTIVE;
1606 call->error = conn->error;
1608 call->app.mode = RX_MODE_ERROR;
1610 call->app.mode = RX_MODE_SENDING;
1612 #ifdef AFS_RXERRQ_ENV
1613 /* remember how many network errors the peer has when we started, so if
1614 * more errors are encountered after the call starts, we know the other endpoint won't be
1615 * responding to us */
1616 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1619 /* remember start time for call in case we have hard dead time limit */
1620 call->queueTime = queueTime;
1621 clock_GetTime(&call->startTime);
1622 call->app.bytesSent = 0;
1623 call->app.bytesRcvd = 0;
1625 /* Turn on busy protocol. */
1626 rxi_KeepAliveOn(call);
1628 /* Attempt MTU discovery */
1629 rxi_GrowMTUOn(call);
1632 * We are no longer the active thread in rx_NewCall
1634 MUTEX_ENTER(&conn->conn_data_lock);
1635 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1636 MUTEX_EXIT(&conn->conn_data_lock);
1639 * Wake up anyone else who might be giving us a chance to
1640 * run (see code above that avoids resource starvation).
1642 #ifdef RX_ENABLE_LOCKS
1643 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1644 osi_Panic("rx_NewCall call about to be used without an empty tq");
1647 CV_BROADCAST(&conn->conn_call_cv);
1651 MUTEX_EXIT(&conn->conn_call_lock);
1652 MUTEX_EXIT(&call->lock);
1655 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1660 rxi_HasActiveCalls(struct rx_connection *aconn)
1663 struct rx_call *tcall;
1667 for (i = 0; i < RX_MAXCALLS; i++) {
1668 if ((tcall = aconn->call[i])) {
1669 if ((tcall->state == RX_STATE_ACTIVE)
1670 || (tcall->state == RX_STATE_PRECALL)) {
1681 rxi_GetCallNumberVector(struct rx_connection *aconn,
1682 afs_int32 * aint32s)
1685 struct rx_call *tcall;
1689 MUTEX_ENTER(&aconn->conn_call_lock);
1690 for (i = 0; i < RX_MAXCALLS; i++) {
1691 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1692 aint32s[i] = aconn->callNumber[i] + 1;
1694 aint32s[i] = aconn->callNumber[i];
1696 MUTEX_EXIT(&aconn->conn_call_lock);
1702 rxi_SetCallNumberVector(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 aconn->callNumber[i] = aint32s[i] - 1;
1715 aconn->callNumber[i] = aint32s[i];
1717 MUTEX_EXIT(&aconn->conn_call_lock);
1722 /* Advertise a new service. A service is named locally by a UDP port
1723 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1726 char *serviceName; Name for identification purposes (e.g. the
1727 service name might be used for probing for
1730 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1731 char *serviceName, struct rx_securityClass **securityObjects,
1732 int nSecurityObjects,
1733 afs_int32(*serviceProc) (struct rx_call * acall))
1735 osi_socket socket = OSI_NULLSOCKET;
1736 struct rx_service *tservice;
1742 if (serviceId == 0) {
1744 "rx_NewService: service id for service %s is not non-zero.\n",
1751 "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",
1759 tservice = rxi_AllocService();
1762 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1764 for (i = 0; i < RX_MAX_SERVICES; i++) {
1765 struct rx_service *service = rx_services[i];
1767 if (port == service->servicePort && host == service->serviceHost) {
1768 if (service->serviceId == serviceId) {
1769 /* The identical service has already been
1770 * installed; if the caller was intending to
1771 * change the security classes used by this
1772 * service, he/she loses. */
1774 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1775 serviceName, serviceId, service->serviceName);
1777 rxi_FreeService(tservice);
1780 /* Different service, same port: re-use the socket
1781 * which is bound to the same port */
1782 socket = service->socket;
1785 if (socket == OSI_NULLSOCKET) {
1786 /* If we don't already have a socket (from another
1787 * service on same port) get a new one */
1788 socket = rxi_GetHostUDPSocket(host, port);
1789 if (socket == OSI_NULLSOCKET) {
1791 rxi_FreeService(tservice);
1796 service->socket = socket;
1797 service->serviceHost = host;
1798 service->servicePort = port;
1799 service->serviceId = serviceId;
1800 service->serviceName = serviceName;
1801 service->nSecurityObjects = nSecurityObjects;
1802 service->securityObjects = securityObjects;
1803 service->minProcs = 0;
1804 service->maxProcs = 1;
1805 service->idleDeadTime = 60;
1806 service->idleDeadErr = 0;
1807 service->connDeadTime = rx_connDeadTime;
1808 service->executeRequestProc = serviceProc;
1809 service->checkReach = 0;
1810 service->nSpecific = 0;
1811 service->specific = NULL;
1812 rx_services[i] = service; /* not visible until now */
1818 rxi_FreeService(tservice);
1819 (osi_Msg "rx_NewService: cannot support > %d services\n",
1824 /* Set configuration options for all of a service's security objects */
1827 rx_SetSecurityConfiguration(struct rx_service *service,
1828 rx_securityConfigVariables type,
1832 for (i = 0; i<service->nSecurityObjects; i++) {
1833 if (service->securityObjects[i]) {
1834 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1842 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1843 struct rx_securityClass **securityObjects, int nSecurityObjects,
1844 afs_int32(*serviceProc) (struct rx_call * acall))
1846 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1849 /* Generic request processing loop. This routine should be called
1850 * by the implementation dependent rx_ServerProc. If socketp is
1851 * non-null, it will be set to the file descriptor that this thread
1852 * is now listening on. If socketp is null, this routine will never
1855 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1857 struct rx_call *call;
1859 struct rx_service *tservice = NULL;
1866 call = rx_GetCall(threadID, tservice, socketp);
1867 if (socketp && *socketp != OSI_NULLSOCKET) {
1868 /* We are now a listener thread */
1874 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1875 #ifdef RX_ENABLE_LOCKS
1877 #endif /* RX_ENABLE_LOCKS */
1878 afs_termState = AFSOP_STOP_AFS;
1879 afs_osi_Wakeup(&afs_termState);
1880 #ifdef RX_ENABLE_LOCKS
1882 #endif /* RX_ENABLE_LOCKS */
1887 /* if server is restarting( typically smooth shutdown) then do not
1888 * allow any new calls.
1891 if (rx_tranquil && (call != NULL)) {
1895 MUTEX_ENTER(&call->lock);
1897 rxi_CallError(call, RX_RESTARTING);
1898 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1900 MUTEX_EXIT(&call->lock);
1905 tservice = call->conn->service;
1907 if (tservice->beforeProc)
1908 (*tservice->beforeProc) (call);
1910 code = tservice->executeRequestProc(call);
1912 if (tservice->afterProc)
1913 (*tservice->afterProc) (call, code);
1915 rx_EndCall(call, code);
1917 if (tservice->postProc)
1918 (*tservice->postProc) (code);
1920 if (rx_stats_active) {
1921 MUTEX_ENTER(&rx_stats_mutex);
1923 MUTEX_EXIT(&rx_stats_mutex);
1930 rx_WakeupServerProcs(void)
1932 struct rx_serverQueueEntry *np, *tqp;
1933 struct opr_queue *cursor;
1937 MUTEX_ENTER(&rx_serverPool_lock);
1939 #ifdef RX_ENABLE_LOCKS
1940 if (rx_waitForPacket)
1941 CV_BROADCAST(&rx_waitForPacket->cv);
1942 #else /* RX_ENABLE_LOCKS */
1943 if (rx_waitForPacket)
1944 osi_rxWakeup(rx_waitForPacket);
1945 #endif /* RX_ENABLE_LOCKS */
1946 MUTEX_ENTER(&freeSQEList_lock);
1947 for (np = rx_FreeSQEList; np; np = tqp) {
1948 tqp = *(struct rx_serverQueueEntry **)np;
1949 #ifdef RX_ENABLE_LOCKS
1950 CV_BROADCAST(&np->cv);
1951 #else /* RX_ENABLE_LOCKS */
1953 #endif /* RX_ENABLE_LOCKS */
1955 MUTEX_EXIT(&freeSQEList_lock);
1956 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1957 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1958 #ifdef RX_ENABLE_LOCKS
1959 CV_BROADCAST(&np->cv);
1960 #else /* RX_ENABLE_LOCKS */
1962 #endif /* RX_ENABLE_LOCKS */
1964 MUTEX_EXIT(&rx_serverPool_lock);
1969 * One thing that seems to happen is that all the server threads get
1970 * tied up on some empty or slow call, and then a whole bunch of calls
1971 * arrive at once, using up the packet pool, so now there are more
1972 * empty calls. The most critical resources here are server threads
1973 * and the free packet pool. The "doreclaim" code seems to help in
1974 * general. I think that eventually we arrive in this state: there
1975 * are lots of pending calls which do have all their packets present,
1976 * so they won't be reclaimed, are multi-packet calls, so they won't
1977 * be scheduled until later, and thus are tying up most of the free
1978 * packet pool for a very long time.
1980 * 1. schedule multi-packet calls if all the packets are present.
1981 * Probably CPU-bound operation, useful to return packets to pool.
1982 * Do what if there is a full window, but the last packet isn't here?
1983 * 3. preserve one thread which *only* runs "best" calls, otherwise
1984 * it sleeps and waits for that type of call.
1985 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1986 * the current dataquota business is badly broken. The quota isn't adjusted
1987 * to reflect how many packets are presently queued for a running call.
1988 * So, when we schedule a queued call with a full window of packets queued
1989 * up for it, that *should* free up a window full of packets for other 2d-class
1990 * calls to be able to use from the packet pool. But it doesn't.
1992 * NB. Most of the time, this code doesn't run -- since idle server threads
1993 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1994 * as a new call arrives.
1996 /* Sleep until a call arrives. Returns a pointer to the call, ready
1997 * for an rx_Read. */
1998 #ifdef RX_ENABLE_LOCKS
2000 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2002 struct rx_serverQueueEntry *sq;
2003 struct rx_call *call = (struct rx_call *)0;
2004 struct rx_service *service = NULL;
2006 MUTEX_ENTER(&freeSQEList_lock);
2008 if ((sq = rx_FreeSQEList)) {
2009 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2010 MUTEX_EXIT(&freeSQEList_lock);
2011 } else { /* otherwise allocate a new one and return that */
2012 MUTEX_EXIT(&freeSQEList_lock);
2013 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2014 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2015 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2018 MUTEX_ENTER(&rx_serverPool_lock);
2019 if (cur_service != NULL) {
2020 ReturnToServerPool(cur_service);
2023 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2024 struct rx_call *tcall, *choice2 = NULL;
2025 struct opr_queue *cursor;
2027 /* Scan for eligible incoming calls. A call is not eligible
2028 * if the maximum number of calls for its service type are
2029 * already executing */
2030 /* One thread will process calls FCFS (to prevent starvation),
2031 * while the other threads may run ahead looking for calls which
2032 * have all their input data available immediately. This helps
2033 * keep threads from blocking, waiting for data from the client. */
2034 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2035 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2037 service = tcall->conn->service;
2038 if (!QuotaOK(service)) {
2041 MUTEX_ENTER(&rx_pthread_mutex);
2042 if (tno == rxi_fcfs_thread_num
2043 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2044 MUTEX_EXIT(&rx_pthread_mutex);
2045 /* If we're the fcfs thread , then we'll just use
2046 * this call. If we haven't been able to find an optimal
2047 * choice, and we're at the end of the list, then use a
2048 * 2d choice if one has been identified. Otherwise... */
2049 call = (choice2 ? choice2 : tcall);
2050 service = call->conn->service;
2052 MUTEX_EXIT(&rx_pthread_mutex);
2053 if (!opr_queue_IsEmpty(&tcall->rq)) {
2054 struct rx_packet *rp;
2055 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2057 if (rp->header.seq == 1) {
2059 || (rp->header.flags & RX_LAST_PACKET)) {
2061 } else if (rxi_2dchoice && !choice2
2062 && !(tcall->flags & RX_CALL_CLEARED)
2063 && (tcall->rprev > rxi_HardAckRate)) {
2073 ReturnToServerPool(service);
2079 opr_queue_Remove(&call->entry);
2080 MUTEX_EXIT(&rx_serverPool_lock);
2081 MUTEX_ENTER(&call->lock);
2083 if (call->flags & RX_CALL_WAIT_PROC) {
2084 call->flags &= ~RX_CALL_WAIT_PROC;
2085 rx_atomic_dec(&rx_nWaiting);
2088 if (call->state != RX_STATE_PRECALL || call->error) {
2089 MUTEX_EXIT(&call->lock);
2090 MUTEX_ENTER(&rx_serverPool_lock);
2091 ReturnToServerPool(service);
2096 if (opr_queue_IsEmpty(&call->rq)
2097 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2098 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2100 CLEAR_CALL_QUEUE_LOCK(call);
2103 /* If there are no eligible incoming calls, add this process
2104 * to the idle server queue, to wait for one */
2108 *socketp = OSI_NULLSOCKET;
2110 sq->socketp = socketp;
2111 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2112 #ifndef AFS_AIX41_ENV
2113 rx_waitForPacket = sq;
2114 #endif /* AFS_AIX41_ENV */
2116 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2118 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2119 MUTEX_EXIT(&rx_serverPool_lock);
2120 return (struct rx_call *)0;
2123 } while (!(call = sq->newcall)
2124 && !(socketp && *socketp != OSI_NULLSOCKET));
2125 MUTEX_EXIT(&rx_serverPool_lock);
2127 MUTEX_ENTER(&call->lock);
2133 MUTEX_ENTER(&freeSQEList_lock);
2134 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2135 rx_FreeSQEList = sq;
2136 MUTEX_EXIT(&freeSQEList_lock);
2139 clock_GetTime(&call->startTime);
2140 call->state = RX_STATE_ACTIVE;
2141 call->app.mode = RX_MODE_RECEIVING;
2142 #ifdef RX_KERNEL_TRACE
2143 if (ICL_SETACTIVE(afs_iclSetp)) {
2144 int glockOwner = ISAFS_GLOCK();
2147 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2148 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2155 rxi_calltrace(RX_CALL_START, call);
2156 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2157 call->conn->service->servicePort, call->conn->service->serviceId,
2160 MUTEX_EXIT(&call->lock);
2161 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2163 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2168 #else /* RX_ENABLE_LOCKS */
2170 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2172 struct rx_serverQueueEntry *sq;
2173 struct rx_call *call = (struct rx_call *)0, *choice2;
2174 struct rx_service *service = NULL;
2178 MUTEX_ENTER(&freeSQEList_lock);
2180 if ((sq = rx_FreeSQEList)) {
2181 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2182 MUTEX_EXIT(&freeSQEList_lock);
2183 } else { /* otherwise allocate a new one and return that */
2184 MUTEX_EXIT(&freeSQEList_lock);
2185 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2186 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2187 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2189 MUTEX_ENTER(&sq->lock);
2191 if (cur_service != NULL) {
2192 cur_service->nRequestsRunning--;
2193 MUTEX_ENTER(&rx_quota_mutex);
2194 if (cur_service->nRequestsRunning < cur_service->minProcs)
2197 MUTEX_EXIT(&rx_quota_mutex);
2199 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2200 struct rx_call *tcall;
2201 struct opr_queue *cursor;
2202 /* Scan for eligible incoming calls. A call is not eligible
2203 * if the maximum number of calls for its service type are
2204 * already executing */
2205 /* One thread will process calls FCFS (to prevent starvation),
2206 * while the other threads may run ahead looking for calls which
2207 * have all their input data available immediately. This helps
2208 * keep threads from blocking, waiting for data from the client. */
2209 choice2 = (struct rx_call *)0;
2210 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2211 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2212 service = tcall->conn->service;
2213 if (QuotaOK(service)) {
2214 MUTEX_ENTER(&rx_pthread_mutex);
2215 /* XXX - If tcall->entry.next is NULL, then we're no longer
2216 * on a queue at all. This shouldn't happen. */
2217 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2218 MUTEX_EXIT(&rx_pthread_mutex);
2219 /* If we're the fcfs thread, then we'll just use
2220 * this call. If we haven't been able to find an optimal
2221 * choice, and we're at the end of the list, then use a
2222 * 2d choice if one has been identified. Otherwise... */
2223 call = (choice2 ? choice2 : tcall);
2224 service = call->conn->service;
2226 MUTEX_EXIT(&rx_pthread_mutex);
2227 if (!opr_queue_IsEmpty(&tcall->rq)) {
2228 struct rx_packet *rp;
2229 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2231 if (rp->header.seq == 1
2233 || (rp->header.flags & RX_LAST_PACKET))) {
2235 } else if (rxi_2dchoice && !choice2
2236 && !(tcall->flags & RX_CALL_CLEARED)
2237 && (tcall->rprev > rxi_HardAckRate)) {
2250 opr_queue_Remove(&call->entry);
2251 /* we can't schedule a call if there's no data!!! */
2252 /* send an ack if there's no data, if we're missing the
2253 * first packet, or we're missing something between first
2254 * and last -- there's a "hole" in the incoming data. */
2255 if (opr_queue_IsEmpty(&call->rq)
2256 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2257 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2258 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2260 call->flags &= (~RX_CALL_WAIT_PROC);
2261 service->nRequestsRunning++;
2262 /* just started call in minProcs pool, need fewer to maintain
2264 MUTEX_ENTER(&rx_quota_mutex);
2265 if (service->nRequestsRunning <= service->minProcs)
2268 MUTEX_EXIT(&rx_quota_mutex);
2269 rx_atomic_dec(&rx_nWaiting);
2270 /* MUTEX_EXIT(&call->lock); */
2272 /* If there are no eligible incoming calls, add this process
2273 * to the idle server queue, to wait for one */
2276 *socketp = OSI_NULLSOCKET;
2278 sq->socketp = socketp;
2279 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2283 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2285 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2286 return (struct rx_call *)0;
2289 } while (!(call = sq->newcall)
2290 && !(socketp && *socketp != OSI_NULLSOCKET));
2292 MUTEX_EXIT(&sq->lock);
2294 MUTEX_ENTER(&freeSQEList_lock);
2295 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2296 rx_FreeSQEList = sq;
2297 MUTEX_EXIT(&freeSQEList_lock);
2300 clock_GetTime(&call->startTime);
2301 call->state = RX_STATE_ACTIVE;
2302 call->app.mode = RX_MODE_RECEIVING;
2303 #ifdef RX_KERNEL_TRACE
2304 if (ICL_SETACTIVE(afs_iclSetp)) {
2305 int glockOwner = ISAFS_GLOCK();
2308 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2309 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2316 rxi_calltrace(RX_CALL_START, call);
2317 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2318 call->conn->service->servicePort, call->conn->service->serviceId,
2321 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2328 #endif /* RX_ENABLE_LOCKS */
2332 /* Establish a procedure to be called when a packet arrives for a
2333 * call. This routine will be called at most once after each call,
2334 * and will also be called if there is an error condition on the or
2335 * the call is complete. Used by multi rx to build a selection
2336 * function which determines which of several calls is likely to be a
2337 * good one to read from.
2338 * NOTE: the way this is currently implemented it is probably only a
2339 * good idea to (1) use it immediately after a newcall (clients only)
2340 * and (2) only use it once. Other uses currently void your warranty
2343 rx_SetArrivalProc(struct rx_call *call,
2344 void (*proc) (struct rx_call * call,
2347 void * handle, int arg)
2349 call->arrivalProc = proc;
2350 call->arrivalProcHandle = handle;
2351 call->arrivalProcArg = arg;
2354 /* Call is finished (possibly prematurely). Return rc to the peer, if
2355 * appropriate, and return the final error code from the conversation
2359 rx_EndCall(struct rx_call *call, afs_int32 rc)
2361 struct rx_connection *conn = call->conn;
2365 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2366 call, rc, call->error, call->abortCode));
2369 MUTEX_ENTER(&call->lock);
2371 if (rc == 0 && call->error == 0) {
2372 call->abortCode = 0;
2373 call->abortCount = 0;
2376 call->arrivalProc = (void (*)())0;
2377 if (rc && call->error == 0) {
2378 rxi_CallError(call, rc);
2379 call->app.mode = RX_MODE_ERROR;
2380 /* Send an abort message to the peer if this error code has
2381 * only just been set. If it was set previously, assume the
2382 * peer has already been sent the error code or will request it
2384 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2386 if (conn->type == RX_SERVER_CONNECTION) {
2387 /* Make sure reply or at least dummy reply is sent */
2388 if (call->app.mode == RX_MODE_RECEIVING) {
2389 MUTEX_EXIT(&call->lock);
2390 rxi_WriteProc(call, 0, 0);
2391 MUTEX_ENTER(&call->lock);
2393 if (call->app.mode == RX_MODE_SENDING) {
2394 MUTEX_EXIT(&call->lock);
2395 rxi_FlushWrite(call);
2396 MUTEX_ENTER(&call->lock);
2398 rxi_calltrace(RX_CALL_END, call);
2399 /* Call goes to hold state until reply packets are acknowledged */
2400 if (call->tfirst + call->nSoftAcked < call->tnext) {
2401 call->state = RX_STATE_HOLD;
2403 call->state = RX_STATE_DALLY;
2404 rxi_ClearTransmitQueue(call, 0);
2405 rxi_rto_cancel(call);
2406 rxi_CancelKeepAliveEvent(call);
2408 } else { /* Client connection */
2410 /* Make sure server receives input packets, in the case where
2411 * no reply arguments are expected */
2413 if ((call->app.mode == RX_MODE_SENDING)
2414 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2415 MUTEX_EXIT(&call->lock);
2416 (void)rxi_ReadProc(call, &dummy, 1);
2417 MUTEX_ENTER(&call->lock);
2420 /* If we had an outstanding delayed ack, be nice to the server
2421 * and force-send it now.
2423 if (call->delayedAckEvent) {
2424 rxi_CancelDelayedAckEvent(call);
2425 rxi_SendDelayedAck(NULL, call, NULL, 0);
2428 /* We need to release the call lock since it's lower than the
2429 * conn_call_lock and we don't want to hold the conn_call_lock
2430 * over the rx_ReadProc call. The conn_call_lock needs to be held
2431 * here for the case where rx_NewCall is perusing the calls on
2432 * the connection structure. We don't want to signal until
2433 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2434 * have checked this call, found it active and by the time it
2435 * goes to sleep, will have missed the signal.
2437 MUTEX_EXIT(&call->lock);
2438 MUTEX_ENTER(&conn->conn_call_lock);
2439 MUTEX_ENTER(&call->lock);
2441 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2442 conn->lastBusy[call->channel] = 0;
2445 MUTEX_ENTER(&conn->conn_data_lock);
2446 conn->flags |= RX_CONN_BUSY;
2447 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2448 MUTEX_EXIT(&conn->conn_data_lock);
2449 #ifdef RX_ENABLE_LOCKS
2450 CV_BROADCAST(&conn->conn_call_cv);
2455 #ifdef RX_ENABLE_LOCKS
2457 MUTEX_EXIT(&conn->conn_data_lock);
2459 #endif /* RX_ENABLE_LOCKS */
2460 call->state = RX_STATE_DALLY;
2462 error = call->error;
2464 /* currentPacket, nLeft, and NFree must be zeroed here, because
2465 * ResetCall cannot: ResetCall may be called at splnet(), in the
2466 * kernel version, and may interrupt the macros rx_Read or
2467 * rx_Write, which run at normal priority for efficiency. */
2468 if (call->app.currentPacket) {
2469 #ifdef RX_TRACK_PACKETS
2470 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2472 rxi_FreePacket(call->app.currentPacket);
2473 call->app.currentPacket = (struct rx_packet *)0;
2476 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2478 /* Free any packets from the last call to ReadvProc/WritevProc */
2479 #ifdef RXDEBUG_PACKET
2481 #endif /* RXDEBUG_PACKET */
2482 rxi_FreePackets(0, &call->app.iovq);
2483 MUTEX_EXIT(&call->lock);
2485 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2486 if (conn->type == RX_CLIENT_CONNECTION) {
2487 MUTEX_ENTER(&conn->conn_data_lock);
2488 conn->flags &= ~RX_CONN_BUSY;
2489 MUTEX_EXIT(&conn->conn_data_lock);
2490 MUTEX_EXIT(&conn->conn_call_lock);
2494 * Map errors to the local host's errno.h format.
2496 error = ntoh_syserr_conv(error);
2498 /* If the caller said the call failed with some error, we had better
2499 * return an error code. */
2500 osi_Assert(!rc || error);
2504 #if !defined(KERNEL)
2506 /* Call this routine when shutting down a server or client (especially
2507 * clients). This will allow Rx to gracefully garbage collect server
2508 * connections, and reduce the number of retries that a server might
2509 * make to a dead client.
2510 * This is not quite right, since some calls may still be ongoing and
2511 * we can't lock them to destroy them. */
2515 struct rx_connection **conn_ptr, **conn_end;
2518 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2519 return; /* Already shutdown. */
2521 rxi_DeleteCachedConnections();
2522 if (rx_connHashTable) {
2523 MUTEX_ENTER(&rx_connHashTable_lock);
2524 for (conn_ptr = &rx_connHashTable[0], conn_end =
2525 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2527 struct rx_connection *conn, *next;
2528 for (conn = *conn_ptr; conn; conn = next) {
2530 if (conn->type == RX_CLIENT_CONNECTION) {
2531 MUTEX_ENTER(&rx_refcnt_mutex);
2533 MUTEX_EXIT(&rx_refcnt_mutex);
2534 #ifdef RX_ENABLE_LOCKS
2535 rxi_DestroyConnectionNoLock(conn);
2536 #else /* RX_ENABLE_LOCKS */
2537 rxi_DestroyConnection(conn);
2538 #endif /* RX_ENABLE_LOCKS */
2542 #ifdef RX_ENABLE_LOCKS
2543 while (rx_connCleanup_list) {
2544 struct rx_connection *conn;
2545 conn = rx_connCleanup_list;
2546 rx_connCleanup_list = rx_connCleanup_list->next;
2547 MUTEX_EXIT(&rx_connHashTable_lock);
2548 rxi_CleanupConnection(conn);
2549 MUTEX_ENTER(&rx_connHashTable_lock);
2551 MUTEX_EXIT(&rx_connHashTable_lock);
2552 #endif /* RX_ENABLE_LOCKS */
2557 afs_winsockCleanup();
2563 /* if we wakeup packet waiter too often, can get in loop with two
2564 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2566 rxi_PacketsUnWait(void)
2568 if (!rx_waitingForPackets) {
2572 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2573 return; /* still over quota */
2576 rx_waitingForPackets = 0;
2577 #ifdef RX_ENABLE_LOCKS
2578 CV_BROADCAST(&rx_waitingForPackets_cv);
2580 osi_rxWakeup(&rx_waitingForPackets);
2586 /* ------------------Internal interfaces------------------------- */
2588 /* Return this process's service structure for the
2589 * specified socket and service */
2590 static struct rx_service *
2591 rxi_FindService(osi_socket socket, u_short serviceId)
2593 struct rx_service **sp;
2594 for (sp = &rx_services[0]; *sp; sp++) {
2595 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2601 #ifdef RXDEBUG_PACKET
2602 #ifdef KDUMP_RX_LOCK
2603 static struct rx_call_rx_lock *rx_allCallsp = 0;
2605 static struct rx_call *rx_allCallsp = 0;
2607 #endif /* RXDEBUG_PACKET */
2609 /* Allocate a call structure, for the indicated channel of the
2610 * supplied connection. The mode and state of the call must be set by
2611 * the caller. Returns the call with mutex locked. */
2612 static struct rx_call *
2613 rxi_NewCall(struct rx_connection *conn, int channel)
2615 struct rx_call *call;
2616 #ifdef RX_ENABLE_LOCKS
2617 struct rx_call *cp; /* Call pointer temp */
2618 struct opr_queue *cursor;
2621 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2623 /* Grab an existing call structure, or allocate a new one.
2624 * Existing call structures are assumed to have been left reset by
2626 MUTEX_ENTER(&rx_freeCallQueue_lock);
2628 #ifdef RX_ENABLE_LOCKS
2630 * EXCEPT that the TQ might not yet be cleared out.
2631 * Skip over those with in-use TQs.
2634 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2635 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2636 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2642 #else /* RX_ENABLE_LOCKS */
2643 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2644 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2645 #endif /* RX_ENABLE_LOCKS */
2646 opr_queue_Remove(&call->entry);
2647 if (rx_stats_active)
2648 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2649 MUTEX_EXIT(&rx_freeCallQueue_lock);
2650 MUTEX_ENTER(&call->lock);
2651 CLEAR_CALL_QUEUE_LOCK(call);
2652 #ifdef RX_ENABLE_LOCKS
2653 /* Now, if TQ wasn't cleared earlier, do it now. */
2654 rxi_WaitforTQBusy(call);
2655 if (call->flags & RX_CALL_TQ_CLEARME) {
2656 rxi_ClearTransmitQueue(call, 1);
2657 /*queue_Init(&call->tq);*/
2659 #endif /* RX_ENABLE_LOCKS */
2660 /* Bind the call to its connection structure */
2662 rxi_ResetCall(call, 1);
2665 call = rxi_Alloc(sizeof(struct rx_call));
2666 #ifdef RXDEBUG_PACKET
2667 call->allNextp = rx_allCallsp;
2668 rx_allCallsp = call;
2670 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2671 #else /* RXDEBUG_PACKET */
2672 rx_atomic_inc(&rx_stats.nCallStructs);
2673 #endif /* RXDEBUG_PACKET */
2675 MUTEX_EXIT(&rx_freeCallQueue_lock);
2676 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2677 MUTEX_ENTER(&call->lock);
2678 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2679 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2680 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2682 /* Initialize once-only items */
2683 opr_queue_Init(&call->tq);
2684 opr_queue_Init(&call->rq);
2685 opr_queue_Init(&call->app.iovq);
2686 #ifdef RXDEBUG_PACKET
2687 call->rqc = call->tqc = call->iovqc = 0;
2688 #endif /* RXDEBUG_PACKET */
2689 /* Bind the call to its connection structure (prereq for reset) */
2691 rxi_ResetCall(call, 1);
2693 call->channel = channel;
2694 call->callNumber = &conn->callNumber[channel];
2695 call->rwind = conn->rwind[channel];
2696 call->twind = conn->twind[channel];
2697 /* Note that the next expected call number is retained (in
2698 * conn->callNumber[i]), even if we reallocate the call structure
2700 conn->call[channel] = call;
2701 /* if the channel's never been used (== 0), we should start at 1, otherwise
2702 * the call number is valid from the last time this channel was used */
2703 if (*call->callNumber == 0)
2704 *call->callNumber = 1;
2709 /* A call has been inactive long enough that so we can throw away
2710 * state, including the call structure, which is placed on the call
2713 * call->lock amd rx_refcnt_mutex are held upon entry.
2714 * haveCTLock is set when called from rxi_ReapConnections.
2716 * return 1 if the call is freed, 0 if not.
2719 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2721 int channel = call->channel;
2722 struct rx_connection *conn = call->conn;
2723 u_char state = call->state;
2726 * We are setting the state to RX_STATE_RESET to
2727 * ensure that no one else will attempt to use this
2728 * call once we drop the refcnt lock. We must drop
2729 * the refcnt lock before calling rxi_ResetCall
2730 * because it cannot be held across acquiring the
2731 * freepktQ lock. NewCall does the same.
2733 call->state = RX_STATE_RESET;
2734 MUTEX_EXIT(&rx_refcnt_mutex);
2735 rxi_ResetCall(call, 0);
2737 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2739 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2740 (*call->callNumber)++;
2742 if (call->conn->call[channel] == call)
2743 call->conn->call[channel] = 0;
2744 MUTEX_EXIT(&conn->conn_call_lock);
2747 * We couldn't obtain the conn_call_lock so we can't
2748 * disconnect the call from the connection. Set the
2749 * call state to dally so that the call can be reused.
2751 MUTEX_ENTER(&rx_refcnt_mutex);
2752 call->state = RX_STATE_DALLY;
2756 MUTEX_ENTER(&rx_freeCallQueue_lock);
2757 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2758 #ifdef RX_ENABLE_LOCKS
2759 /* A call may be free even though its transmit queue is still in use.
2760 * Since we search the call list from head to tail, put busy calls at
2761 * the head of the list, and idle calls at the tail.
2763 if (call->flags & RX_CALL_TQ_BUSY)
2764 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2766 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2767 #else /* RX_ENABLE_LOCKS */
2768 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2769 #endif /* RX_ENABLE_LOCKS */
2770 if (rx_stats_active)
2771 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2772 MUTEX_EXIT(&rx_freeCallQueue_lock);
2774 /* Destroy the connection if it was previously slated for
2775 * destruction, i.e. the Rx client code previously called
2776 * rx_DestroyConnection (client connections), or
2777 * rxi_ReapConnections called the same routine (server
2778 * connections). Only do this, however, if there are no
2779 * outstanding calls. Note that for fine grain locking, there appears
2780 * to be a deadlock in that rxi_FreeCall has a call locked and
2781 * DestroyConnectionNoLock locks each call in the conn. But note a
2782 * few lines up where we have removed this call from the conn.
2783 * If someone else destroys a connection, they either have no
2784 * call lock held or are going through this section of code.
2786 MUTEX_ENTER(&conn->conn_data_lock);
2787 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2788 MUTEX_ENTER(&rx_refcnt_mutex);
2790 MUTEX_EXIT(&rx_refcnt_mutex);
2791 MUTEX_EXIT(&conn->conn_data_lock);
2792 #ifdef RX_ENABLE_LOCKS
2794 rxi_DestroyConnectionNoLock(conn);
2796 rxi_DestroyConnection(conn);
2797 #else /* RX_ENABLE_LOCKS */
2798 rxi_DestroyConnection(conn);
2799 #endif /* RX_ENABLE_LOCKS */
2801 MUTEX_EXIT(&conn->conn_data_lock);
2803 MUTEX_ENTER(&rx_refcnt_mutex);
2807 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2808 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2811 rxi_Alloc(size_t size)
2815 if (rx_stats_active) {
2816 rx_atomic_add(&rxi_Allocsize, (int) size);
2817 rx_atomic_inc(&rxi_Alloccnt);
2821 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2822 afs_osi_Alloc_NoSleep(size);
2827 osi_Panic("rxi_Alloc error");
2833 rxi_Free(void *addr, size_t size)
2835 if (rx_stats_active) {
2836 rx_atomic_sub(&rxi_Allocsize, (int) size);
2837 rx_atomic_dec(&rxi_Alloccnt);
2839 osi_Free(addr, size);
2843 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2845 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2846 struct rx_peer *next = NULL;
2850 MUTEX_ENTER(&rx_peerHashTable_lock);
2852 peer_ptr = &rx_peerHashTable[0];
2853 peer_end = &rx_peerHashTable[rx_hashTableSize];
2856 for ( ; peer_ptr < peer_end; peer_ptr++) {
2859 for ( ; peer; peer = next) {
2861 if (host == peer->host)
2866 hashIndex = PEER_HASH(host, port);
2867 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2868 if ((peer->host == host) && (peer->port == port))
2873 MUTEX_ENTER(&rx_peerHashTable_lock);
2878 MUTEX_EXIT(&rx_peerHashTable_lock);
2880 MUTEX_ENTER(&peer->peer_lock);
2881 /* We don't handle dropping below min, so don't */
2882 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2883 peer->ifMTU=MIN(mtu, peer->ifMTU);
2884 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2885 /* if we tweaked this down, need to tune our peer MTU too */
2886 peer->MTU = MIN(peer->MTU, peer->natMTU);
2887 /* if we discovered a sub-1500 mtu, degrade */
2888 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2889 peer->maxDgramPackets = 1;
2890 /* We no longer have valid peer packet information */
2891 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2892 peer->maxPacketSize = 0;
2893 MUTEX_EXIT(&peer->peer_lock);
2895 MUTEX_ENTER(&rx_peerHashTable_lock);
2897 if (host && !port) {
2899 /* pick up where we left off */
2903 MUTEX_EXIT(&rx_peerHashTable_lock);
2906 #ifdef AFS_RXERRQ_ENV
2908 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2910 int hashIndex = PEER_HASH(host, port);
2911 struct rx_peer *peer;
2913 MUTEX_ENTER(&rx_peerHashTable_lock);
2915 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2916 if (peer->host == host && peer->port == port) {
2922 MUTEX_EXIT(&rx_peerHashTable_lock);
2925 rx_atomic_inc(&peer->neterrs);
2926 MUTEX_ENTER(&peer->peer_lock);
2927 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2928 peer->last_err_type = err->ee_type;
2929 peer->last_err_code = err->ee_code;
2930 MUTEX_EXIT(&peer->peer_lock);
2932 MUTEX_ENTER(&rx_peerHashTable_lock);
2934 MUTEX_EXIT(&rx_peerHashTable_lock);
2939 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2941 # ifdef AFS_ADAPT_PMTU
2942 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2943 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2947 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2948 switch (err->ee_code) {
2949 case ICMP_NET_UNREACH:
2950 case ICMP_HOST_UNREACH:
2951 case ICMP_PORT_UNREACH:
2954 rxi_SetPeerDead(err, addr, port);
2961 rxi_TranslateICMP(int type, int code)
2964 case ICMP_DEST_UNREACH:
2966 case ICMP_NET_UNREACH:
2967 return "Destination Net Unreachable";
2968 case ICMP_HOST_UNREACH:
2969 return "Destination Host Unreachable";
2970 case ICMP_PROT_UNREACH:
2971 return "Destination Protocol Unreachable";
2972 case ICMP_PORT_UNREACH:
2973 return "Destination Port Unreachable";
2975 return "Destination Net Prohibited";
2977 return "Destination Host Prohibited";
2983 #endif /* AFS_RXERRQ_ENV */
2986 * Get the last network error for a connection
2988 * A "network error" here means an error retrieved from ICMP, or some other
2989 * mechanism outside of Rx that informs us of errors in network reachability.
2991 * If a peer associated with the given Rx connection has received a network
2992 * error recently, this function allows the caller to know what error
2993 * specifically occurred. This can be useful to know, since e.g. ICMP errors
2994 * can cause calls to that peer to be quickly aborted. So, this function can
2995 * help see why a call was aborted due to network errors.
2997 * If we have received traffic from a peer since the last network error, we
2998 * treat that peer as if we had not received an network error for it.
3000 * @param[in] conn The Rx connection to examine
3001 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
3002 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
3003 * @param[out] err_type The type of the last error
3004 * @param[out] err_code The code of the last error
3005 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
3007 * @return If we have an error
3008 * @retval -1 No error to get; 'out' params are undefined
3009 * @retval 0 We have an error; 'out' params contain the last error
3012 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
3013 int *err_code, const char **msg)
3015 #ifdef AFS_RXERRQ_ENV
3016 struct rx_peer *peer = conn->peer;
3017 if (rx_atomic_read(&peer->neterrs)) {
3018 MUTEX_ENTER(&peer->peer_lock);
3019 *err_origin = peer->last_err_origin;
3020 *err_type = peer->last_err_type;
3021 *err_code = peer->last_err_code;
3022 MUTEX_EXIT(&peer->peer_lock);
3025 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3026 *msg = rxi_TranslateICMP(*err_type, *err_code);
3035 /* Find the peer process represented by the supplied (host,port)
3036 * combination. If there is no appropriate active peer structure, a
3037 * new one will be allocated and initialized
3040 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3044 hashIndex = PEER_HASH(host, port);
3045 MUTEX_ENTER(&rx_peerHashTable_lock);
3046 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3047 if ((pp->host == host) && (pp->port == port))
3052 pp = rxi_AllocPeer(); /* This bzero's *pp */
3053 pp->host = host; /* set here or in InitPeerParams is zero */
3055 #ifdef AFS_RXERRQ_ENV
3056 rx_atomic_set(&pp->neterrs, 0);
3058 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3059 opr_queue_Init(&pp->rpcStats);
3060 pp->next = rx_peerHashTable[hashIndex];
3061 rx_peerHashTable[hashIndex] = pp;
3062 rxi_InitPeerParams(pp);
3063 if (rx_stats_active)
3064 rx_atomic_inc(&rx_stats.nPeerStructs);
3070 MUTEX_EXIT(&rx_peerHashTable_lock);
3075 /* Find the connection at (host, port) started at epoch, and with the
3076 * given connection id. Creates the server connection if necessary.
3077 * The type specifies whether a client connection or a server
3078 * connection is desired. In both cases, (host, port) specify the
3079 * peer's (host, pair) pair. Client connections are not made
3080 * automatically by this routine. The parameter socket gives the
3081 * socket descriptor on which the packet was received. This is used,
3082 * in the case of server connections, to check that *new* connections
3083 * come via a valid (port, serviceId). Finally, the securityIndex
3084 * parameter must match the existing index for the connection. If a
3085 * server connection is created, it will be created using the supplied
3086 * index, if the index is valid for this service */
3087 static struct rx_connection *
3088 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3089 u_short port, u_short serviceId, afs_uint32 cid,
3090 afs_uint32 epoch, int type, u_int securityIndex,
3091 int *unknownService)
3093 int hashindex, flag, i;
3094 struct rx_connection *conn;
3095 *unknownService = 0;
3096 hashindex = CONN_HASH(host, port, cid, epoch, type);
3097 MUTEX_ENTER(&rx_connHashTable_lock);
3098 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3099 rx_connHashTable[hashindex],
3102 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3103 && (epoch == conn->epoch)) {
3104 struct rx_peer *pp = conn->peer;
3105 if (securityIndex != conn->securityIndex) {
3106 /* this isn't supposed to happen, but someone could forge a packet
3107 * like this, and there seems to be some CM bug that makes this
3108 * happen from time to time -- in which case, the fileserver
3110 MUTEX_EXIT(&rx_connHashTable_lock);
3111 return (struct rx_connection *)0;
3113 if (pp->host == host && pp->port == port)
3115 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3117 /* So what happens when it's a callback connection? */
3118 if ( /*type == RX_CLIENT_CONNECTION && */
3119 (conn->epoch & 0x80000000))
3123 /* the connection rxLastConn that was used the last time is not the
3124 ** one we are looking for now. Hence, start searching in the hash */
3126 conn = rx_connHashTable[hashindex];
3131 struct rx_service *service;
3132 if (type == RX_CLIENT_CONNECTION) {
3133 MUTEX_EXIT(&rx_connHashTable_lock);
3134 return (struct rx_connection *)0;
3136 service = rxi_FindService(socket, serviceId);
3137 if (!service || (securityIndex >= service->nSecurityObjects)
3138 || (service->securityObjects[securityIndex] == 0)) {
3139 MUTEX_EXIT(&rx_connHashTable_lock);
3140 *unknownService = 1;
3141 return (struct rx_connection *)0;
3143 conn = rxi_AllocConnection(); /* This bzero's the connection */
3144 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3145 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3146 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3147 conn->next = rx_connHashTable[hashindex];
3148 rx_connHashTable[hashindex] = conn;
3149 conn->peer = rxi_FindPeer(host, port, 1);
3150 conn->type = RX_SERVER_CONNECTION;
3151 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3152 conn->epoch = epoch;
3153 conn->cid = cid & RX_CIDMASK;
3154 conn->ackRate = RX_FAST_ACK_RATE;
3155 conn->service = service;
3156 conn->serviceId = serviceId;
3157 conn->securityIndex = securityIndex;
3158 conn->securityObject = service->securityObjects[securityIndex];
3159 conn->nSpecific = 0;
3160 conn->specific = NULL;
3161 rx_SetConnDeadTime(conn, service->connDeadTime);
3162 conn->idleDeadTime = service->idleDeadTime;
3163 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3164 for (i = 0; i < RX_MAXCALLS; i++) {
3165 conn->twind[i] = rx_initSendWindow;
3166 conn->rwind[i] = rx_initReceiveWindow;
3168 /* Notify security object of the new connection */
3169 RXS_NewConnection(conn->securityObject, conn);
3170 /* XXXX Connection timeout? */
3171 if (service->newConnProc)
3172 (*service->newConnProc) (conn);
3173 if (rx_stats_active)
3174 rx_atomic_inc(&rx_stats.nServerConns);
3177 MUTEX_ENTER(&rx_refcnt_mutex);
3179 MUTEX_EXIT(&rx_refcnt_mutex);
3181 rxLastConn = conn; /* store this connection as the last conn used */
3182 MUTEX_EXIT(&rx_connHashTable_lock);
3187 * Timeout a call on a busy call channel if appropriate.
3189 * @param[in] call The busy call.
3191 * @pre 'call' is marked as busy (namely,
3192 * call->conn->lastBusy[call->channel] != 0)
3194 * @pre call->lock is held
3195 * @pre rxi_busyChannelError is nonzero
3197 * @note call->lock is dropped and reacquired
3200 rxi_CheckBusy(struct rx_call *call)
3202 struct rx_connection *conn = call->conn;
3203 int channel = call->channel;
3204 int freechannel = 0;
3207 MUTEX_EXIT(&call->lock);
3209 MUTEX_ENTER(&conn->conn_call_lock);
3211 /* Are there any other call slots on this conn that we should try? Look for
3212 * slots that are empty and are either non-busy, or were marked as busy
3213 * longer than conn->secondsUntilDead seconds before this call started. */
3215 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3217 /* only look at channels that aren't us */
3221 if (conn->lastBusy[i]) {
3222 /* if this channel looked busy too recently, don't look at it */
3223 if (conn->lastBusy[i] >= call->startTime.sec) {
3226 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3231 if (conn->call[i]) {
3232 struct rx_call *tcall = conn->call[i];
3233 MUTEX_ENTER(&tcall->lock);
3234 if (tcall->state == RX_STATE_DALLY) {
3237 MUTEX_EXIT(&tcall->lock);
3243 MUTEX_ENTER(&call->lock);
3245 /* Since the call->lock has been released it is possible that the call may
3246 * no longer be busy (the call channel cannot have been reallocated as we
3247 * haven't dropped the conn_call_lock) Therefore, we must confirm
3248 * that the call state has not changed when deciding whether or not to
3249 * force this application thread to retry by forcing a Timeout error. */
3251 if (freechannel && (call->flags & RX_CALL_PEER_BUSY)) {
3252 /* Since 'freechannel' is set, there exists another channel in this
3253 * rx_conn that the application thread might be able to use. We know
3254 * that we have the correct call since callNumber is unchanged, and we
3255 * know that the call is still busy. So, set the call error state to
3256 * rxi_busyChannelError so the application can retry the request,
3257 * presumably on a less-busy call channel. */
3259 rxi_CallError(call, RX_CALL_BUSY);
3261 MUTEX_EXIT(&conn->conn_call_lock);
3265 * Abort the call if the server is over the busy threshold. This
3266 * can be used without requiring a call structure be initialised,
3267 * or connected to a particular channel
3270 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3271 struct rx_packet *np)
3273 if ((rx_BusyThreshold > 0) &&
3274 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3275 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3276 rx_BusyError, np, 0);
3277 if (rx_stats_active)
3278 rx_atomic_inc(&rx_stats.nBusies);
3285 static_inline struct rx_call *
3286 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3289 struct rx_call *call;
3291 channel = np->header.cid & RX_CHANNELMASK;
3292 MUTEX_ENTER(&conn->conn_call_lock);
3293 call = conn->call[channel];
3294 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3295 MUTEX_EXIT(&conn->conn_call_lock);
3296 if (rx_stats_active)
3297 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3301 MUTEX_ENTER(&call->lock);
3302 MUTEX_EXIT(&conn->conn_call_lock);
3304 if ((call->state == RX_STATE_DALLY)
3305 && np->header.type == RX_PACKET_TYPE_ACK) {
3306 if (rx_stats_active)
3307 rx_atomic_inc(&rx_stats.ignorePacketDally);
3308 MUTEX_EXIT(&call->lock);
3315 static_inline struct rx_call *
3316 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3317 struct rx_connection *conn)
3320 struct rx_call *call;
3322 channel = np->header.cid & RX_CHANNELMASK;
3323 MUTEX_ENTER(&conn->conn_call_lock);
3324 call = conn->call[channel];
3327 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3328 MUTEX_EXIT(&conn->conn_call_lock);
3332 call = rxi_NewCall(conn, channel); /* returns locked call */
3333 *call->callNumber = np->header.callNumber;
3334 MUTEX_EXIT(&conn->conn_call_lock);
3336 call->state = RX_STATE_PRECALL;
3337 clock_GetTime(&call->queueTime);
3338 call->app.bytesSent = 0;
3339 call->app.bytesRcvd = 0;
3340 rxi_KeepAliveOn(call);
3345 if (np->header.callNumber == conn->callNumber[channel]) {
3346 MUTEX_ENTER(&call->lock);
3347 MUTEX_EXIT(&conn->conn_call_lock);
3351 if (np->header.callNumber < conn->callNumber[channel]) {
3352 MUTEX_EXIT(&conn->conn_call_lock);
3353 if (rx_stats_active)
3354 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3358 MUTEX_ENTER(&call->lock);
3359 MUTEX_EXIT(&conn->conn_call_lock);
3361 /* Wait until the transmit queue is idle before deciding
3362 * whether to reset the current call. Chances are that the
3363 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3366 #ifdef RX_ENABLE_LOCKS
3367 if (call->state == RX_STATE_ACTIVE && !call->error) {
3368 rxi_WaitforTQBusy(call);
3369 /* If we entered error state while waiting,
3370 * must call rxi_CallError to permit rxi_ResetCall
3371 * to processed when the tqWaiter count hits zero.
3374 rxi_CallError(call, call->error);
3375 MUTEX_EXIT(&call->lock);
3379 #endif /* RX_ENABLE_LOCKS */
3380 /* If the new call cannot be taken right now send a busy and set
3381 * the error condition in this call, so that it terminates as
3382 * quickly as possible */
3383 if (call->state == RX_STATE_ACTIVE) {
3384 rxi_CallError(call, RX_CALL_DEAD);
3385 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3387 MUTEX_EXIT(&call->lock);
3391 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3392 MUTEX_EXIT(&call->lock);
3396 rxi_ResetCall(call, 0);
3397 /* The conn_call_lock is not held but no one else should be
3398 * using this call channel while we are processing this incoming
3399 * packet. This assignment should be safe.
3401 *call->callNumber = np->header.callNumber;
3402 call->state = RX_STATE_PRECALL;
3403 clock_GetTime(&call->queueTime);
3404 call->app.bytesSent = 0;
3405 call->app.bytesRcvd = 0;
3406 rxi_KeepAliveOn(call);
3412 /* There are two packet tracing routines available for testing and monitoring
3413 * Rx. One is called just after every packet is received and the other is
3414 * called just before every packet is sent. Received packets, have had their
3415 * headers decoded, and packets to be sent have not yet had their headers
3416 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3417 * containing the network address. Both can be modified. The return value, if
3418 * non-zero, indicates that the packet should be dropped. */
3420 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3421 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3423 /* A packet has been received off the interface. Np is the packet, socket is
3424 * the socket number it was received from (useful in determining which service
3425 * this packet corresponds to), and (host, port) reflect the host,port of the
3426 * sender. This call returns the packet to the caller if it is finished with
3427 * it, rather than de-allocating it, just as a small performance hack */
3430 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3431 afs_uint32 host, u_short port, int *tnop,
3432 struct rx_call **newcallp)
3434 struct rx_call *call;
3435 struct rx_connection *conn;
3437 int unknownService = 0;
3441 struct rx_packet *tnp;
3444 /* We don't print out the packet until now because (1) the time may not be
3445 * accurate enough until now in the lwp implementation (rx_Listener only gets
3446 * the time after the packet is read) and (2) from a protocol point of view,
3447 * this is the first time the packet has been seen */
3448 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3449 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3450 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3451 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3452 np->header.epoch, np->header.cid, np->header.callNumber,
3453 np->header.seq, np->header.flags, np));
3456 /* Account for connectionless packets */
3457 if (rx_stats_active &&
3458 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3459 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3460 struct rx_peer *peer;
3462 /* Try to look up the peer structure, but don't create one */
3463 peer = rxi_FindPeer(host, port, 0);
3465 /* Since this may not be associated with a connection, it may have
3466 * no refCount, meaning we could race with ReapConnections
3469 if (peer && (peer->refCount > 0)) {
3470 #ifdef AFS_RXERRQ_ENV
3471 if (rx_atomic_read(&peer->neterrs)) {
3472 rx_atomic_set(&peer->neterrs, 0);
3475 MUTEX_ENTER(&peer->peer_lock);
3476 peer->bytesReceived += np->length;
3477 MUTEX_EXIT(&peer->peer_lock);
3481 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3482 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3485 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3486 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3489 /* If an input tracer function is defined, call it with the packet and
3490 * network address. Note this function may modify its arguments. */
3491 if (rx_justReceived) {
3492 struct sockaddr_in addr;
3494 addr.sin_family = AF_INET;
3495 addr.sin_port = port;
3496 addr.sin_addr.s_addr = host;
3497 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3498 addr.sin_len = sizeof(addr);
3499 #endif /* AFS_OSF_ENV */
3500 drop = (*rx_justReceived) (np, &addr);
3501 /* drop packet if return value is non-zero */
3504 port = addr.sin_port; /* in case fcn changed addr */
3505 host = addr.sin_addr.s_addr;
3509 /* If packet was not sent by the client, then *we* must be the client */
3510 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3511 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3513 /* Find the connection (or fabricate one, if we're the server & if
3514 * necessary) associated with this packet */
3516 rxi_FindConnection(socket, host, port, np->header.serviceId,
3517 np->header.cid, np->header.epoch, type,
3518 np->header.securityIndex, &unknownService);
3520 /* To avoid having 2 connections just abort at each other,
3521 don't abort an abort. */
3523 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3524 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3529 #ifdef AFS_RXERRQ_ENV
3530 if (rx_atomic_read(&conn->peer->neterrs)) {
3531 rx_atomic_set(&conn->peer->neterrs, 0);
3535 /* If we're doing statistics, then account for the incoming packet */
3536 if (rx_stats_active) {
3537 MUTEX_ENTER(&conn->peer->peer_lock);
3538 conn->peer->bytesReceived += np->length;
3539 MUTEX_EXIT(&conn->peer->peer_lock);
3542 /* If the connection is in an error state, send an abort packet and ignore
3543 * the incoming packet */
3545 /* Don't respond to an abort packet--we don't want loops! */
3546 MUTEX_ENTER(&conn->conn_data_lock);
3547 if (np->header.type != RX_PACKET_TYPE_ABORT)
3548 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3549 putConnection(conn);
3550 MUTEX_EXIT(&conn->conn_data_lock);
3554 /* Check for connection-only requests (i.e. not call specific). */
3555 if (np->header.callNumber == 0) {
3556 switch (np->header.type) {
3557 case RX_PACKET_TYPE_ABORT: {
3558 /* What if the supplied error is zero? */
3559 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3560 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3561 rxi_ConnectionError(conn, errcode);
3562 putConnection(conn);
3565 case RX_PACKET_TYPE_CHALLENGE:
3566 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3567 putConnection(conn);
3569 case RX_PACKET_TYPE_RESPONSE:
3570 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3571 putConnection(conn);
3573 case RX_PACKET_TYPE_PARAMS:
3574 case RX_PACKET_TYPE_PARAMS + 1:
3575 case RX_PACKET_TYPE_PARAMS + 2:
3576 /* ignore these packet types for now */
3577 putConnection(conn);
3581 /* Should not reach here, unless the peer is broken: send an
3583 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3584 MUTEX_ENTER(&conn->conn_data_lock);
3585 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3586 putConnection(conn);
3587 MUTEX_EXIT(&conn->conn_data_lock);
3592 if (type == RX_SERVER_CONNECTION)
3593 call = rxi_ReceiveServerCall(socket, np, conn);
3595 call = rxi_ReceiveClientCall(np, conn);
3598 putConnection(conn);
3602 MUTEX_ASSERT(&call->lock);
3603 /* Set remote user defined status from packet */
3604 call->remoteStatus = np->header.userStatus;
3606 /* Now do packet type-specific processing */
3607 switch (np->header.type) {
3608 case RX_PACKET_TYPE_DATA:
3609 /* If we're a client, and receiving a response, then all the packets
3610 * we transmitted packets are implicitly acknowledged. */
3611 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3612 rxi_AckAllInTransmitQueue(call);
3614 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3617 case RX_PACKET_TYPE_ACK:
3618 /* Respond immediately to ack packets requesting acknowledgement
3620 if (np->header.flags & RX_REQUEST_ACK) {
3622 (void)rxi_SendCallAbort(call, 0, 1, 0);
3624 (void)rxi_SendAck(call, 0, np->header.serial,
3625 RX_ACK_PING_RESPONSE, 1);
3627 np = rxi_ReceiveAckPacket(call, np, 1);
3629 case RX_PACKET_TYPE_ABORT: {
3630 /* An abort packet: reset the call, passing the error up to the user. */
3631 /* What if error is zero? */
3632 /* What if the error is -1? the application will treat it as a timeout. */
3633 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3634 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3635 rxi_CallError(call, errdata);
3636 MUTEX_EXIT(&call->lock);
3637 putConnection(conn);
3638 return np; /* xmitting; drop packet */
3640 case RX_PACKET_TYPE_BUSY: {
3641 struct clock busyTime;
3643 clock_GetTime(&busyTime);
3645 MUTEX_EXIT(&call->lock);
3647 MUTEX_ENTER(&conn->conn_call_lock);
3648 MUTEX_ENTER(&call->lock);
3649 conn->lastBusy[call->channel] = busyTime.sec;
3650 call->flags |= RX_CALL_PEER_BUSY;
3651 MUTEX_EXIT(&call->lock);
3652 MUTEX_EXIT(&conn->conn_call_lock);
3654 putConnection(conn);
3658 case RX_PACKET_TYPE_ACKALL:
3659 /* All packets acknowledged, so we can drop all packets previously
3660 * readied for sending */
3661 rxi_AckAllInTransmitQueue(call);
3664 /* Should not reach here, unless the peer is broken: send an abort
3666 rxi_CallError(call, RX_PROTOCOL_ERROR);
3667 np = rxi_SendCallAbort(call, np, 1, 0);
3670 /* Note when this last legitimate packet was received, for keep-alive
3671 * processing. Note, we delay getting the time until now in the hope that
3672 * the packet will be delivered to the user before any get time is required
3673 * (if not, then the time won't actually be re-evaluated here). */
3674 call->lastReceiveTime = clock_Sec();
3675 /* we've received a legit packet, so the channel is not busy */
3676 call->flags &= ~RX_CALL_PEER_BUSY;
3677 MUTEX_EXIT(&call->lock);
3678 putConnection(conn);
3682 /* return true if this is an "interesting" connection from the point of view
3683 of someone trying to debug the system */
3685 rxi_IsConnInteresting(struct rx_connection *aconn)
3688 struct rx_call *tcall;
3690 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3693 for (i = 0; i < RX_MAXCALLS; i++) {
3694 tcall = aconn->call[i];
3696 if ((tcall->state == RX_STATE_PRECALL)
3697 || (tcall->state == RX_STATE_ACTIVE))
3699 if ((tcall->app.mode == RX_MODE_SENDING)
3700 || (tcall->app.mode == RX_MODE_RECEIVING))
3708 /* if this is one of the last few packets AND it wouldn't be used by the
3709 receiving call to immediately satisfy a read request, then drop it on
3710 the floor, since accepting it might prevent a lock-holding thread from
3711 making progress in its reading. If a call has been cleared while in
3712 the precall state then ignore all subsequent packets until the call
3713 is assigned to a thread. */
3716 TooLow(struct rx_packet *ap, struct rx_call *acall)
3720 MUTEX_ENTER(&rx_quota_mutex);
3721 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3722 && (acall->state == RX_STATE_PRECALL))
3723 || ((rx_nFreePackets < rxi_dataQuota + 2)
3724 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3725 && (acall->flags & RX_CALL_READER_WAIT)))) {
3728 MUTEX_EXIT(&rx_quota_mutex);
3734 * Clear the attach wait flag on a connection and proceed.
3736 * Any processing waiting for a connection to be attached should be
3737 * unblocked. We clear the flag and do any other needed tasks.
3740 * the conn to unmark waiting for attach
3742 * @pre conn's conn_data_lock must be locked before calling this function
3746 rxi_ConnClearAttachWait(struct rx_connection *conn)
3748 /* Indicate that rxi_CheckReachEvent is no longer running by
3749 * clearing the flag. Must be atomic under conn_data_lock to
3750 * avoid a new call slipping by: rxi_CheckConnReach holds
3751 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3753 conn->flags &= ~RX_CONN_ATTACHWAIT;
3754 if (conn->flags & RX_CONN_NAT_PING) {
3755 conn->flags &= ~RX_CONN_NAT_PING;
3756 rxi_ScheduleNatKeepAliveEvent(conn);
3761 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3763 struct rx_connection *conn = arg1;
3764 struct rx_call *acall = arg2;
3765 struct rx_call *call = acall;
3766 struct clock when, now;
3769 MUTEX_ENTER(&conn->conn_data_lock);
3772 rxevent_Put(&conn->checkReachEvent);
3774 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3776 putConnection(conn);
3778 MUTEX_EXIT(&conn->conn_data_lock);
3782 MUTEX_ENTER(&conn->conn_call_lock);
3783 MUTEX_ENTER(&conn->conn_data_lock);
3784 for (i = 0; i < RX_MAXCALLS; i++) {
3785 struct rx_call *tc = conn->call[i];
3786 if (tc && tc->state == RX_STATE_PRECALL) {
3792 rxi_ConnClearAttachWait(conn);
3793 MUTEX_EXIT(&conn->conn_data_lock);
3794 MUTEX_EXIT(&conn->conn_call_lock);
3799 MUTEX_ENTER(&call->lock);
3800 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3802 MUTEX_EXIT(&call->lock);
3804 clock_GetTime(&now);
3806 when.sec += RX_CHECKREACH_TIMEOUT;
3807 MUTEX_ENTER(&conn->conn_data_lock);
3808 if (!conn->checkReachEvent) {
3809 MUTEX_ENTER(&rx_refcnt_mutex);
3811 MUTEX_EXIT(&rx_refcnt_mutex);
3812 conn->checkReachEvent = rxevent_Post(&when, &now,
3813 rxi_CheckReachEvent, conn,
3816 MUTEX_EXIT(&conn->conn_data_lock);
3822 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3824 struct rx_service *service = conn->service;
3825 struct rx_peer *peer = conn->peer;
3826 afs_uint32 now, lastReach;
3828 if (service->checkReach == 0)
3832 MUTEX_ENTER(&peer->peer_lock);
3833 lastReach = peer->lastReachTime;
3834 MUTEX_EXIT(&peer->peer_lock);
3835 if (now - lastReach < RX_CHECKREACH_TTL)
3838 MUTEX_ENTER(&conn->conn_data_lock);
3839 if (conn->flags & RX_CONN_ATTACHWAIT) {
3840 MUTEX_EXIT(&conn->conn_data_lock);
3843 conn->flags |= RX_CONN_ATTACHWAIT;
3844 MUTEX_EXIT(&conn->conn_data_lock);
3845 if (!conn->checkReachEvent)
3846 rxi_CheckReachEvent(NULL, conn, call, 0);
3851 /* try to attach call, if authentication is complete */
3853 TryAttach(struct rx_call *acall, osi_socket socket,
3854 int *tnop, struct rx_call **newcallp,
3857 struct rx_connection *conn = acall->conn;
3859 if (conn->type == RX_SERVER_CONNECTION
3860 && acall->state == RX_STATE_PRECALL) {
3861 /* Don't attach until we have any req'd. authentication. */
3862 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3863 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3864 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3865 /* Note: this does not necessarily succeed; there
3866 * may not any proc available
3869 rxi_ChallengeOn(acall->conn);
3874 /* A data packet has been received off the interface. This packet is
3875 * appropriate to the call (the call is in the right state, etc.). This
3876 * routine can return a packet to the caller, for re-use */
3878 static struct rx_packet *
3879 rxi_ReceiveDataPacket(struct rx_call *call,
3880 struct rx_packet *np, int istack,
3881 osi_socket socket, afs_uint32 host, u_short port,
3882 int *tnop, struct rx_call **newcallp)
3884 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3889 afs_uint32 serial=0, flags=0;
3891 struct rx_packet *tnp;
3892 if (rx_stats_active)
3893 rx_atomic_inc(&rx_stats.dataPacketsRead);
3896 /* If there are no packet buffers, drop this new packet, unless we can find
3897 * packet buffers from inactive calls */
3899 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3900 MUTEX_ENTER(&rx_freePktQ_lock);
3901 rxi_NeedMorePackets = TRUE;
3902 MUTEX_EXIT(&rx_freePktQ_lock);
3903 if (rx_stats_active)
3904 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3905 rxi_calltrace(RX_TRACE_DROP, call);
3906 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3907 /* We used to clear the receive queue here, in an attempt to free
3908 * packets. However this is unsafe if the queue has received a
3909 * soft ACK for the final packet */
3910 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3916 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3917 * packet is one of several packets transmitted as a single
3918 * datagram. Do not send any soft or hard acks until all packets
3919 * in a jumbogram have been processed. Send negative acks right away.
3921 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3922 /* tnp is non-null when there are more packets in the
3923 * current jumbo gram */
3930 seq = np->header.seq;
3931 serial = np->header.serial;
3932 flags = np->header.flags;
3934 /* If the call is in an error state, send an abort message */
3936 return rxi_SendCallAbort(call, np, istack, 0);
3938 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3939 * AFS 3.5 jumbogram. */
3940 if (flags & RX_JUMBO_PACKET) {
3941 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3946 if (np->header.spare != 0) {
3947 MUTEX_ENTER(&call->conn->conn_data_lock);
3948 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3949 MUTEX_EXIT(&call->conn->conn_data_lock);
3952 /* The usual case is that this is the expected next packet */
3953 if (seq == call->rnext) {
3955 /* Check to make sure it is not a duplicate of one already queued */
3956 if (!opr_queue_IsEmpty(&call->rq)
3957 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3958 if (rx_stats_active)
3959 rx_atomic_inc(&rx_stats.dupPacketsRead);
3960 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3961 rxi_CancelDelayedAckEvent(call);
3962 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3968 /* It's the next packet. Stick it on the receive queue
3969 * for this call. Set newPackets to make sure we wake
3970 * the reader once all packets have been processed */
3971 #ifdef RX_TRACK_PACKETS
3972 np->flags |= RX_PKTFLAG_RQ;
3974 opr_queue_Prepend(&call->rq, &np->entry);
3975 #ifdef RXDEBUG_PACKET
3977 #endif /* RXDEBUG_PACKET */
3979 np = NULL; /* We can't use this anymore */
3982 /* If an ack is requested then set a flag to make sure we
3983 * send an acknowledgement for this packet */
3984 if (flags & RX_REQUEST_ACK) {
3985 ackNeeded = RX_ACK_REQUESTED;
3988 /* Keep track of whether we have received the last packet */
3989 if (flags & RX_LAST_PACKET) {
3990 call->flags |= RX_CALL_HAVE_LAST;
3994 /* Check whether we have all of the packets for this call */
3995 if (call->flags & RX_CALL_HAVE_LAST) {
3996 afs_uint32 tseq; /* temporary sequence number */
3997 struct opr_queue *cursor;
3999 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
4000 struct rx_packet *tp;
4002 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4003 if (tseq != tp->header.seq)
4005 if (tp->header.flags & RX_LAST_PACKET) {
4006 call->flags |= RX_CALL_RECEIVE_DONE;
4013 /* Provide asynchronous notification for those who want it
4014 * (e.g. multi rx) */
4015 if (call->arrivalProc) {
4016 (*call->arrivalProc) (call, call->arrivalProcHandle,
4017 call->arrivalProcArg);
4018 call->arrivalProc = (void (*)())0;
4021 /* Update last packet received */
4024 /* If there is no server process serving this call, grab
4025 * one, if available. We only need to do this once. If a
4026 * server thread is available, this thread becomes a server
4027 * thread and the server thread becomes a listener thread. */
4029 TryAttach(call, socket, tnop, newcallp, 0);
4032 /* This is not the expected next packet. */
4034 /* Determine whether this is a new or old packet, and if it's
4035 * a new one, whether it fits into the current receive window.
4036 * Also figure out whether the packet was delivered in sequence.
4037 * We use the prev variable to determine whether the new packet
4038 * is the successor of its immediate predecessor in the
4039 * receive queue, and the missing flag to determine whether
4040 * any of this packets predecessors are missing. */
4042 afs_uint32 prev; /* "Previous packet" sequence number */
4043 struct opr_queue *cursor;
4044 int missing; /* Are any predecessors missing? */
4046 /* If the new packet's sequence number has been sent to the
4047 * application already, then this is a duplicate */
4048 if (seq < call->rnext) {
4049 if (rx_stats_active)
4050 rx_atomic_inc(&rx_stats.dupPacketsRead);
4051 rxi_CancelDelayedAckEvent(call);
4052 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4058 /* If the sequence number is greater than what can be
4059 * accomodated by the current window, then send a negative
4060 * acknowledge and drop the packet */
4061 if ((call->rnext + call->rwind) <= seq) {
4062 rxi_CancelDelayedAckEvent(call);
4063 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4070 /* Look for the packet in the queue of old received packets */
4071 prev = call->rnext - 1;
4073 for (opr_queue_Scan(&call->rq, cursor)) {
4074 struct rx_packet *tp
4075 = opr_queue_Entry(cursor, struct rx_packet, entry);
4077 /*Check for duplicate packet */
4078 if (seq == tp->header.seq) {
4079 if (rx_stats_active)
4080 rx_atomic_inc(&rx_stats.dupPacketsRead);
4081 rxi_CancelDelayedAckEvent(call);
4082 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4088 /* If we find a higher sequence packet, break out and
4089 * insert the new packet here. */
4090 if (seq < tp->header.seq)
4092 /* Check for missing packet */
4093 if (tp->header.seq != prev + 1) {
4097 prev = tp->header.seq;
4100 /* Keep track of whether we have received the last packet. */
4101 if (flags & RX_LAST_PACKET) {
4102 call->flags |= RX_CALL_HAVE_LAST;
4105 /* It's within the window: add it to the the receive queue.
4106 * tp is left by the previous loop either pointing at the
4107 * packet before which to insert the new packet, or at the
4108 * queue head if the queue is empty or the packet should be
4110 #ifdef RX_TRACK_PACKETS
4111 np->flags |= RX_PKTFLAG_RQ;
4113 #ifdef RXDEBUG_PACKET
4115 #endif /* RXDEBUG_PACKET */
4116 opr_queue_InsertBefore(cursor, &np->entry);
4120 /* Check whether we have all of the packets for this call */
4121 if ((call->flags & RX_CALL_HAVE_LAST)
4122 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4123 afs_uint32 tseq; /* temporary sequence number */
4126 for (opr_queue_Scan(&call->rq, cursor)) {
4127 struct rx_packet *tp
4128 = opr_queue_Entry(cursor, struct rx_packet, entry);
4129 if (tseq != tp->header.seq)
4131 if (tp->header.flags & RX_LAST_PACKET) {
4132 call->flags |= RX_CALL_RECEIVE_DONE;
4139 /* We need to send an ack of the packet is out of sequence,
4140 * or if an ack was requested by the peer. */
4141 if (seq != prev + 1 || missing) {
4142 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4143 } else if (flags & RX_REQUEST_ACK) {
4144 ackNeeded = RX_ACK_REQUESTED;
4147 /* Acknowledge the last packet for each call */
4148 if (flags & RX_LAST_PACKET) {
4159 * If the receiver is waiting for an iovec, fill the iovec
4160 * using the data from the receive queue */
4161 if (call->flags & RX_CALL_IOVEC_WAIT) {
4162 didHardAck = rxi_FillReadVec(call, serial);
4163 /* the call may have been aborted */
4172 /* Wakeup the reader if any */
4173 if ((call->flags & RX_CALL_READER_WAIT)
4174 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4175 || (call->iovNext >= call->iovMax)
4176 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4177 call->flags &= ~RX_CALL_READER_WAIT;
4178 #ifdef RX_ENABLE_LOCKS
4179 CV_BROADCAST(&call->cv_rq);
4181 osi_rxWakeup(&call->rq);
4187 * Send an ack when requested by the peer, or once every
4188 * rxi_SoftAckRate packets until the last packet has been
4189 * received. Always send a soft ack for the last packet in
4190 * the server's reply. */
4192 rxi_CancelDelayedAckEvent(call);
4193 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4194 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4195 rxi_CancelDelayedAckEvent(call);
4196 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4197 } else if (call->nSoftAcks) {
4198 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4199 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4201 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4202 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4203 rxi_CancelDelayedAckEvent(call);
4210 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4212 struct rx_peer *peer = conn->peer;
4214 MUTEX_ENTER(&peer->peer_lock);
4215 peer->lastReachTime = clock_Sec();
4216 MUTEX_EXIT(&peer->peer_lock);
4218 MUTEX_ENTER(&conn->conn_data_lock);
4219 if (conn->flags & RX_CONN_ATTACHWAIT) {
4222 rxi_ConnClearAttachWait(conn);
4223 MUTEX_EXIT(&conn->conn_data_lock);
4225 for (i = 0; i < RX_MAXCALLS; i++) {
4226 struct rx_call *call = conn->call[i];
4229 MUTEX_ENTER(&call->lock);
4230 /* tnop can be null if newcallp is null */
4231 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4233 MUTEX_EXIT(&call->lock);
4237 MUTEX_EXIT(&conn->conn_data_lock);
4240 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4242 rx_ack_reason(int reason)
4245 case RX_ACK_REQUESTED:
4247 case RX_ACK_DUPLICATE:
4249 case RX_ACK_OUT_OF_SEQUENCE:
4251 case RX_ACK_EXCEEDS_WINDOW:
4253 case RX_ACK_NOSPACE:
4257 case RX_ACK_PING_RESPONSE:
4270 /* The real smarts of the whole thing. */
4271 static struct rx_packet *
4272 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4275 struct rx_ackPacket *ap;
4277 struct rx_packet *tp;
4278 struct rx_connection *conn = call->conn;
4279 struct rx_peer *peer = conn->peer;
4280 struct opr_queue *cursor;
4281 struct clock now; /* Current time, for RTT calculations */
4289 int newAckCount = 0;
4290 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4291 int pktsize = 0; /* Set if we need to update the peer mtu */
4292 int conn_data_locked = 0;
4294 if (rx_stats_active)
4295 rx_atomic_inc(&rx_stats.ackPacketsRead);
4296 ap = (struct rx_ackPacket *)rx_DataOf(np);
4297 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4299 return np; /* truncated ack packet */
4301 /* depends on ack packet struct */
4302 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4303 first = ntohl(ap->firstPacket);
4304 prev = ntohl(ap->previousPacket);
4305 serial = ntohl(ap->serial);
4308 * Ignore ack packets received out of order while protecting
4309 * against peers that set the previousPacket field to a packet
4310 * serial number instead of a sequence number.
4312 if (first < call->tfirst ||
4313 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4320 if (np->header.flags & RX_SLOW_START_OK) {
4321 call->flags |= RX_CALL_SLOW_START_OK;
4324 if (ap->reason == RX_ACK_PING_RESPONSE)
4325 rxi_UpdatePeerReach(conn, call);
4327 if (conn->lastPacketSizeSeq) {
4328 MUTEX_ENTER(&conn->conn_data_lock);
4329 conn_data_locked = 1;
4330 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4331 pktsize = conn->lastPacketSize;
4332 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4335 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4336 if (!conn_data_locked) {
4337 MUTEX_ENTER(&conn->conn_data_lock);
4338 conn_data_locked = 1;
4340 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4341 /* process mtu ping ack */
4342 pktsize = conn->lastPingSize;
4343 conn->lastPingSizeSer = conn->lastPingSize = 0;
4347 if (conn_data_locked) {
4348 MUTEX_EXIT(&conn->conn_data_lock);
4349 conn_data_locked = 0;
4353 if (rxdebug_active) {
4357 len = _snprintf(msg, sizeof(msg),
4358 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4359 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4360 ntohl(ap->serial), ntohl(ap->previousPacket),
4361 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4362 ap->nAcks, ntohs(ap->bufferSpace) );
4366 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4367 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4371 OutputDebugString(msg);
4373 #else /* AFS_NT40_ENV */
4376 "RACK: reason %x previous %u seq %u serial %u first %u",
4377 ap->reason, ntohl(ap->previousPacket),
4378 (unsigned int)np->header.seq, (unsigned int)serial,
4379 ntohl(ap->firstPacket));
4382 for (offset = 0; offset < nAcks; offset++)
4383 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4388 #endif /* AFS_NT40_ENV */
4391 MUTEX_ENTER(&peer->peer_lock);
4394 * Start somewhere. Can't assume we can send what we can receive,
4395 * but we are clearly receiving.
4397 if (!peer->maxPacketSize)
4398 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4400 if (pktsize > peer->maxPacketSize) {
4401 peer->maxPacketSize = pktsize;
4402 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4403 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4404 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4405 rxi_ScheduleGrowMTUEvent(call, 1);
4410 clock_GetTime(&now);
4412 /* The transmit queue splits into 4 sections.
4414 * The first section is packets which have now been acknowledged
4415 * by a window size change in the ack. These have reached the
4416 * application layer, and may be discarded. These are packets
4417 * with sequence numbers < ap->firstPacket.
4419 * The second section is packets which have sequence numbers in
4420 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4421 * contents of the packet's ack array determines whether these
4422 * packets are acknowledged or not.
4424 * The third section is packets which fall above the range
4425 * addressed in the ack packet. These have not yet been received
4428 * The four section is packets which have not yet been transmitted.
4429 * These packets will have a header.serial of 0.
4432 /* First section - implicitly acknowledged packets that can be
4436 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4437 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4438 struct rx_packet *next;
4440 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4441 call->tfirst = tp->header.seq + 1;
4443 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4445 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4448 #ifdef RX_ENABLE_LOCKS
4449 /* XXX Hack. Because we have to release the global call lock when sending
4450 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4451 * in rxi_Start sending packets out because packets may move to the
4452 * freePacketQueue as result of being here! So we drop these packets until
4453 * we're safely out of the traversing. Really ugly!
4454 * To make it even uglier, if we're using fine grain locking, we can
4455 * set the ack bits in the packets and have rxi_Start remove the packets
4456 * when it's done transmitting.
4458 if (call->flags & RX_CALL_TQ_BUSY) {
4459 tp->flags |= RX_PKTFLAG_ACKED;
4460 call->flags |= RX_CALL_TQ_SOME_ACKED;
4462 #endif /* RX_ENABLE_LOCKS */
4464 opr_queue_Remove(&tp->entry);
4465 #ifdef RX_TRACK_PACKETS
4466 tp->flags &= ~RX_PKTFLAG_TQ;
4468 #ifdef RXDEBUG_PACKET
4470 #endif /* RXDEBUG_PACKET */
4471 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4476 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4478 /* Second section of the queue - packets for which we are receiving
4481 * Go through the explicit acks/nacks and record the results in
4482 * the waiting packets. These are packets that can't be released
4483 * yet, even with a positive acknowledge. This positive
4484 * acknowledge only means the packet has been received by the
4485 * peer, not that it will be retained long enough to be sent to
4486 * the peer's upper level. In addition, reset the transmit timers
4487 * of any missing packets (those packets that must be missing
4488 * because this packet was out of sequence) */
4490 call->nSoftAcked = 0;
4492 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4493 && tp->header.seq < first + nAcks) {
4494 /* Set the acknowledge flag per packet based on the
4495 * information in the ack packet. An acknowlegded packet can
4496 * be downgraded when the server has discarded a packet it
4497 * soacked previously, or when an ack packet is received
4498 * out of sequence. */
4499 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4500 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4502 tp->flags |= RX_PKTFLAG_ACKED;
4503 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4510 } else /* RX_ACK_TYPE_NACK */ {
4511 tp->flags &= ~RX_PKTFLAG_ACKED;
4515 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4518 /* We don't need to take any action with the 3rd or 4th section in the
4519 * queue - they're not addressed by the contents of this ACK packet.
4522 /* If the window has been extended by this acknowledge packet,
4523 * then wakeup a sender waiting in alloc for window space, or try
4524 * sending packets now, if he's been sitting on packets due to
4525 * lack of window space */
4526 if (call->tnext < (call->tfirst + call->twind)) {
4527 #ifdef RX_ENABLE_LOCKS
4528 CV_SIGNAL(&call->cv_twind);
4530 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4531 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4532 osi_rxWakeup(&call->twind);
4535 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4536 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4540 /* if the ack packet has a receivelen field hanging off it,
4541 * update our state */
4542 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4545 /* If the ack packet has a "recommended" size that is less than
4546 * what I am using now, reduce my size to match */
4547 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4548 (int)sizeof(afs_int32), &tSize);
4549 tSize = (afs_uint32) ntohl(tSize);
4550 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4552 /* Get the maximum packet size to send to this peer */
4553 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4555 tSize = (afs_uint32) ntohl(tSize);
4556 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4557 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4559 /* sanity check - peer might have restarted with different params.
4560 * If peer says "send less", dammit, send less... Peer should never
4561 * be unable to accept packets of the size that prior AFS versions would
4562 * send without asking. */
4563 if (peer->maxMTU != tSize) {
4564 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4566 peer->maxMTU = tSize;
4567 peer->MTU = MIN(tSize, peer->MTU);
4568 call->MTU = MIN(call->MTU, tSize);
4571 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4574 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4575 (int)sizeof(afs_int32), &tSize);
4576 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4577 if (tSize < call->twind) { /* smaller than our send */
4578 call->twind = tSize; /* window, we must send less... */
4579 call->ssthresh = MIN(call->twind, call->ssthresh);
4580 call->conn->twind[call->channel] = call->twind;
4583 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4584 * network MTU confused with the loopback MTU. Calculate the
4585 * maximum MTU here for use in the slow start code below.
4587 /* Did peer restart with older RX version? */
4588 if (peer->maxDgramPackets > 1) {
4589 peer->maxDgramPackets = 1;
4591 } else if (np->length >=
4592 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4595 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4596 sizeof(afs_int32), &tSize);
4597 tSize = (afs_uint32) ntohl(tSize);
4599 * As of AFS 3.5 we set the send window to match the receive window.
4601 if (tSize < call->twind) {
4602 call->twind = tSize;
4603 call->conn->twind[call->channel] = call->twind;
4604 call->ssthresh = MIN(call->twind, call->ssthresh);
4605 } else if (tSize > call->twind) {
4606 call->twind = tSize;
4607 call->conn->twind[call->channel] = call->twind;
4611 * As of AFS 3.5, a jumbogram is more than one fixed size
4612 * packet transmitted in a single UDP datagram. If the remote
4613 * MTU is smaller than our local MTU then never send a datagram
4614 * larger than the natural MTU.
4617 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4618 (int)sizeof(afs_int32), &tSize);
4619 maxDgramPackets = (afs_uint32) ntohl(tSize);
4620 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4622 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4623 if (maxDgramPackets > 1) {
4624 peer->maxDgramPackets = maxDgramPackets;
4625 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4627 peer->maxDgramPackets = 1;
4628 call->MTU = peer->natMTU;
4630 } else if (peer->maxDgramPackets > 1) {
4631 /* Restarted with lower version of RX */
4632 peer->maxDgramPackets = 1;
4634 } else if (peer->maxDgramPackets > 1
4635 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4636 /* Restarted with lower version of RX */
4637 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4638 peer->natMTU = OLD_MAX_PACKET_SIZE;
4639 peer->MTU = OLD_MAX_PACKET_SIZE;
4640 peer->maxDgramPackets = 1;
4641 peer->nDgramPackets = 1;
4643 call->MTU = OLD_MAX_PACKET_SIZE;
4648 * Calculate how many datagrams were successfully received after
4649 * the first missing packet and adjust the negative ack counter
4654 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4655 if (call->nNacks < nNacked) {
4656 call->nNacks = nNacked;
4659 call->nAcks += newAckCount;
4663 /* If the packet contained new acknowledgements, rather than just
4664 * being a duplicate of one we have previously seen, then we can restart
4667 if (newAckCount > 0)
4668 rxi_rto_packet_acked(call, istack);
4670 if (call->flags & RX_CALL_FAST_RECOVER) {
4671 if (newAckCount == 0) {
4672 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4674 call->flags &= ~RX_CALL_FAST_RECOVER;
4675 call->cwind = call->nextCwind;
4676 call->nextCwind = 0;
4679 call->nCwindAcks = 0;
4680 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4681 /* Three negative acks in a row trigger congestion recovery */
4682 call->flags |= RX_CALL_FAST_RECOVER;
4683 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4685 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4686 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4687 call->nextCwind = call->ssthresh;
4690 peer->MTU = call->MTU;
4691 peer->cwind = call->nextCwind;
4692 peer->nDgramPackets = call->nDgramPackets;
4694 call->congestSeq = peer->congestSeq;
4696 /* Reset the resend times on the packets that were nacked
4697 * so we will retransmit as soon as the window permits
4701 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4702 struct rx_packet *tp =
4703 opr_queue_Entry(cursor, struct rx_packet, entry);
4705 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4706 tp->flags &= ~RX_PKTFLAG_SENT;
4708 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4713 /* If cwind is smaller than ssthresh, then increase
4714 * the window one packet for each ack we receive (exponential
4716 * If cwind is greater than or equal to ssthresh then increase
4717 * the congestion window by one packet for each cwind acks we
4718 * receive (linear growth). */
4719 if (call->cwind < call->ssthresh) {
4721 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4722 call->nCwindAcks = 0;
4724 call->nCwindAcks += newAckCount;
4725 if (call->nCwindAcks >= call->cwind) {
4726 call->nCwindAcks = 0;
4727 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4731 * If we have received several acknowledgements in a row then
4732 * it is time to increase the size of our datagrams
4734 if ((int)call->nAcks > rx_nDgramThreshold) {
4735 if (peer->maxDgramPackets > 1) {
4736 if (call->nDgramPackets < peer->maxDgramPackets) {
4737 call->nDgramPackets++;
4739 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4740 } else if (call->MTU < peer->maxMTU) {
4741 /* don't upgrade if we can't handle it */
4742 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4743 call->MTU = peer->ifMTU;
4745 call->MTU += peer->natMTU;
4746 call->MTU = MIN(call->MTU, peer->maxMTU);
4753 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4755 /* Servers need to hold the call until all response packets have
4756 * been acknowledged. Soft acks are good enough since clients
4757 * are not allowed to clear their receive queues. */
4758 if (call->state == RX_STATE_HOLD
4759 && call->tfirst + call->nSoftAcked >= call->tnext) {
4760 call->state = RX_STATE_DALLY;
4761 rxi_ClearTransmitQueue(call, 0);
4762 rxi_CancelKeepAliveEvent(call);
4763 } else if (!opr_queue_IsEmpty(&call->tq)) {
4764 rxi_Start(call, istack);
4770 * Schedule a connection abort to be sent after some delay.
4772 * @param[in] conn The connection to send the abort on.
4773 * @param[in] msec The number of milliseconds to wait before sending.
4775 * @pre conn_data_lock must be held
4778 rxi_SendConnectionAbortLater(struct rx_connection *conn, int msec)
4780 struct clock when, now;
4784 if (!conn->delayedAbortEvent) {
4785 clock_GetTime(&now);
4787 clock_Addmsec(&when, msec);
4788 conn->delayedAbortEvent =
4789 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
4793 /* Received a response to a challenge packet */
4794 static struct rx_packet *
4795 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4796 struct rx_packet *np, int istack)
4800 /* Ignore the packet if we're the client */
4801 if (conn->type == RX_CLIENT_CONNECTION)
4804 /* If already authenticated, ignore the packet (it's probably a retry) */
4805 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4808 /* Otherwise, have the security object evaluate the response packet */
4809 error = RXS_CheckResponse(conn->securityObject, conn, np);
4811 /* If the response is invalid, reset the connection, sending
4812 * an abort to the peer. Send the abort with a 1 second delay,
4813 * to avoid a peer hammering us by constantly recreating a
4814 * connection with bad credentials. */
4815 rxi_ConnectionError(conn, error);
4816 MUTEX_ENTER(&conn->conn_data_lock);
4817 rxi_SendConnectionAbortLater(conn, 1000);
4818 MUTEX_EXIT(&conn->conn_data_lock);
4821 /* If the response is valid, any calls waiting to attach
4822 * servers can now do so */
4825 for (i = 0; i < RX_MAXCALLS; i++) {
4826 struct rx_call *call = conn->call[i];
4828 MUTEX_ENTER(&call->lock);
4829 if (call->state == RX_STATE_PRECALL)
4830 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4831 /* tnop can be null if newcallp is null */
4832 MUTEX_EXIT(&call->lock);
4836 /* Update the peer reachability information, just in case
4837 * some calls went into attach-wait while we were waiting
4838 * for authentication..
4840 rxi_UpdatePeerReach(conn, NULL);
4845 /* A client has received an authentication challenge: the security
4846 * object is asked to cough up a respectable response packet to send
4847 * back to the server. The server is responsible for retrying the
4848 * challenge if it fails to get a response. */
4850 static struct rx_packet *
4851 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4852 struct rx_packet *np, int istack)
4856 /* Ignore the challenge if we're the server */
4857 if (conn->type == RX_SERVER_CONNECTION)
4860 /* Ignore the challenge if the connection is otherwise idle; someone's
4861 * trying to use us as an oracle. */
4862 if (!rxi_HasActiveCalls(conn))
4865 /* Send the security object the challenge packet. It is expected to fill
4866 * in the response. */
4867 error = RXS_GetResponse(conn->securityObject, conn, np);
4869 /* If the security object is unable to return a valid response, reset the
4870 * connection and send an abort to the peer. Otherwise send the response
4871 * packet to the peer connection. */
4873 rxi_ConnectionError(conn, error);
4874 MUTEX_ENTER(&conn->conn_data_lock);
4875 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4876 MUTEX_EXIT(&conn->conn_data_lock);
4878 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4879 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4885 /* Find an available server process to service the current request in
4886 * the given call structure. If one isn't available, queue up this
4887 * call so it eventually gets one */
4889 rxi_AttachServerProc(struct rx_call *call,
4890 osi_socket socket, int *tnop,
4891 struct rx_call **newcallp)
4893 struct rx_serverQueueEntry *sq;
4894 struct rx_service *service = call->conn->service;
4897 /* May already be attached */
4898 if (call->state == RX_STATE_ACTIVE)
4901 MUTEX_ENTER(&rx_serverPool_lock);
4903 haveQuota = QuotaOK(service);
4904 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4905 /* If there are no processes available to service this call,
4906 * put the call on the incoming call queue (unless it's
4907 * already on the queue).
4909 #ifdef RX_ENABLE_LOCKS
4911 ReturnToServerPool(service);
4912 #endif /* RX_ENABLE_LOCKS */
4914 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4915 call->flags |= RX_CALL_WAIT_PROC;
4916 rx_atomic_inc(&rx_nWaiting);
4917 rx_atomic_inc(&rx_nWaited);
4918 rxi_calltrace(RX_CALL_ARRIVAL, call);
4919 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4920 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4923 sq = opr_queue_Last(&rx_idleServerQueue,
4924 struct rx_serverQueueEntry, entry);
4926 /* If hot threads are enabled, and both newcallp and sq->socketp
4927 * are non-null, then this thread will process the call, and the
4928 * idle server thread will start listening on this threads socket.
4930 opr_queue_Remove(&sq->entry);
4932 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4935 *sq->socketp = socket;
4936 clock_GetTime(&call->startTime);
4937 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4941 if (call->flags & RX_CALL_WAIT_PROC) {
4942 /* Conservative: I don't think this should happen */
4943 call->flags &= ~RX_CALL_WAIT_PROC;
4944 rx_atomic_dec(&rx_nWaiting);
4945 if (opr_queue_IsOnQueue(&call->entry)) {
4946 opr_queue_Remove(&call->entry);
4949 call->state = RX_STATE_ACTIVE;
4950 call->app.mode = RX_MODE_RECEIVING;
4951 #ifdef RX_KERNEL_TRACE
4953 int glockOwner = ISAFS_GLOCK();
4956 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4957 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4963 if (call->flags & RX_CALL_CLEARED) {
4964 /* send an ack now to start the packet flow up again */
4965 call->flags &= ~RX_CALL_CLEARED;
4966 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4968 #ifdef RX_ENABLE_LOCKS
4971 service->nRequestsRunning++;
4972 MUTEX_ENTER(&rx_quota_mutex);
4973 if (service->nRequestsRunning <= service->minProcs)
4976 MUTEX_EXIT(&rx_quota_mutex);
4980 MUTEX_EXIT(&rx_serverPool_lock);
4983 /* Delay the sending of an acknowledge event for a short while, while
4984 * a new call is being prepared (in the case of a client) or a reply
4985 * is being prepared (in the case of a server). Rather than sending
4986 * an ack packet, an ACKALL packet is sent. */
4988 rxi_AckAll(struct rx_call *call)
4990 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4992 call->flags |= RX_CALL_ACKALL_SENT;
4996 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4999 struct rx_call *call = arg1;
5000 #ifdef RX_ENABLE_LOCKS
5002 MUTEX_ENTER(&call->lock);
5003 if (event == call->delayedAckEvent)
5004 rxevent_Put(&call->delayedAckEvent);
5005 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
5007 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5009 MUTEX_EXIT(&call->lock);
5010 #else /* RX_ENABLE_LOCKS */
5012 rxevent_Put(&call->delayedAckEvent);
5013 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5014 #endif /* RX_ENABLE_LOCKS */
5017 #ifdef RX_ENABLE_LOCKS
5018 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5019 * clearing them out.
5022 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5024 struct opr_queue *cursor;
5027 for (opr_queue_Scan(&call->tq, cursor)) {
5029 = opr_queue_Entry(cursor, struct rx_packet, entry);
5031 p->flags |= RX_PKTFLAG_ACKED;
5036 call->flags |= RX_CALL_TQ_CLEARME;
5037 call->flags |= RX_CALL_TQ_SOME_ACKED;
5040 rxi_rto_cancel(call);
5042 call->tfirst = call->tnext;
5043 call->nSoftAcked = 0;
5045 if (call->flags & RX_CALL_FAST_RECOVER) {
5046 call->flags &= ~RX_CALL_FAST_RECOVER;
5047 call->cwind = call->nextCwind;
5048 call->nextCwind = 0;
5051 CV_SIGNAL(&call->cv_twind);
5053 #endif /* RX_ENABLE_LOCKS */
5056 * Acknowledge the whole transmit queue.
5058 * If we're running without locks, or the transmit queue isn't busy, then
5059 * we can just clear the queue now. Otherwise, we have to mark all of the
5060 * packets as acknowledged, and let rxi_Start clear it later on
5063 rxi_AckAllInTransmitQueue(struct rx_call *call)
5065 #ifdef RX_ENABLE_LOCKS
5066 if (call->flags & RX_CALL_TQ_BUSY) {
5067 rxi_SetAcksInTransmitQueue(call);
5071 rxi_ClearTransmitQueue(call, 0);
5073 /* Clear out the transmit queue for the current call (all packets have
5074 * been received by peer) */
5076 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5078 #ifdef RX_ENABLE_LOCKS
5079 struct opr_queue *cursor;
5080 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5082 for (opr_queue_Scan(&call->tq, cursor)) {
5084 = opr_queue_Entry(cursor, struct rx_packet, entry);
5086 p->flags |= RX_PKTFLAG_ACKED;
5090 call->flags |= RX_CALL_TQ_CLEARME;
5091 call->flags |= RX_CALL_TQ_SOME_ACKED;
5094 #endif /* RX_ENABLE_LOCKS */
5095 #ifdef RXDEBUG_PACKET
5097 #endif /* RXDEBUG_PACKET */
5098 rxi_FreePackets(0, &call->tq);
5099 rxi_WakeUpTransmitQueue(call);
5100 #ifdef RX_ENABLE_LOCKS
5101 call->flags &= ~RX_CALL_TQ_CLEARME;
5105 rxi_rto_cancel(call);
5106 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5107 call->nSoftAcked = 0;
5109 if (call->flags & RX_CALL_FAST_RECOVER) {
5110 call->flags &= ~RX_CALL_FAST_RECOVER;
5111 call->cwind = call->nextCwind;
5113 #ifdef RX_ENABLE_LOCKS
5114 CV_SIGNAL(&call->cv_twind);
5116 osi_rxWakeup(&call->twind);
5121 rxi_ClearReceiveQueue(struct rx_call *call)
5123 if (!opr_queue_IsEmpty(&call->rq)) {
5126 count = rxi_FreePackets(0, &call->rq);
5127 rx_packetReclaims += count;
5128 #ifdef RXDEBUG_PACKET
5130 if ( call->rqc != 0 )
5131 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5133 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5135 if (call->state == RX_STATE_PRECALL) {
5136 call->flags |= RX_CALL_CLEARED;
5140 /* Send an abort packet for the specified call */
5141 static struct rx_packet *
5142 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5143 int istack, int force)
5145 afs_int32 error, cerror;
5146 struct clock when, now;
5151 switch (call->error) {
5154 cerror = RX_CALL_TIMEOUT;
5157 cerror = call->error;
5160 /* Clients should never delay abort messages */
5161 if (rx_IsClientConn(call->conn))
5164 if (call->abortCode != cerror) {
5165 call->abortCode = cerror;
5166 call->abortCount = 0;
5169 if (force || rxi_callAbortThreshhold == 0
5170 || call->abortCount < rxi_callAbortThreshhold) {
5171 rxi_CancelDelayedAbortEvent(call);
5172 error = htonl(cerror);
5175 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5176 (char *)&error, sizeof(error), istack);
5177 } else if (!call->delayedAbortEvent) {
5178 clock_GetTime(&now);
5180 clock_Addmsec(&when, rxi_callAbortDelay);
5181 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5182 call->delayedAbortEvent =
5183 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5189 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5191 if (call->delayedAbortEvent) {
5192 rxevent_Cancel(&call->delayedAbortEvent);
5193 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5197 /* Send an abort packet for the specified connection. Packet is an
5198 * optional pointer to a packet that can be used to send the abort.
5199 * Once the number of abort messages reaches the threshhold, an
5200 * event is scheduled to send the abort. Setting the force flag
5201 * overrides sending delayed abort messages.
5203 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5204 * to send the abort packet.
5207 rxi_SendConnectionAbort(struct rx_connection *conn,
5208 struct rx_packet *packet, int istack, int force)
5215 /* Clients should never delay abort messages */
5216 if (rx_IsClientConn(conn))
5219 if (force || rxi_connAbortThreshhold == 0
5220 || conn->abortCount < rxi_connAbortThreshhold) {
5222 rxevent_Cancel(&conn->delayedAbortEvent);
5223 error = htonl(conn->error);
5225 MUTEX_EXIT(&conn->conn_data_lock);
5227 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5228 RX_PACKET_TYPE_ABORT, (char *)&error,
5229 sizeof(error), istack);
5230 MUTEX_ENTER(&conn->conn_data_lock);
5232 rxi_SendConnectionAbortLater(conn, rxi_connAbortDelay);
5237 /* Associate an error all of the calls owned by a connection. Called
5238 * with error non-zero. This is only for really fatal things, like
5239 * bad authentication responses. The connection itself is set in
5240 * error at this point, so that future packets received will be
5243 rxi_ConnectionError(struct rx_connection *conn,
5249 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5251 MUTEX_ENTER(&conn->conn_data_lock);
5252 rxevent_Cancel(&conn->challengeEvent);
5253 rxevent_Cancel(&conn->natKeepAliveEvent);
5254 if (conn->checkReachEvent) {
5255 rxevent_Cancel(&conn->checkReachEvent);
5256 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5257 putConnection(conn);
5259 MUTEX_EXIT(&conn->conn_data_lock);
5260 for (i = 0; i < RX_MAXCALLS; i++) {
5261 struct rx_call *call = conn->call[i];
5263 MUTEX_ENTER(&call->lock);
5264 rxi_CallError(call, error);
5265 MUTEX_EXIT(&call->lock);
5268 conn->error = error;
5269 if (rx_stats_active)
5270 rx_atomic_inc(&rx_stats.fatalErrors);
5275 * Interrupt an in-progress call with the specified error and wakeup waiters.
5277 * @param[in] call The call to interrupt
5278 * @param[in] error The error code to send to the peer
5281 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5283 MUTEX_ENTER(&call->lock);
5284 rxi_CallError(call, error);
5285 rxi_SendCallAbort(call, NULL, 0, 1);
5286 MUTEX_EXIT(&call->lock);
5290 rxi_CallError(struct rx_call *call, afs_int32 error)
5292 MUTEX_ASSERT(&call->lock);
5293 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5295 error = call->error;
5297 #ifdef RX_ENABLE_LOCKS
5298 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5299 rxi_ResetCall(call, 0);
5302 rxi_ResetCall(call, 0);
5304 call->error = error;
5307 /* Reset various fields in a call structure, and wakeup waiting
5308 * processes. Some fields aren't changed: state & mode are not
5309 * touched (these must be set by the caller), and bufptr, nLeft, and
5310 * nFree are not reset, since these fields are manipulated by
5311 * unprotected macros, and may only be reset by non-interrupting code.
5315 rxi_ResetCall(struct rx_call *call, int newcall)
5318 struct rx_peer *peer;
5319 struct rx_packet *packet;
5321 MUTEX_ASSERT(&call->lock);
5322 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5324 /* Notify anyone who is waiting for asynchronous packet arrival */
5325 if (call->arrivalProc) {
5326 (*call->arrivalProc) (call, call->arrivalProcHandle,
5327 call->arrivalProcArg);
5328 call->arrivalProc = (void (*)())0;
5332 rxi_CancelGrowMTUEvent(call);
5334 if (call->delayedAbortEvent) {
5335 rxi_CancelDelayedAbortEvent(call);
5336 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5338 rxi_SendCallAbort(call, packet, 0, 1);
5339 rxi_FreePacket(packet);
5344 * Update the peer with the congestion information in this call
5345 * so other calls on this connection can pick up where this call
5346 * left off. If the congestion sequence numbers don't match then
5347 * another call experienced a retransmission.
5349 peer = call->conn->peer;
5350 MUTEX_ENTER(&peer->peer_lock);
5352 if (call->congestSeq == peer->congestSeq) {
5353 peer->cwind = MAX(peer->cwind, call->cwind);
5354 peer->MTU = MAX(peer->MTU, call->MTU);
5355 peer->nDgramPackets =
5356 MAX(peer->nDgramPackets, call->nDgramPackets);
5359 call->abortCode = 0;
5360 call->abortCount = 0;
5362 if (peer->maxDgramPackets > 1) {
5363 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5365 call->MTU = peer->MTU;
5367 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5368 call->ssthresh = rx_maxSendWindow;
5369 call->nDgramPackets = peer->nDgramPackets;
5370 call->congestSeq = peer->congestSeq;
5371 call->rtt = peer->rtt;
5372 call->rtt_dev = peer->rtt_dev;
5373 clock_Zero(&call->rto);
5374 clock_Addmsec(&call->rto,
5375 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5376 MUTEX_EXIT(&peer->peer_lock);
5378 flags = call->flags;
5379 rxi_WaitforTQBusy(call);
5381 rxi_ClearTransmitQueue(call, 1);
5382 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5383 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5387 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5388 /* The call channel is still busy; resetting the call doesn't change
5389 * that. However, if 'newcall' is set, we are processing a call
5390 * structure that has either been recycled from the free list, or has
5391 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5392 * 'newcall' is set, since it describes a completely different call
5393 * channel which we do not care about. */
5394 call->flags |= RX_CALL_PEER_BUSY;
5397 rxi_ClearReceiveQueue(call);
5398 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5402 call->twind = call->conn->twind[call->channel];
5403 call->rwind = call->conn->rwind[call->channel];
5404 call->nSoftAcked = 0;
5405 call->nextCwind = 0;
5408 call->nCwindAcks = 0;
5409 call->nSoftAcks = 0;
5410 call->nHardAcks = 0;
5412 call->tfirst = call->rnext = call->tnext = 1;
5415 call->lastAcked = 0;
5416 call->localStatus = call->remoteStatus = 0;
5417 call->lastSendData = 0;
5419 if (flags & RX_CALL_READER_WAIT) {
5420 #ifdef RX_ENABLE_LOCKS
5421 CV_BROADCAST(&call->cv_rq);
5423 osi_rxWakeup(&call->rq);
5426 if (flags & RX_CALL_WAIT_PACKETS) {
5427 MUTEX_ENTER(&rx_freePktQ_lock);
5428 rxi_PacketsUnWait(); /* XXX */
5429 MUTEX_EXIT(&rx_freePktQ_lock);
5431 #ifdef RX_ENABLE_LOCKS
5432 CV_SIGNAL(&call->cv_twind);
5434 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5435 osi_rxWakeup(&call->twind);
5438 if (flags & RX_CALL_WAIT_PROC) {
5439 rx_atomic_dec(&rx_nWaiting);
5441 #ifdef RX_ENABLE_LOCKS
5442 /* The following ensures that we don't mess with any queue while some
5443 * other thread might also be doing so. The call_queue_lock field is
5444 * is only modified under the call lock. If the call is in the process
5445 * of being removed from a queue, the call is not locked until the
5446 * the queue lock is dropped and only then is the call_queue_lock field
5447 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5448 * Note that any other routine which removes a call from a queue has to
5449 * obtain the queue lock before examing the queue and removing the call.
5451 if (call->call_queue_lock) {
5452 MUTEX_ENTER(call->call_queue_lock);
5453 if (opr_queue_IsOnQueue(&call->entry)) {
5454 opr_queue_Remove(&call->entry);
5456 MUTEX_EXIT(call->call_queue_lock);
5457 CLEAR_CALL_QUEUE_LOCK(call);
5459 #else /* RX_ENABLE_LOCKS */
5460 if (opr_queue_IsOnQueue(&call->entry)) {
5461 opr_queue_Remove(&call->entry);
5463 #endif /* RX_ENABLE_LOCKS */
5465 rxi_CancelKeepAliveEvent(call);
5466 rxi_CancelDelayedAckEvent(call);
5469 /* Send an acknowledge for the indicated packet (seq,serial) of the
5470 * indicated call, for the indicated reason (reason). This
5471 * acknowledge will specifically acknowledge receiving the packet, and
5472 * will also specify which other packets for this call have been
5473 * received. This routine returns the packet that was used to the
5474 * caller. The caller is responsible for freeing it or re-using it.
5475 * This acknowledgement also returns the highest sequence number
5476 * actually read out by the higher level to the sender; the sender
5477 * promises to keep around packets that have not been read by the
5478 * higher level yet (unless, of course, the sender decides to abort
5479 * the call altogether). Any of p, seq, serial, pflags, or reason may
5480 * be set to zero without ill effect. That is, if they are zero, they
5481 * will not convey any information.
5482 * NOW there is a trailer field, after the ack where it will safely be
5483 * ignored by mundanes, which indicates the maximum size packet this
5484 * host can swallow. */
5486 struct rx_packet *optionalPacket; use to send ack (or null)
5487 int seq; Sequence number of the packet we are acking
5488 int serial; Serial number of the packet
5489 int pflags; Flags field from packet header
5490 int reason; Reason an acknowledge was prompted
5494 rxi_SendAck(struct rx_call *call,
5495 struct rx_packet *optionalPacket, int serial, int reason,
5498 struct rx_ackPacket *ap;
5499 struct rx_packet *p;
5500 struct opr_queue *cursor;
5503 afs_uint32 padbytes = 0;
5504 #ifdef RX_ENABLE_TSFPQ
5505 struct rx_ts_info_t * rx_ts_info;
5509 * Open the receive window once a thread starts reading packets
5511 if (call->rnext > 1) {
5512 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5515 /* Don't attempt to grow MTU if this is a critical ping */
5516 if (reason == RX_ACK_MTU) {
5517 /* keep track of per-call attempts, if we're over max, do in small
5518 * otherwise in larger? set a size to increment by, decrease
5521 if (call->conn->peer->maxPacketSize &&
5522 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5524 padbytes = call->conn->peer->maxPacketSize+16;
5526 padbytes = call->conn->peer->maxMTU + 128;
5528 /* do always try a minimum size ping */
5529 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5531 /* subtract the ack payload */
5532 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5533 reason = RX_ACK_PING;
5536 call->nHardAcks = 0;
5537 call->nSoftAcks = 0;
5538 if (call->rnext > call->lastAcked)
5539 call->lastAcked = call->rnext;
5543 rx_computelen(p, p->length); /* reset length, you never know */
5544 } /* where that's been... */
5545 #ifdef RX_ENABLE_TSFPQ
5547 RX_TS_INFO_GET(rx_ts_info);
5548 if ((p = rx_ts_info->local_special_packet)) {
5549 rx_computelen(p, p->length);
5550 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5551 rx_ts_info->local_special_packet = p;
5552 } else { /* We won't send the ack, but don't panic. */
5553 return optionalPacket;
5557 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5558 /* We won't send the ack, but don't panic. */
5559 return optionalPacket;
5564 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5567 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5568 #ifndef RX_ENABLE_TSFPQ
5569 if (!optionalPacket)
5572 return optionalPacket;
5574 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5575 if (rx_Contiguous(p) < templ) {
5576 #ifndef RX_ENABLE_TSFPQ
5577 if (!optionalPacket)
5580 return optionalPacket;
5585 /* MTUXXX failing to send an ack is very serious. We should */
5586 /* try as hard as possible to send even a partial ack; it's */
5587 /* better than nothing. */
5588 ap = (struct rx_ackPacket *)rx_DataOf(p);
5589 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5590 ap->reason = reason;
5592 /* The skew computation used to be bogus, I think it's better now. */
5593 /* We should start paying attention to skew. XXX */
5594 ap->serial = htonl(serial);
5595 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5598 * First packet not yet forwarded to reader. When ACKALL has been
5599 * sent the peer has been told that all received packets will be
5600 * delivered to the reader. The value 'rnext' is used internally
5601 * to refer to the next packet in the receive queue that must be
5602 * delivered to the reader. From the perspective of the peer it
5603 * already has so report the last sequence number plus one if there
5604 * are packets in the receive queue awaiting processing.
5606 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5607 !opr_queue_IsEmpty(&call->rq)) {
5608 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5610 ap->firstPacket = htonl(call->rnext);
5612 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5614 /* No fear of running out of ack packet here because there can only
5615 * be at most one window full of unacknowledged packets. The window
5616 * size must be constrained to be less than the maximum ack size,
5617 * of course. Also, an ack should always fit into a single packet
5618 * -- it should not ever be fragmented. */
5620 for (opr_queue_Scan(&call->rq, cursor)) {
5621 struct rx_packet *rqp
5622 = opr_queue_Entry(cursor, struct rx_packet, entry);
5624 if (!rqp || !call->rq.next
5625 || (rqp->header.seq > (call->rnext + call->rwind))) {
5626 #ifndef RX_ENABLE_TSFPQ
5627 if (!optionalPacket)
5630 rxi_CallError(call, RX_CALL_DEAD);
5631 return optionalPacket;
5634 while (rqp->header.seq > call->rnext + offset)
5635 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5636 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5638 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5639 #ifndef RX_ENABLE_TSFPQ
5640 if (!optionalPacket)
5643 rxi_CallError(call, RX_CALL_DEAD);
5644 return optionalPacket;
5650 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5652 /* these are new for AFS 3.3 */
5653 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5654 templ = htonl(templ);
5655 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5656 templ = htonl(call->conn->peer->ifMTU);
5657 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5658 sizeof(afs_int32), &templ);
5660 /* new for AFS 3.4 */
5661 templ = htonl(call->rwind);
5662 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5663 sizeof(afs_int32), &templ);
5665 /* new for AFS 3.5 */
5666 templ = htonl(call->conn->peer->ifDgramPackets);
5667 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5668 sizeof(afs_int32), &templ);
5670 p->header.serviceId = call->conn->serviceId;
5671 p->header.cid = (call->conn->cid | call->channel);
5672 p->header.callNumber = *call->callNumber;
5674 p->header.securityIndex = call->conn->securityIndex;
5675 p->header.epoch = call->conn->epoch;
5676 p->header.type = RX_PACKET_TYPE_ACK;
5677 p->header.flags = RX_SLOW_START_OK;
5678 if (reason == RX_ACK_PING) {
5679 p->header.flags |= RX_REQUEST_ACK;
5681 p->length = padbytes +
5682 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5685 /* not fast but we can potentially use this if truncated
5686 * fragments are delivered to figure out the mtu.
5688 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5689 sizeof(afs_int32), sizeof(afs_int32),
5693 if (call->conn->type == RX_CLIENT_CONNECTION)
5694 p->header.flags |= RX_CLIENT_INITIATED;
5698 if (rxdebug_active) {
5702 len = _snprintf(msg, sizeof(msg),
5703 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5704 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5705 ntohl(ap->serial), ntohl(ap->previousPacket),
5706 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5707 ap->nAcks, ntohs(ap->bufferSpace) );
5711 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5712 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5716 OutputDebugString(msg);
5718 #else /* AFS_NT40_ENV */
5720 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5721 ap->reason, ntohl(ap->previousPacket),
5722 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5724 for (offset = 0; offset < ap->nAcks; offset++)
5725 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5730 #endif /* AFS_NT40_ENV */
5733 int i, nbytes = p->length;
5735 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5736 if (nbytes <= p->wirevec[i].iov_len) {
5739 savelen = p->wirevec[i].iov_len;
5741 p->wirevec[i].iov_len = nbytes;
5743 rxi_Send(call, p, istack);
5744 p->wirevec[i].iov_len = savelen;
5748 nbytes -= p->wirevec[i].iov_len;
5751 if (rx_stats_active)
5752 rx_atomic_inc(&rx_stats.ackPacketsSent);
5753 #ifndef RX_ENABLE_TSFPQ
5754 if (!optionalPacket)
5757 return optionalPacket; /* Return packet for re-use by caller */
5761 struct rx_packet **list;
5766 /* Send all of the packets in the list in single datagram */
5768 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5769 int istack, int moreFlag)
5775 struct rx_connection *conn = call->conn;
5776 struct rx_peer *peer = conn->peer;
5778 MUTEX_ENTER(&peer->peer_lock);
5779 peer->nSent += xmit->len;
5780 if (xmit->resending)
5781 peer->reSends += xmit->len;
5782 MUTEX_EXIT(&peer->peer_lock);
5784 if (rx_stats_active) {
5785 if (xmit->resending)
5786 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5788 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5791 clock_GetTime(&now);
5793 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5797 /* Set the packet flags and schedule the resend events */
5798 /* Only request an ack for the last packet in the list */
5799 for (i = 0; i < xmit->len; i++) {
5800 struct rx_packet *packet = xmit->list[i];
5802 /* Record the time sent */
5803 packet->timeSent = now;
5804 packet->flags |= RX_PKTFLAG_SENT;
5806 /* Ask for an ack on retransmitted packets, on every other packet
5807 * if the peer doesn't support slow start. Ask for an ack on every
5808 * packet until the congestion window reaches the ack rate. */
5809 if (packet->header.serial) {
5812 packet->firstSent = now;
5813 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5814 || (!(call->flags & RX_CALL_SLOW_START_OK)
5815 && (packet->header.seq & 1)))) {
5820 /* Tag this packet as not being the last in this group,
5821 * for the receiver's benefit */
5822 if (i < xmit->len - 1 || moreFlag) {
5823 packet->header.flags |= RX_MORE_PACKETS;
5828 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5831 /* Since we're about to send a data packet to the peer, it's
5832 * safe to nuke any scheduled end-of-packets ack */
5833 rxi_CancelDelayedAckEvent(call);
5835 MUTEX_EXIT(&call->lock);
5836 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5837 if (xmit->len > 1) {
5838 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5840 rxi_SendPacket(call, conn, xmit->list[0], istack);
5842 MUTEX_ENTER(&call->lock);
5843 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5845 /* Tell the RTO calculation engine that we have sent a packet, and
5846 * if it was the last one */
5847 rxi_rto_packet_sent(call, lastPacket, istack);
5849 /* Update last send time for this call (for keep-alive
5850 * processing), and for the connection (so that we can discover
5851 * idle connections) */
5852 conn->lastSendTime = call->lastSendTime = clock_Sec();
5853 /* Let a set of retransmits trigger an idle timeout */
5854 if (!xmit->resending)
5855 call->lastSendData = call->lastSendTime;
5858 /* When sending packets we need to follow these rules:
5859 * 1. Never send more than maxDgramPackets in a jumbogram.
5860 * 2. Never send a packet with more than two iovecs in a jumbogram.
5861 * 3. Never send a retransmitted packet in a jumbogram.
5862 * 4. Never send more than cwind/4 packets in a jumbogram
5863 * We always keep the last list we should have sent so we
5864 * can set the RX_MORE_PACKETS flags correctly.
5868 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5873 struct xmitlist working;
5874 struct xmitlist last;
5876 struct rx_peer *peer = call->conn->peer;
5877 int morePackets = 0;
5879 memset(&last, 0, sizeof(struct xmitlist));
5880 working.list = &list[0];
5882 working.resending = 0;
5884 recovery = call->flags & RX_CALL_FAST_RECOVER;
5886 for (i = 0; i < len; i++) {
5887 /* Does the current packet force us to flush the current list? */
5889 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5890 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5892 /* This sends the 'last' list and then rolls the current working
5893 * set into the 'last' one, and resets the working set */
5896 rxi_SendList(call, &last, istack, 1);
5897 /* If the call enters an error state stop sending, or if
5898 * we entered congestion recovery mode, stop sending */
5900 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5905 working.resending = 0;
5906 working.list = &list[i];
5908 /* Add the current packet to the list if it hasn't been acked.
5909 * Otherwise adjust the list pointer to skip the current packet. */
5910 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5913 if (list[i]->header.serial)
5914 working.resending = 1;
5916 /* Do we need to flush the list? */
5917 if (working.len >= (int)peer->maxDgramPackets
5918 || working.len >= (int)call->nDgramPackets
5919 || working.len >= (int)call->cwind
5920 || list[i]->header.serial
5921 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5923 rxi_SendList(call, &last, istack, 1);
5924 /* If the call enters an error state stop sending, or if
5925 * we entered congestion recovery mode, stop sending */
5927 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5932 working.resending = 0;
5933 working.list = &list[i + 1];
5936 if (working.len != 0) {
5937 osi_Panic("rxi_SendList error");
5939 working.list = &list[i + 1];
5943 /* Send the whole list when the call is in receive mode, when
5944 * the call is in eof mode, when we are in fast recovery mode,
5945 * and when we have the last packet */
5946 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5947 * the listener or event threads
5949 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5950 || (call->flags & RX_CALL_FLUSH)
5951 || (call->flags & RX_CALL_FAST_RECOVER)) {
5952 /* Check for the case where the current list contains
5953 * an acked packet. Since we always send retransmissions
5954 * in a separate packet, we only need to check the first
5955 * packet in the list */
5956 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5960 rxi_SendList(call, &last, istack, morePackets);
5961 /* If the call enters an error state stop sending, or if
5962 * we entered congestion recovery mode, stop sending */
5964 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5968 rxi_SendList(call, &working, istack, 0);
5970 } else if (last.len > 0) {
5971 rxi_SendList(call, &last, istack, 0);
5972 /* Packets which are in 'working' are not sent by this call */
5977 * Check if the peer for the given call is known to be dead
5979 * If the call's peer appears dead (it has encountered fatal network errors
5980 * since the call started) the call is killed with RX_CALL_DEAD if the call
5981 * is active. Otherwise, we do nothing.
5983 * @param[in] call The call to check
5986 * @retval 0 The call is fine, and we haven't done anything to the call
5987 * @retval nonzero The call's peer appears dead, and the call has been
5988 * terminated if it was active
5990 * @pre call->lock must be locked
5993 rxi_CheckPeerDead(struct rx_call *call)
5995 #ifdef AFS_RXERRQ_ENV
5998 if (call->state == RX_STATE_DALLY) {
6002 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
6003 if (call->neterr_gen < peererrs) {
6004 /* we have received network errors since this call started; kill
6006 if (call->state == RX_STATE_ACTIVE) {
6007 rxi_CallError(call, RX_CALL_DEAD);
6011 if (call->neterr_gen > peererrs) {
6012 /* someone has reset the number of peer errors; set the call error gen
6013 * so we can detect if more errors are encountered */
6014 call->neterr_gen = peererrs;
6021 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
6023 struct rx_call *call = arg0;
6024 struct rx_peer *peer;
6025 struct opr_queue *cursor;
6026 struct clock maxTimeout = { 60, 0 };
6028 MUTEX_ENTER(&call->lock);
6030 peer = call->conn->peer;
6032 /* Make sure that the event pointer is removed from the call
6033 * structure, since there is no longer a per-call retransmission
6035 if (event == call->resendEvent) {
6036 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6037 rxevent_Put(&call->resendEvent);
6040 rxi_CheckPeerDead(call);
6042 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6043 rxi_CheckBusy(call);
6046 if (opr_queue_IsEmpty(&call->tq)) {
6047 /* Nothing to do. This means that we've been raced, and that an
6048 * ACK has come in between when we were triggered, and when we
6049 * actually got to run. */
6053 /* We're in loss recovery */
6054 call->flags |= RX_CALL_FAST_RECOVER;
6056 /* Mark all of the pending packets in the queue as being lost */
6057 for (opr_queue_Scan(&call->tq, cursor)) {
6058 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6059 if (!(p->flags & RX_PKTFLAG_ACKED))
6060 p->flags &= ~RX_PKTFLAG_SENT;
6063 /* We're resending, so we double the timeout of the call. This will be
6064 * dropped back down by the first successful ACK that we receive.
6066 * We apply a maximum value here of 60 seconds
6068 clock_Add(&call->rto, &call->rto);
6069 if (clock_Gt(&call->rto, &maxTimeout))
6070 call->rto = maxTimeout;
6072 /* Packet loss is most likely due to congestion, so drop our window size
6073 * and start again from the beginning */
6074 if (peer->maxDgramPackets >1) {
6075 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6076 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6078 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6079 call->nDgramPackets = 1;
6081 call->nextCwind = 1;
6084 MUTEX_ENTER(&peer->peer_lock);
6085 peer->MTU = call->MTU;
6086 peer->cwind = call->cwind;
6087 peer->nDgramPackets = 1;
6089 call->congestSeq = peer->congestSeq;
6090 MUTEX_EXIT(&peer->peer_lock);
6092 rxi_Start(call, istack);
6095 MUTEX_EXIT(&call->lock);
6098 /* This routine is called when new packets are readied for
6099 * transmission and when retransmission may be necessary, or when the
6100 * transmission window or burst count are favourable. This should be
6101 * better optimized for new packets, the usual case, now that we've
6102 * got rid of queues of send packets. XXXXXXXXXXX */
6104 rxi_Start(struct rx_call *call, int istack)
6106 struct opr_queue *cursor;
6107 #ifdef RX_ENABLE_LOCKS
6108 struct opr_queue *store;
6114 #ifdef RX_ENABLE_LOCKS
6115 if (rx_stats_active)
6116 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6121 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6122 /* Send (or resend) any packets that need it, subject to
6123 * window restrictions and congestion burst control
6124 * restrictions. Ask for an ack on the last packet sent in
6125 * this burst. For now, we're relying upon the window being
6126 * considerably bigger than the largest number of packets that
6127 * are typically sent at once by one initial call to
6128 * rxi_Start. This is probably bogus (perhaps we should ask
6129 * for an ack when we're half way through the current
6130 * window?). Also, for non file transfer applications, this
6131 * may end up asking for an ack for every packet. Bogus. XXXX
6134 * But check whether we're here recursively, and let the other guy
6137 #ifdef RX_ENABLE_LOCKS
6138 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6139 call->flags |= RX_CALL_TQ_BUSY;
6141 #endif /* RX_ENABLE_LOCKS */
6143 #ifdef RX_ENABLE_LOCKS
6144 call->flags &= ~RX_CALL_NEED_START;
6145 #endif /* RX_ENABLE_LOCKS */
6147 maxXmitPackets = MIN(call->twind, call->cwind);
6148 for (opr_queue_Scan(&call->tq, cursor)) {
6150 = opr_queue_Entry(cursor, struct rx_packet, entry);
6152 if (p->flags & RX_PKTFLAG_ACKED) {
6153 /* Since we may block, don't trust this */
6154 if (rx_stats_active)
6155 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6156 continue; /* Ignore this packet if it has been acknowledged */
6159 /* Turn off all flags except these ones, which are the same
6160 * on each transmission */
6161 p->header.flags &= RX_PRESET_FLAGS;
6163 if (p->header.seq >=
6164 call->tfirst + MIN((int)call->twind,
6165 (int)(call->nSoftAcked +
6167 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6168 /* Note: if we're waiting for more window space, we can
6169 * still send retransmits; hence we don't return here, but
6170 * break out to schedule a retransmit event */
6171 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6172 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6177 /* Transmit the packet if it needs to be sent. */
6178 if (!(p->flags & RX_PKTFLAG_SENT)) {
6179 if (nXmitPackets == maxXmitPackets) {
6180 rxi_SendXmitList(call, call->xmitList,
6181 nXmitPackets, istack);
6184 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6185 *(call->callNumber), p));
6186 call->xmitList[nXmitPackets++] = p;
6188 } /* end of the queue_Scan */
6190 /* xmitList now hold pointers to all of the packets that are
6191 * ready to send. Now we loop to send the packets */
6192 if (nXmitPackets > 0) {
6193 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6197 #ifdef RX_ENABLE_LOCKS
6199 /* We went into the error state while sending packets. Now is
6200 * the time to reset the call. This will also inform the using
6201 * process that the call is in an error state.
6203 if (rx_stats_active)
6204 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6205 call->flags &= ~RX_CALL_TQ_BUSY;
6206 rxi_WakeUpTransmitQueue(call);
6207 rxi_CallError(call, call->error);
6211 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6213 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6214 /* Some packets have received acks. If they all have, we can clear
6215 * the transmit queue.
6218 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6220 = opr_queue_Entry(cursor, struct rx_packet, entry);
6222 if (p->header.seq < call->tfirst
6223 && (p->flags & RX_PKTFLAG_ACKED)) {
6224 opr_queue_Remove(&p->entry);
6225 #ifdef RX_TRACK_PACKETS
6226 p->flags &= ~RX_PKTFLAG_TQ;
6228 #ifdef RXDEBUG_PACKET
6236 call->flags |= RX_CALL_TQ_CLEARME;
6238 if (call->flags & RX_CALL_TQ_CLEARME)
6239 rxi_ClearTransmitQueue(call, 1);
6240 } while (call->flags & RX_CALL_NEED_START);
6242 * TQ references no longer protected by this flag; they must remain
6243 * protected by the call lock.
6245 call->flags &= ~RX_CALL_TQ_BUSY;
6246 rxi_WakeUpTransmitQueue(call);
6248 call->flags |= RX_CALL_NEED_START;
6250 #endif /* RX_ENABLE_LOCKS */
6252 rxi_rto_cancel(call);
6256 /* Also adjusts the keep alive parameters for the call, to reflect
6257 * that we have just sent a packet (so keep alives aren't sent
6260 rxi_Send(struct rx_call *call, struct rx_packet *p,
6263 struct rx_connection *conn = call->conn;
6265 /* Stamp each packet with the user supplied status */
6266 p->header.userStatus = call->localStatus;
6268 /* Allow the security object controlling this call's security to
6269 * make any last-minute changes to the packet */
6270 RXS_SendPacket(conn->securityObject, call, p);
6272 /* Since we're about to send SOME sort of packet to the peer, it's
6273 * safe to nuke any scheduled end-of-packets ack */
6274 rxi_CancelDelayedAckEvent(call);
6276 /* Actually send the packet, filling in more connection-specific fields */
6277 MUTEX_EXIT(&call->lock);
6278 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6279 rxi_SendPacket(call, conn, p, istack);
6280 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6281 MUTEX_ENTER(&call->lock);
6283 /* Update last send time for this call (for keep-alive
6284 * processing), and for the connection (so that we can discover
6285 * idle connections) */
6286 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6287 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6288 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6290 conn->lastSendTime = call->lastSendTime = clock_Sec();
6291 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6292 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6293 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6294 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6295 RX_ACK_PING_RESPONSE)))
6296 call->lastSendData = call->lastSendTime;
6300 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6301 * that things are fine. Also called periodically to guarantee that nothing
6302 * falls through the cracks (e.g. (error + dally) connections have keepalive
6303 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6305 * haveCTLock Set if calling from rxi_ReapConnections
6308 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6310 struct rx_connection *conn = call->conn;
6312 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6313 afs_uint32 fudgeFactor;
6316 int idle_timeout = 0;
6317 afs_int32 clock_diff = 0;
6319 if (rxi_CheckPeerDead(call)) {
6325 /* Large swings in the clock can have a significant impact on
6326 * the performance of RX call processing. Forward clock shifts
6327 * will result in premature event triggering or timeouts.
6328 * Backward shifts can result in calls not completing until
6329 * the clock catches up with the original start clock value.
6331 * If a backward clock shift of more than five minutes is noticed,
6332 * just fail the call.
6334 if (now < call->lastSendTime)
6335 clock_diff = call->lastSendTime - now;
6336 if (now < call->startWait)
6337 clock_diff = MAX(clock_diff, call->startWait - now);
6338 if (now < call->lastReceiveTime)
6339 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6340 if (clock_diff > 5 * 60)
6342 if (call->state == RX_STATE_ACTIVE)
6343 rxi_CallError(call, RX_CALL_TIMEOUT);
6347 #ifdef RX_ENABLE_LOCKS
6348 if (call->flags & RX_CALL_TQ_BUSY) {
6349 /* Call is active and will be reset by rxi_Start if it's
6350 * in an error state.
6355 /* RTT + 8*MDEV, rounded up to the next second. */
6356 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6357 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6359 deadTime = conn->secondsUntilDead + fudgeFactor;
6360 /* These are computed to the second (+- 1 second). But that's
6361 * good enough for these values, which should be a significant
6362 * number of seconds. */
6363 if (now > (call->lastReceiveTime + deadTime)) {
6364 if (call->state == RX_STATE_ACTIVE) {
6365 #ifdef AFS_ADAPT_PMTU
6366 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6368 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6369 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6370 ip_stack_t *ipst = ns->netstack_ip;
6372 ire = ire_cache_lookup(conn->peer->host
6373 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6375 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6377 # if defined(GLOBAL_NETSTACKID)
6384 if (ire && ire->ire_max_frag > 0)
6385 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6387 # if defined(GLOBAL_NETSTACKID)
6391 #endif /* AFS_ADAPT_PMTU */
6392 cerror = RX_CALL_DEAD;
6395 #ifdef RX_ENABLE_LOCKS
6396 /* Cancel pending events */
6397 rxi_CancelDelayedAckEvent(call);
6398 rxi_rto_cancel(call);
6399 rxi_CancelKeepAliveEvent(call);
6400 rxi_CancelGrowMTUEvent(call);
6401 MUTEX_ENTER(&rx_refcnt_mutex);
6402 /* if rxi_FreeCall returns 1 it has freed the call */
6403 if (call->refCount == 0 &&
6404 rxi_FreeCall(call, haveCTLock))
6406 MUTEX_EXIT(&rx_refcnt_mutex);
6409 MUTEX_EXIT(&rx_refcnt_mutex);
6411 #else /* RX_ENABLE_LOCKS */
6412 rxi_FreeCall(call, 0);
6414 #endif /* RX_ENABLE_LOCKS */
6416 /* Non-active calls are destroyed if they are not responding
6417 * to pings; active calls are simply flagged in error, so the
6418 * attached process can die reasonably gracefully. */
6421 if (conn->idleDeadDetection) {
6422 if (conn->idleDeadTime) {
6423 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6427 /* see if we have a non-activity timeout */
6428 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6429 (call->flags & RX_CALL_READER_WAIT)) {
6430 if (call->state == RX_STATE_ACTIVE) {
6431 cerror = RX_CALL_TIMEOUT;
6436 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6437 if (call->state == RX_STATE_ACTIVE) {
6438 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6446 if (conn->hardDeadTime) {
6447 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6450 /* see if we have a hard timeout */
6452 && (now > (hardDeadTime + call->startTime.sec))) {
6453 if (call->state == RX_STATE_ACTIVE)
6454 rxi_CallError(call, RX_CALL_TIMEOUT);
6459 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6460 call->lastReceiveTime) {
6461 int oldMTU = conn->peer->ifMTU;
6463 /* if we thought we could send more, perhaps things got worse */
6464 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6465 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6466 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6467 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6469 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6471 /* minimum capped in SetPeerMtu */
6472 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6475 conn->lastPacketSize = 0;
6477 /* needed so ResetCall doesn't clobber us. */
6478 call->MTU = conn->peer->ifMTU;
6480 /* if we never succeeded, let the error pass out as-is */
6481 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6482 cerror = conn->msgsizeRetryErr;
6485 rxi_CallError(call, cerror);
6490 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6491 void *dummy, int dummy2)
6493 struct rx_connection *conn = arg1;
6494 struct rx_header theader;
6495 char tbuffer[1 + sizeof(struct rx_header)];
6496 struct sockaddr_in taddr;
6499 struct iovec tmpiov[2];
6502 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6505 tp = &tbuffer[sizeof(struct rx_header)];
6506 taddr.sin_family = AF_INET;
6507 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6508 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6509 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6510 taddr.sin_len = sizeof(struct sockaddr_in);
6512 memset(&theader, 0, sizeof(theader));
6513 theader.epoch = htonl(999);
6515 theader.callNumber = 0;
6518 theader.type = RX_PACKET_TYPE_VERSION;
6519 theader.flags = RX_LAST_PACKET;
6520 theader.serviceId = 0;
6522 memcpy(tbuffer, &theader, sizeof(theader));
6523 memcpy(tp, &a, sizeof(a));
6524 tmpiov[0].iov_base = tbuffer;
6525 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6527 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6529 MUTEX_ENTER(&conn->conn_data_lock);
6530 MUTEX_ENTER(&rx_refcnt_mutex);
6531 /* Only reschedule ourselves if the connection would not be destroyed */
6532 if (conn->refCount <= 1) {
6533 rxevent_Put(&conn->natKeepAliveEvent);
6534 MUTEX_EXIT(&rx_refcnt_mutex);
6535 MUTEX_EXIT(&conn->conn_data_lock);
6536 rx_DestroyConnection(conn); /* drop the reference for this */
6538 conn->refCount--; /* drop the reference for this */
6539 MUTEX_EXIT(&rx_refcnt_mutex);
6540 rxevent_Put(&conn->natKeepAliveEvent);
6541 rxi_ScheduleNatKeepAliveEvent(conn);
6542 MUTEX_EXIT(&conn->conn_data_lock);
6547 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6549 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6550 struct clock when, now;
6551 clock_GetTime(&now);
6553 when.sec += conn->secondsUntilNatPing;
6554 MUTEX_ENTER(&rx_refcnt_mutex);
6555 conn->refCount++; /* hold a reference for this */
6556 MUTEX_EXIT(&rx_refcnt_mutex);
6557 conn->natKeepAliveEvent =
6558 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6563 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6565 MUTEX_ENTER(&conn->conn_data_lock);
6566 conn->secondsUntilNatPing = seconds;
6568 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6569 rxi_ScheduleNatKeepAliveEvent(conn);
6571 conn->flags |= RX_CONN_NAT_PING;
6573 MUTEX_EXIT(&conn->conn_data_lock);
6576 /* When a call is in progress, this routine is called occasionally to
6577 * make sure that some traffic has arrived (or been sent to) the peer.
6578 * If nothing has arrived in a reasonable amount of time, the call is
6579 * declared dead; if nothing has been sent for a while, we send a
6580 * keep-alive packet (if we're actually trying to keep the call alive)
6583 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6586 struct rx_call *call = arg1;
6587 struct rx_connection *conn;
6590 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6591 MUTEX_ENTER(&call->lock);
6593 if (event == call->keepAliveEvent)
6594 rxevent_Put(&call->keepAliveEvent);
6598 if (rxi_CheckCall(call, 0)) {
6599 MUTEX_EXIT(&call->lock);
6603 /* Don't try to keep alive dallying calls */
6604 if (call->state == RX_STATE_DALLY) {
6605 MUTEX_EXIT(&call->lock);
6610 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6611 /* Don't try to send keepalives if there is unacknowledged data */
6612 /* the rexmit code should be good enough, this little hack
6613 * doesn't quite work XXX */
6614 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6616 rxi_ScheduleKeepAliveEvent(call);
6617 MUTEX_EXIT(&call->lock);
6620 /* Does what's on the nameplate. */
6622 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6624 struct rx_call *call = arg1;
6625 struct rx_connection *conn;
6627 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6628 MUTEX_ENTER(&call->lock);
6630 if (event == call->growMTUEvent)
6631 rxevent_Put(&call->growMTUEvent);
6633 if (rxi_CheckCall(call, 0)) {
6634 MUTEX_EXIT(&call->lock);
6638 /* Don't bother with dallying calls */
6639 if (call->state == RX_STATE_DALLY) {
6640 MUTEX_EXIT(&call->lock);
6647 * keep being scheduled, just don't do anything if we're at peak,
6648 * or we're not set up to be properly handled (idle timeout required)
6650 if ((conn->peer->maxPacketSize != 0) &&
6651 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6652 conn->idleDeadDetection)
6653 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6654 rxi_ScheduleGrowMTUEvent(call, 0);
6655 MUTEX_EXIT(&call->lock);
6659 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6661 if (!call->keepAliveEvent) {
6662 struct clock when, now;
6663 clock_GetTime(&now);
6665 when.sec += call->conn->secondsUntilPing;
6666 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6667 call->keepAliveEvent =
6668 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6673 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6674 if (call->keepAliveEvent) {
6675 rxevent_Cancel(&call->keepAliveEvent);
6676 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6681 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6683 if (!call->growMTUEvent) {
6684 struct clock when, now;
6686 clock_GetTime(&now);
6689 if (call->conn->secondsUntilPing)
6690 secs = (6*call->conn->secondsUntilPing)-1;
6692 if (call->conn->secondsUntilDead)
6693 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6697 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6698 call->growMTUEvent =
6699 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6704 rxi_CancelGrowMTUEvent(struct rx_call *call)
6706 if (call->growMTUEvent) {
6707 rxevent_Cancel(&call->growMTUEvent);
6708 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6713 * Increment the counter for the next connection ID, handling overflow.
6716 update_nextCid(void)
6718 /* Overflow is technically undefined behavior; avoid it. */
6719 if (rx_nextCid > MAX_AFS_INT32 - (1 << RX_CIDSHIFT))
6720 rx_nextCid = -1 * ((MAX_AFS_INT32 / RX_CIDSHIFT) * RX_CIDSHIFT);
6722 rx_nextCid += 1 << RX_CIDSHIFT;
6726 rxi_KeepAliveOn(struct rx_call *call)
6728 /* Pretend last packet received was received now--i.e. if another
6729 * packet isn't received within the keep alive time, then the call
6730 * will die; Initialize last send time to the current time--even
6731 * if a packet hasn't been sent yet. This will guarantee that a
6732 * keep-alive is sent within the ping time */
6733 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6734 rxi_ScheduleKeepAliveEvent(call);
6738 rx_KeepAliveOff(struct rx_call *call)
6740 MUTEX_ENTER(&call->lock);
6741 rxi_CancelKeepAliveEvent(call);
6742 MUTEX_EXIT(&call->lock);
6746 rx_KeepAliveOn(struct rx_call *call)
6748 MUTEX_ENTER(&call->lock);
6749 rxi_KeepAliveOn(call);
6750 MUTEX_EXIT(&call->lock);
6754 rxi_GrowMTUOn(struct rx_call *call)
6756 struct rx_connection *conn = call->conn;
6757 MUTEX_ENTER(&conn->conn_data_lock);
6758 conn->lastPingSizeSer = conn->lastPingSize = 0;
6759 MUTEX_EXIT(&conn->conn_data_lock);
6760 rxi_ScheduleGrowMTUEvent(call, 1);
6763 /* This routine is called to send connection abort messages
6764 * that have been delayed to throttle looping clients. */
6766 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6769 struct rx_connection *conn = arg1;
6772 struct rx_packet *packet;
6774 MUTEX_ENTER(&conn->conn_data_lock);
6775 rxevent_Put(&conn->delayedAbortEvent);
6776 error = htonl(conn->error);
6778 MUTEX_EXIT(&conn->conn_data_lock);
6779 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6782 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6783 RX_PACKET_TYPE_ABORT, (char *)&error,
6785 rxi_FreePacket(packet);
6789 /* This routine is called to send call abort messages
6790 * that have been delayed to throttle looping clients. */
6792 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6795 struct rx_call *call = arg1;
6798 struct rx_packet *packet;
6800 MUTEX_ENTER(&call->lock);
6801 rxevent_Put(&call->delayedAbortEvent);
6802 error = htonl(call->error);
6804 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6807 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6808 (char *)&error, sizeof(error), 0);
6809 rxi_FreePacket(packet);
6811 MUTEX_EXIT(&call->lock);
6812 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6815 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6816 * seconds) to ask the client to authenticate itself. The routine
6817 * issues a challenge to the client, which is obtained from the
6818 * security object associated with the connection */
6820 rxi_ChallengeEvent(struct rxevent *event,
6821 void *arg0, void *arg1, int tries)
6823 struct rx_connection *conn = arg0;
6826 rxevent_Put(&conn->challengeEvent);
6828 /* If there are no active calls it is not worth re-issuing the
6829 * challenge. If the client issues another call on this connection
6830 * the challenge can be requested at that time.
6832 if (!rxi_HasActiveCalls(conn))
6835 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6836 struct rx_packet *packet;
6837 struct clock when, now;
6840 /* We've failed to authenticate for too long.
6841 * Reset any calls waiting for authentication;
6842 * they are all in RX_STATE_PRECALL.
6846 MUTEX_ENTER(&conn->conn_call_lock);
6847 for (i = 0; i < RX_MAXCALLS; i++) {
6848 struct rx_call *call = conn->call[i];
6850 MUTEX_ENTER(&call->lock);
6851 if (call->state == RX_STATE_PRECALL) {
6852 rxi_CallError(call, RX_CALL_DEAD);
6853 rxi_SendCallAbort(call, NULL, 0, 0);
6855 MUTEX_EXIT(&call->lock);
6858 MUTEX_EXIT(&conn->conn_call_lock);
6862 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6864 /* If there's no packet available, do this later. */
6865 RXS_GetChallenge(conn->securityObject, conn, packet);
6866 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6867 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6868 rxi_FreePacket(packet);
6870 clock_GetTime(&now);
6872 when.sec += RX_CHALLENGE_TIMEOUT;
6873 conn->challengeEvent =
6874 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6879 /* Call this routine to start requesting the client to authenticate
6880 * itself. This will continue until authentication is established,
6881 * the call times out, or an invalid response is returned. The
6882 * security object associated with the connection is asked to create
6883 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6884 * defined earlier. */
6886 rxi_ChallengeOn(struct rx_connection *conn)
6888 if (!conn->challengeEvent) {
6889 RXS_CreateChallenge(conn->securityObject, conn);
6890 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6895 /* rxi_ComputeRoundTripTime is called with peer locked. */
6896 /* peer may be null */
6898 rxi_ComputeRoundTripTime(struct rx_packet *p,
6899 struct rx_ackPacket *ack,
6900 struct rx_call *call,
6901 struct rx_peer *peer,
6904 struct clock thisRtt, *sentp;
6908 /* If the ACK is delayed, then do nothing */
6909 if (ack->reason == RX_ACK_DELAY)
6912 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6913 * their RTT multiple times, so only include the RTT of the last packet
6915 if (p->flags & RX_JUMBO_PACKET)
6918 /* Use the serial number to determine which transmission the ACK is for,
6919 * and set the sent time to match this. If we have no serial number, then
6920 * only use the ACK for RTT calculations if the packet has not been
6924 serial = ntohl(ack->serial);
6926 if (serial == p->header.serial) {
6927 sentp = &p->timeSent;
6928 } else if (serial == p->firstSerial) {
6929 sentp = &p->firstSent;
6930 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6931 sentp = &p->firstSent;
6935 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6936 sentp = &p->firstSent;
6943 if (clock_Lt(&thisRtt, sentp))
6944 return; /* somebody set the clock back, don't count this time. */
6946 clock_Sub(&thisRtt, sentp);
6947 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6948 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6950 if (clock_IsZero(&thisRtt)) {
6952 * The actual round trip time is shorter than the
6953 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6954 * Since we can't tell which at the moment we will assume 1ms.
6956 thisRtt.usec = 1000;
6959 if (rx_stats_active) {
6960 MUTEX_ENTER(&rx_stats_mutex);
6961 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6962 rx_stats.minRtt = thisRtt;
6963 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6964 if (thisRtt.sec > 60) {
6965 MUTEX_EXIT(&rx_stats_mutex);
6966 return; /* somebody set the clock ahead */
6968 rx_stats.maxRtt = thisRtt;
6970 clock_Add(&rx_stats.totalRtt, &thisRtt);
6971 rx_atomic_inc(&rx_stats.nRttSamples);
6972 MUTEX_EXIT(&rx_stats_mutex);
6975 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6977 /* Apply VanJacobson round-trip estimations */
6982 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6983 * srtt is stored as fixed point with 3 bits after the binary
6984 * point (i.e., scaled by 8). The following magic is
6985 * equivalent to the smoothing algorithm in rfc793 with an
6986 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6987 * srtt'*8 = rtt + srtt*7
6988 * srtt'*8 = srtt*8 + rtt - srtt
6989 * srtt' = srtt + rtt/8 - srtt/8
6990 * srtt' = srtt + (rtt - srtt)/8
6993 delta = _8THMSEC(&thisRtt) - call->rtt;
6994 call->rtt += (delta >> 3);
6997 * We accumulate a smoothed rtt variance (actually, a smoothed
6998 * mean difference), then set the retransmit timer to smoothed
6999 * rtt + 4 times the smoothed variance (was 2x in van's original
7000 * paper, but 4x works better for me, and apparently for him as
7002 * rttvar is stored as
7003 * fixed point with 2 bits after the binary point (scaled by
7004 * 4). The following is equivalent to rfc793 smoothing with
7005 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
7006 * rttvar'*4 = rttvar*3 + |delta|
7007 * rttvar'*4 = rttvar*4 + |delta| - rttvar
7008 * rttvar' = rttvar + |delta|/4 - rttvar/4
7009 * rttvar' = rttvar + (|delta| - rttvar)/4
7010 * This replaces rfc793's wired-in beta.
7011 * dev*4 = dev*4 + (|actual - expected| - dev)
7017 delta -= (call->rtt_dev << 1);
7018 call->rtt_dev += (delta >> 3);
7020 /* I don't have a stored RTT so I start with this value. Since I'm
7021 * probably just starting a call, and will be pushing more data down
7022 * this, I expect congestion to increase rapidly. So I fudge a
7023 * little, and I set deviance to half the rtt. In practice,
7024 * deviance tends to approach something a little less than
7025 * half the smoothed rtt. */
7026 call->rtt = _8THMSEC(&thisRtt) + 8;
7027 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
7029 /* the smoothed RTT time is RTT + 4*MDEV
7031 * We allow a user specified minimum to be set for this, to allow clamping
7032 * at a minimum value in the same way as TCP. In addition, we have to allow
7033 * for the possibility that this packet is answered by a delayed ACK, so we
7034 * add on a fixed 200ms to account for that timer expiring.
7037 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
7038 rx_minPeerTimeout) + 200;
7039 clock_Zero(&call->rto);
7040 clock_Addmsec(&call->rto, rtt_timeout);
7042 /* Update the peer, so any new calls start with our values */
7043 peer->rtt_dev = call->rtt_dev;
7044 peer->rtt = call->rtt;
7046 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
7047 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
7051 /* Find all server connections that have not been active for a long time, and
7054 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7057 struct clock now, when;
7058 struct rxevent *event;
7059 clock_GetTime(&now);
7061 /* Find server connection structures that haven't been used for
7062 * greater than rx_idleConnectionTime */
7064 struct rx_connection **conn_ptr, **conn_end;
7065 int i, havecalls = 0;
7066 MUTEX_ENTER(&rx_connHashTable_lock);
7067 for (conn_ptr = &rx_connHashTable[0], conn_end =
7068 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7070 struct rx_connection *conn, *next;
7071 struct rx_call *call;
7075 for (conn = *conn_ptr; conn; conn = next) {
7076 /* XXX -- Shouldn't the connection be locked? */
7079 for (i = 0; i < RX_MAXCALLS; i++) {
7080 call = conn->call[i];
7084 code = MUTEX_TRYENTER(&call->lock);
7087 result = rxi_CheckCall(call, 1);
7088 MUTEX_EXIT(&call->lock);
7090 /* If CheckCall freed the call, it might
7091 * have destroyed the connection as well,
7092 * which screws up the linked lists.
7098 if (conn->type == RX_SERVER_CONNECTION) {
7099 /* This only actually destroys the connection if
7100 * there are no outstanding calls */
7101 MUTEX_ENTER(&conn->conn_data_lock);
7102 MUTEX_ENTER(&rx_refcnt_mutex);
7103 if (!havecalls && !conn->refCount
7104 && ((conn->lastSendTime + rx_idleConnectionTime) <
7106 conn->refCount++; /* it will be decr in rx_DestroyConn */
7107 MUTEX_EXIT(&rx_refcnt_mutex);
7108 MUTEX_EXIT(&conn->conn_data_lock);
7109 #ifdef RX_ENABLE_LOCKS
7110 rxi_DestroyConnectionNoLock(conn);
7111 #else /* RX_ENABLE_LOCKS */
7112 rxi_DestroyConnection(conn);
7113 #endif /* RX_ENABLE_LOCKS */
7115 #ifdef RX_ENABLE_LOCKS
7117 MUTEX_EXIT(&rx_refcnt_mutex);
7118 MUTEX_EXIT(&conn->conn_data_lock);
7120 #endif /* RX_ENABLE_LOCKS */
7124 #ifdef RX_ENABLE_LOCKS
7125 while (rx_connCleanup_list) {
7126 struct rx_connection *conn;
7127 conn = rx_connCleanup_list;
7128 rx_connCleanup_list = rx_connCleanup_list->next;
7129 MUTEX_EXIT(&rx_connHashTable_lock);
7130 rxi_CleanupConnection(conn);
7131 MUTEX_ENTER(&rx_connHashTable_lock);
7133 MUTEX_EXIT(&rx_connHashTable_lock);
7134 #endif /* RX_ENABLE_LOCKS */
7137 /* Find any peer structures that haven't been used (haven't had an
7138 * associated connection) for greater than rx_idlePeerTime */
7140 struct rx_peer **peer_ptr, **peer_end;
7144 * Why do we need to hold the rx_peerHashTable_lock across
7145 * the incrementing of peer_ptr since the rx_peerHashTable
7146 * array is not changing? We don't.
7148 * By dropping the lock periodically we can permit other
7149 * activities to be performed while a rxi_ReapConnections
7150 * call is in progress. The goal of reap connections
7151 * is to clean up quickly without causing large amounts
7152 * of contention. Therefore, it is important that global
7153 * mutexes not be held for extended periods of time.
7155 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7156 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7158 struct rx_peer *peer, *next, *prev;
7160 MUTEX_ENTER(&rx_peerHashTable_lock);
7161 for (prev = peer = *peer_ptr; peer; peer = next) {
7163 code = MUTEX_TRYENTER(&peer->peer_lock);
7164 if ((code) && (peer->refCount == 0)
7165 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7166 struct opr_queue *cursor, *store;
7170 * now know that this peer object is one to be
7171 * removed from the hash table. Once it is removed
7172 * it can't be referenced by other threads.
7173 * Lets remove it first and decrement the struct
7174 * nPeerStructs count.
7176 if (peer == *peer_ptr) {
7182 if (rx_stats_active)
7183 rx_atomic_dec(&rx_stats.nPeerStructs);
7186 * Now if we hold references on 'prev' and 'next'
7187 * we can safely drop the rx_peerHashTable_lock
7188 * while we destroy this 'peer' object.
7194 MUTEX_EXIT(&rx_peerHashTable_lock);
7196 MUTEX_EXIT(&peer->peer_lock);
7197 MUTEX_DESTROY(&peer->peer_lock);
7199 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7200 unsigned int num_funcs;
7201 struct rx_interface_stat *rpc_stat
7202 = opr_queue_Entry(cursor, struct rx_interface_stat,
7207 opr_queue_Remove(&rpc_stat->entry);
7208 opr_queue_Remove(&rpc_stat->entryPeers);
7210 num_funcs = rpc_stat->stats[0].func_total;
7212 sizeof(rx_interface_stat_t) +
7213 rpc_stat->stats[0].func_total *
7214 sizeof(rx_function_entry_v1_t);
7216 rxi_Free(rpc_stat, space);
7218 MUTEX_ENTER(&rx_rpc_stats);
7219 rxi_rpc_peer_stat_cnt -= num_funcs;
7220 MUTEX_EXIT(&rx_rpc_stats);
7225 * Regain the rx_peerHashTable_lock and
7226 * decrement the reference count on 'prev'
7229 MUTEX_ENTER(&rx_peerHashTable_lock);
7236 MUTEX_EXIT(&peer->peer_lock);
7241 MUTEX_EXIT(&rx_peerHashTable_lock);
7245 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7246 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7247 * GC, just below. Really, we shouldn't have to keep moving packets from
7248 * one place to another, but instead ought to always know if we can
7249 * afford to hold onto a packet in its particular use. */
7250 MUTEX_ENTER(&rx_freePktQ_lock);
7251 if (rx_waitingForPackets) {
7252 rx_waitingForPackets = 0;
7253 #ifdef RX_ENABLE_LOCKS
7254 CV_BROADCAST(&rx_waitingForPackets_cv);
7256 osi_rxWakeup(&rx_waitingForPackets);
7259 MUTEX_EXIT(&rx_freePktQ_lock);
7262 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7263 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7264 rxevent_Put(&event);
7268 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7269 * rx.h is sort of strange this is better. This is called with a security
7270 * object before it is discarded. Each connection using a security object has
7271 * its own refcount to the object so it won't actually be freed until the last
7272 * connection is destroyed.
7274 * This is the only rxs module call. A hold could also be written but no one
7278 rxs_Release(struct rx_securityClass *aobj)
7280 return RXS_Close(aobj);
7288 #define TRACE_OPTION_RX_DEBUG 16
7296 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7297 0, KEY_QUERY_VALUE, &parmKey);
7298 if (code != ERROR_SUCCESS)
7301 dummyLen = sizeof(TraceOption);
7302 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7303 (BYTE *) &TraceOption, &dummyLen);
7304 if (code == ERROR_SUCCESS) {
7305 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7307 RegCloseKey (parmKey);
7308 #endif /* AFS_NT40_ENV */
7313 rx_DebugOnOff(int on)
7317 rxdebug_active = on;
7323 rx_StatsOnOff(int on)
7325 rx_stats_active = on;
7329 /* Don't call this debugging routine directly; use dpf */
7331 rxi_DebugPrint(char *format, ...)
7340 va_start(ap, format);
7342 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7345 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7347 OutputDebugString(msg);
7353 va_start(ap, format);
7355 clock_GetTime(&now);
7356 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7357 (unsigned int)now.usec);
7358 vfprintf(rx_Log, format, ap);
7366 * This function is used to process the rx_stats structure that is local
7367 * to a process as well as an rx_stats structure received from a remote
7368 * process (via rxdebug). Therefore, it needs to do minimal version
7372 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7373 afs_int32 freePackets, char version)
7377 if (size != sizeof(struct rx_statistics)) {
7379 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7380 size, sizeof(struct rx_statistics));
7383 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7386 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7387 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7388 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7389 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7390 s->specialPktAllocFailures);
7392 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7393 s->receivePktAllocFailures, s->sendPktAllocFailures,
7394 s->specialPktAllocFailures);
7398 " greedy %u, " "bogusReads %u (last from host %x), "
7399 "noPackets %u, " "noBuffers %u, " "selects %u, "
7400 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7401 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7402 s->selects, s->sendSelects);
7404 fprintf(file, " packets read: ");
7405 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7406 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7408 fprintf(file, "\n");
7411 " other read counters: data %u, " "ack %u, " "dup %u "
7412 "spurious %u " "dally %u\n", s->dataPacketsRead,
7413 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7414 s->ignorePacketDally);
7416 fprintf(file, " packets sent: ");
7417 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7418 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7420 fprintf(file, "\n");
7423 " other send counters: ack %u, " "data %u (not resends), "
7424 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7425 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7426 s->dataPacketsPushed, s->ignoreAckedPacket);
7429 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7430 s->netSendFailures, (int)s->fatalErrors);
7432 if (s->nRttSamples) {
7433 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7434 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7436 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7437 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7441 " %d server connections, " "%d client connections, "
7442 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7443 s->nServerConns, s->nClientConns, s->nPeerStructs,
7444 s->nCallStructs, s->nFreeCallStructs);
7446 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7447 fprintf(file, " %d clock updates\n", clock_nUpdates);
7451 /* for backward compatibility */
7453 rx_PrintStats(FILE * file)
7455 MUTEX_ENTER(&rx_stats_mutex);
7456 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7457 sizeof(rx_stats), rx_nFreePackets,
7459 MUTEX_EXIT(&rx_stats_mutex);
7463 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7465 fprintf(file, "Peer %x.%d.\n",
7466 ntohl(peer->host), (int)ntohs(peer->port));
7469 " Rtt %d, " "total sent %d, " "resent %d\n",
7470 peer->rtt, peer->nSent, peer->reSends);
7472 fprintf(file, " Packet size %d\n", peer->ifMTU);
7476 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7478 * This mutex protects the following static variables:
7482 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7483 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7485 #define LOCK_RX_DEBUG
7486 #define UNLOCK_RX_DEBUG
7487 #endif /* AFS_PTHREAD_ENV */
7489 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7491 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7492 u_char type, void *inputData, size_t inputLength,
7493 void *outputData, size_t outputLength)
7495 static afs_int32 counter = 100;
7496 time_t waitTime, waitCount;
7497 struct rx_header theader;
7500 struct timeval tv_now, tv_wake, tv_delta;
7501 struct sockaddr_in taddr, faddr;
7515 tp = &tbuffer[sizeof(struct rx_header)];
7516 taddr.sin_family = AF_INET;
7517 taddr.sin_port = remotePort;
7518 taddr.sin_addr.s_addr = remoteAddr;
7519 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7520 taddr.sin_len = sizeof(struct sockaddr_in);
7523 memset(&theader, 0, sizeof(theader));
7524 theader.epoch = htonl(999);
7526 theader.callNumber = htonl(counter);
7529 theader.type = type;
7530 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7531 theader.serviceId = 0;
7533 memcpy(tbuffer, &theader, sizeof(theader));
7534 memcpy(tp, inputData, inputLength);
7536 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7537 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7539 /* see if there's a packet available */
7540 gettimeofday(&tv_wake, NULL);
7541 tv_wake.tv_sec += waitTime;
7544 FD_SET(socket, &imask);
7545 tv_delta.tv_sec = tv_wake.tv_sec;
7546 tv_delta.tv_usec = tv_wake.tv_usec;
7547 gettimeofday(&tv_now, NULL);
7549 if (tv_delta.tv_usec < tv_now.tv_usec) {
7551 tv_delta.tv_usec += 1000000;
7554 tv_delta.tv_usec -= tv_now.tv_usec;
7556 if (tv_delta.tv_sec < tv_now.tv_sec) {
7560 tv_delta.tv_sec -= tv_now.tv_sec;
7563 code = select(0, &imask, 0, 0, &tv_delta);
7564 #else /* AFS_NT40_ENV */
7565 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7566 #endif /* AFS_NT40_ENV */
7567 if (code == 1 && FD_ISSET(socket, &imask)) {
7568 /* now receive a packet */
7569 faddrLen = sizeof(struct sockaddr_in);
7571 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7572 (struct sockaddr *)&faddr, &faddrLen);
7575 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7576 if (counter == ntohl(theader.callNumber))
7584 /* see if we've timed out */
7592 code -= sizeof(struct rx_header);
7593 if (code > outputLength)
7594 code = outputLength;
7595 memcpy(outputData, tp, code);
7598 #endif /* RXDEBUG */
7601 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7602 afs_uint16 remotePort, struct rx_debugStats * stat,
7603 afs_uint32 * supportedValues)
7605 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7607 struct rx_debugIn in;
7609 *supportedValues = 0;
7610 in.type = htonl(RX_DEBUGI_GETSTATS);
7613 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7614 &in, sizeof(in), stat, sizeof(*stat));
7617 * If the call was successful, fixup the version and indicate
7618 * what contents of the stat structure are valid.
7619 * Also do net to host conversion of fields here.
7623 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7624 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7626 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7627 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7629 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7630 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7632 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7633 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7635 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7636 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7638 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7639 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7641 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7642 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7644 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7645 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7647 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7648 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7650 stat->nFreePackets = ntohl(stat->nFreePackets);
7651 stat->packetReclaims = ntohl(stat->packetReclaims);
7652 stat->callsExecuted = ntohl(stat->callsExecuted);
7653 stat->nWaiting = ntohl(stat->nWaiting);
7654 stat->idleThreads = ntohl(stat->idleThreads);
7655 stat->nWaited = ntohl(stat->nWaited);
7656 stat->nPackets = ntohl(stat->nPackets);
7665 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7666 afs_uint16 remotePort, struct rx_statistics * stat,
7667 afs_uint32 * supportedValues)
7669 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7671 struct rx_debugIn in;
7672 afs_int32 *lp = (afs_int32 *) stat;
7676 * supportedValues is currently unused, but added to allow future
7677 * versioning of this function.
7680 *supportedValues = 0;
7681 in.type = htonl(RX_DEBUGI_RXSTATS);
7683 memset(stat, 0, sizeof(*stat));
7685 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7686 &in, sizeof(in), stat, sizeof(*stat));
7691 * Do net to host conversion here
7694 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7705 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7706 afs_uint16 remotePort, size_t version_length,
7709 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7711 return MakeDebugCall(socket, remoteAddr, remotePort,
7712 RX_PACKET_TYPE_VERSION, a, 1, version,
7720 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7721 afs_uint16 remotePort, afs_int32 * nextConnection,
7722 int allConnections, afs_uint32 debugSupportedValues,
7723 struct rx_debugConn * conn,
7724 afs_uint32 * supportedValues)
7726 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7728 struct rx_debugIn in;
7732 * supportedValues is currently unused, but added to allow future
7733 * versioning of this function.
7736 *supportedValues = 0;
7737 if (allConnections) {
7738 in.type = htonl(RX_DEBUGI_GETALLCONN);
7740 in.type = htonl(RX_DEBUGI_GETCONN);
7742 in.index = htonl(*nextConnection);
7743 memset(conn, 0, sizeof(*conn));
7745 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7746 &in, sizeof(in), conn, sizeof(*conn));
7749 *nextConnection += 1;
7752 * Convert old connection format to new structure.
7755 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7756 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7757 #define MOVEvL(a) (conn->a = vL->a)
7759 /* any old or unrecognized version... */
7760 for (i = 0; i < RX_MAXCALLS; i++) {
7761 MOVEvL(callState[i]);
7762 MOVEvL(callMode[i]);
7763 MOVEvL(callFlags[i]);
7764 MOVEvL(callOther[i]);
7766 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7767 MOVEvL(secStats.type);
7768 MOVEvL(secStats.level);
7769 MOVEvL(secStats.flags);
7770 MOVEvL(secStats.expires);
7771 MOVEvL(secStats.packetsReceived);
7772 MOVEvL(secStats.packetsSent);
7773 MOVEvL(secStats.bytesReceived);
7774 MOVEvL(secStats.bytesSent);
7779 * Do net to host conversion here
7781 * I don't convert host or port since we are most likely
7782 * going to want these in NBO.
7784 conn->cid = ntohl(conn->cid);
7785 conn->serial = ntohl(conn->serial);
7786 for (i = 0; i < RX_MAXCALLS; i++) {
7787 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7789 conn->error = ntohl(conn->error);
7790 conn->secStats.flags = ntohl(conn->secStats.flags);
7791 conn->secStats.expires = ntohl(conn->secStats.expires);
7792 conn->secStats.packetsReceived =
7793 ntohl(conn->secStats.packetsReceived);
7794 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7795 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7796 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7797 conn->epoch = ntohl(conn->epoch);
7798 conn->natMTU = ntohl(conn->natMTU);
7807 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7808 afs_uint16 remotePort, afs_int32 * nextPeer,
7809 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7810 afs_uint32 * supportedValues)
7812 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7814 struct rx_debugIn in;
7817 * supportedValues is currently unused, but added to allow future
7818 * versioning of this function.
7821 *supportedValues = 0;
7822 in.type = htonl(RX_DEBUGI_GETPEER);
7823 in.index = htonl(*nextPeer);
7824 memset(peer, 0, sizeof(*peer));
7826 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7827 &in, sizeof(in), peer, sizeof(*peer));
7833 * Do net to host conversion here
7835 * I don't convert host or port since we are most likely
7836 * going to want these in NBO.
7838 peer->ifMTU = ntohs(peer->ifMTU);
7839 peer->idleWhen = ntohl(peer->idleWhen);
7840 peer->refCount = ntohs(peer->refCount);
7841 peer->rtt = ntohl(peer->rtt);
7842 peer->rtt_dev = ntohl(peer->rtt_dev);
7843 peer->timeout.sec = 0;
7844 peer->timeout.usec = 0;
7845 peer->nSent = ntohl(peer->nSent);
7846 peer->reSends = ntohl(peer->reSends);
7847 peer->natMTU = ntohs(peer->natMTU);
7848 peer->maxMTU = ntohs(peer->maxMTU);
7849 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7850 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7851 peer->MTU = ntohs(peer->MTU);
7852 peer->cwind = ntohs(peer->cwind);
7853 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7854 peer->congestSeq = ntohs(peer->congestSeq);
7855 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7856 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7857 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7858 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7867 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7868 struct rx_debugPeer * peerStats)
7871 afs_int32 error = 1; /* default to "did not succeed" */
7872 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7874 MUTEX_ENTER(&rx_peerHashTable_lock);
7875 for(tp = rx_peerHashTable[hashValue];
7876 tp != NULL; tp = tp->next) {
7877 if (tp->host == peerHost)
7883 MUTEX_EXIT(&rx_peerHashTable_lock);
7887 MUTEX_ENTER(&tp->peer_lock);
7888 peerStats->host = tp->host;
7889 peerStats->port = tp->port;
7890 peerStats->ifMTU = tp->ifMTU;
7891 peerStats->idleWhen = tp->idleWhen;
7892 peerStats->refCount = tp->refCount;
7893 peerStats->burstSize = 0;
7894 peerStats->burst = 0;
7895 peerStats->burstWait.sec = 0;
7896 peerStats->burstWait.usec = 0;
7897 peerStats->rtt = tp->rtt;
7898 peerStats->rtt_dev = tp->rtt_dev;
7899 peerStats->timeout.sec = 0;
7900 peerStats->timeout.usec = 0;
7901 peerStats->nSent = tp->nSent;
7902 peerStats->reSends = tp->reSends;
7903 peerStats->natMTU = tp->natMTU;
7904 peerStats->maxMTU = tp->maxMTU;
7905 peerStats->maxDgramPackets = tp->maxDgramPackets;
7906 peerStats->ifDgramPackets = tp->ifDgramPackets;
7907 peerStats->MTU = tp->MTU;
7908 peerStats->cwind = tp->cwind;
7909 peerStats->nDgramPackets = tp->nDgramPackets;
7910 peerStats->congestSeq = tp->congestSeq;
7911 peerStats->bytesSent.high = tp->bytesSent >> 32;
7912 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7913 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7914 peerStats->bytesReceived.low
7915 = tp->bytesReceived & MAX_AFS_UINT32;
7916 MUTEX_EXIT(&tp->peer_lock);
7918 MUTEX_ENTER(&rx_peerHashTable_lock);
7921 MUTEX_EXIT(&rx_peerHashTable_lock);
7929 struct rx_serverQueueEntry *np;
7932 struct rx_call *call;
7933 struct rx_serverQueueEntry *sq;
7936 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7937 return; /* Already shutdown. */
7941 #ifndef AFS_PTHREAD_ENV
7942 FD_ZERO(&rx_selectMask);
7943 #endif /* AFS_PTHREAD_ENV */
7944 rxi_dataQuota = RX_MAX_QUOTA;
7945 #ifndef AFS_PTHREAD_ENV
7947 #endif /* AFS_PTHREAD_ENV */
7950 #ifndef AFS_PTHREAD_ENV
7951 #ifndef AFS_USE_GETTIMEOFDAY
7953 #endif /* AFS_USE_GETTIMEOFDAY */
7954 #endif /* AFS_PTHREAD_ENV */
7956 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7957 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7958 opr_queue_Remove(&call->entry);
7959 rxi_Free(call, sizeof(struct rx_call));
7962 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7963 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7965 opr_queue_Remove(&sq->entry);
7970 struct rx_peer **peer_ptr, **peer_end;
7971 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7972 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7974 struct rx_peer *peer, *next;
7976 MUTEX_ENTER(&rx_peerHashTable_lock);
7977 for (peer = *peer_ptr; peer; peer = next) {
7978 struct opr_queue *cursor, *store;
7981 MUTEX_ENTER(&rx_rpc_stats);
7982 MUTEX_ENTER(&peer->peer_lock);
7983 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7984 unsigned int num_funcs;
7985 struct rx_interface_stat *rpc_stat
7986 = opr_queue_Entry(cursor, struct rx_interface_stat,
7990 opr_queue_Remove(&rpc_stat->entry);
7991 opr_queue_Remove(&rpc_stat->entryPeers);
7992 num_funcs = rpc_stat->stats[0].func_total;
7994 sizeof(rx_interface_stat_t) +
7995 rpc_stat->stats[0].func_total *
7996 sizeof(rx_function_entry_v1_t);
7998 rxi_Free(rpc_stat, space);
8000 /* rx_rpc_stats must be held */
8001 rxi_rpc_peer_stat_cnt -= num_funcs;
8003 MUTEX_EXIT(&peer->peer_lock);
8004 MUTEX_EXIT(&rx_rpc_stats);
8008 if (rx_stats_active)
8009 rx_atomic_dec(&rx_stats.nPeerStructs);
8011 MUTEX_EXIT(&rx_peerHashTable_lock);
8014 for (i = 0; i < RX_MAX_SERVICES; i++) {
8016 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
8018 for (i = 0; i < rx_hashTableSize; i++) {
8019 struct rx_connection *tc, *ntc;
8020 MUTEX_ENTER(&rx_connHashTable_lock);
8021 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8023 for (j = 0; j < RX_MAXCALLS; j++) {
8025 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8028 rxi_Free(tc, sizeof(*tc));
8030 MUTEX_EXIT(&rx_connHashTable_lock);
8033 MUTEX_ENTER(&freeSQEList_lock);
8035 while ((np = rx_FreeSQEList)) {
8036 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8037 MUTEX_DESTROY(&np->lock);
8038 rxi_Free(np, sizeof(*np));
8041 MUTEX_EXIT(&freeSQEList_lock);
8042 MUTEX_DESTROY(&freeSQEList_lock);
8043 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8044 MUTEX_DESTROY(&rx_connHashTable_lock);
8045 MUTEX_DESTROY(&rx_peerHashTable_lock);
8046 MUTEX_DESTROY(&rx_serverPool_lock);
8048 osi_Free(rx_connHashTable,
8049 rx_hashTableSize * sizeof(struct rx_connection *));
8050 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8052 UNPIN(rx_connHashTable,
8053 rx_hashTableSize * sizeof(struct rx_connection *));
8054 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8056 rxi_FreeAllPackets();
8058 MUTEX_ENTER(&rx_quota_mutex);
8059 rxi_dataQuota = RX_MAX_QUOTA;
8060 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8061 MUTEX_EXIT(&rx_quota_mutex);
8067 * Routines to implement connection specific data.
8071 rx_KeyCreate(rx_destructor_t rtn)
8074 MUTEX_ENTER(&rxi_keyCreate_lock);
8075 key = rxi_keyCreate_counter++;
8076 rxi_keyCreate_destructor = (rx_destructor_t *)
8077 realloc((void *)rxi_keyCreate_destructor,
8078 (key + 1) * sizeof(rx_destructor_t));
8079 rxi_keyCreate_destructor[key] = rtn;
8080 MUTEX_EXIT(&rxi_keyCreate_lock);
8085 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8088 MUTEX_ENTER(&conn->conn_data_lock);
8089 if (!conn->specific) {
8090 conn->specific = malloc((key + 1) * sizeof(void *));
8091 for (i = 0; i < key; i++)
8092 conn->specific[i] = NULL;
8093 conn->nSpecific = key + 1;
8094 conn->specific[key] = ptr;
8095 } else if (key >= conn->nSpecific) {
8096 conn->specific = (void **)
8097 realloc(conn->specific, (key + 1) * sizeof(void *));
8098 for (i = conn->nSpecific; i < key; i++)
8099 conn->specific[i] = NULL;
8100 conn->nSpecific = key + 1;
8101 conn->specific[key] = ptr;
8103 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8104 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8105 conn->specific[key] = ptr;
8107 MUTEX_EXIT(&conn->conn_data_lock);
8111 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8114 MUTEX_ENTER(&svc->svc_data_lock);
8115 if (!svc->specific) {
8116 svc->specific = malloc((key + 1) * sizeof(void *));
8117 for (i = 0; i < key; i++)
8118 svc->specific[i] = NULL;
8119 svc->nSpecific = key + 1;
8120 svc->specific[key] = ptr;
8121 } else if (key >= svc->nSpecific) {
8122 svc->specific = (void **)
8123 realloc(svc->specific, (key + 1) * sizeof(void *));
8124 for (i = svc->nSpecific; i < key; i++)
8125 svc->specific[i] = NULL;
8126 svc->nSpecific = key + 1;
8127 svc->specific[key] = ptr;
8129 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8130 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8131 svc->specific[key] = ptr;
8133 MUTEX_EXIT(&svc->svc_data_lock);
8137 rx_GetSpecific(struct rx_connection *conn, int key)
8140 MUTEX_ENTER(&conn->conn_data_lock);
8141 if (key >= conn->nSpecific)
8144 ptr = conn->specific[key];
8145 MUTEX_EXIT(&conn->conn_data_lock);
8150 rx_GetServiceSpecific(struct rx_service *svc, int key)
8153 MUTEX_ENTER(&svc->svc_data_lock);
8154 if (key >= svc->nSpecific)
8157 ptr = svc->specific[key];
8158 MUTEX_EXIT(&svc->svc_data_lock);
8163 #endif /* !KERNEL */
8166 * processStats is a queue used to store the statistics for the local
8167 * process. Its contents are similar to the contents of the rpcStats
8168 * queue on a rx_peer structure, but the actual data stored within
8169 * this queue contains totals across the lifetime of the process (assuming
8170 * the stats have not been reset) - unlike the per peer structures
8171 * which can come and go based upon the peer lifetime.
8174 static struct opr_queue processStats = { &processStats, &processStats };
8177 * peerStats is a queue used to store the statistics for all peer structs.
8178 * Its contents are the union of all the peer rpcStats queues.
8181 static struct opr_queue peerStats = { &peerStats, &peerStats };
8184 * rxi_monitor_processStats is used to turn process wide stat collection
8188 static int rxi_monitor_processStats = 0;
8191 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8194 static int rxi_monitor_peerStats = 0;
8198 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8200 rpc_stat->invocations = 0;
8201 rpc_stat->bytes_sent = 0;
8202 rpc_stat->bytes_rcvd = 0;
8203 rpc_stat->queue_time_sum.sec = 0;
8204 rpc_stat->queue_time_sum.usec = 0;
8205 rpc_stat->queue_time_sum_sqr.sec = 0;
8206 rpc_stat->queue_time_sum_sqr.usec = 0;
8207 rpc_stat->queue_time_min.sec = 9999999;
8208 rpc_stat->queue_time_min.usec = 9999999;
8209 rpc_stat->queue_time_max.sec = 0;
8210 rpc_stat->queue_time_max.usec = 0;
8211 rpc_stat->execution_time_sum.sec = 0;
8212 rpc_stat->execution_time_sum.usec = 0;
8213 rpc_stat->execution_time_sum_sqr.sec = 0;
8214 rpc_stat->execution_time_sum_sqr.usec = 0;
8215 rpc_stat->execution_time_min.sec = 9999999;
8216 rpc_stat->execution_time_min.usec = 9999999;
8217 rpc_stat->execution_time_max.sec = 0;
8218 rpc_stat->execution_time_max.usec = 0;
8222 * Given all of the information for a particular rpc
8223 * call, find or create (if requested) the stat structure for the rpc.
8226 * the queue of stats that will be updated with the new value
8228 * @param rxInterface
8229 * a unique number that identifies the rpc interface
8232 * the total number of functions in this interface. this is only
8233 * required if create is true
8236 * if true, this invocation was made to a server
8239 * the ip address of the remote host. this is only required if create
8240 * and addToPeerList are true
8243 * the port of the remote host. this is only required if create
8244 * and addToPeerList are true
8246 * @param addToPeerList
8247 * if != 0, add newly created stat to the global peer list
8250 * if a new stats structure is allocated, the counter will
8251 * be updated with the new number of allocated stat structures.
8252 * only required if create is true
8255 * if no stats structure exists, allocate one
8259 static rx_interface_stat_p
8260 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8261 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8262 afs_uint32 remotePort, int addToPeerList,
8263 unsigned int *counter, int create)
8265 rx_interface_stat_p rpc_stat = NULL;
8266 struct opr_queue *cursor;
8269 * See if there's already a structure for this interface
8272 for (opr_queue_Scan(stats, cursor)) {
8273 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8275 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8276 && (rpc_stat->stats[0].remote_is_server == isServer))
8280 /* if they didn't ask us to create, we're done */
8282 if (opr_queue_IsEnd(stats, cursor))
8288 /* can't proceed without these */
8289 if (!totalFunc || !counter)
8293 * Didn't find a match so allocate a new structure and add it to the
8297 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8298 || (rpc_stat->stats[0].interfaceId != rxInterface)
8299 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8304 sizeof(rx_interface_stat_t) +
8305 totalFunc * sizeof(rx_function_entry_v1_t);
8307 rpc_stat = rxi_Alloc(space);
8308 if (rpc_stat == NULL)
8311 *counter += totalFunc;
8312 for (i = 0; i < totalFunc; i++) {
8313 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8314 rpc_stat->stats[i].remote_peer = remoteHost;
8315 rpc_stat->stats[i].remote_port = remotePort;
8316 rpc_stat->stats[i].remote_is_server = isServer;
8317 rpc_stat->stats[i].interfaceId = rxInterface;
8318 rpc_stat->stats[i].func_total = totalFunc;
8319 rpc_stat->stats[i].func_index = i;
8321 opr_queue_Prepend(stats, &rpc_stat->entry);
8322 if (addToPeerList) {
8323 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8330 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8332 rx_interface_stat_p rpc_stat;
8335 if (rxInterface == -1)
8338 MUTEX_ENTER(&rx_rpc_stats);
8339 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8342 totalFunc = rpc_stat->stats[0].func_total;
8343 for (i = 0; i < totalFunc; i++)
8344 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8346 MUTEX_EXIT(&rx_rpc_stats);
8351 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8353 rx_interface_stat_p rpc_stat;
8355 struct rx_peer * peer;
8357 if (rxInterface == -1)
8360 peer = rxi_FindPeer(peerHost, peerPort, 0);
8364 MUTEX_ENTER(&rx_rpc_stats);
8365 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8368 totalFunc = rpc_stat->stats[0].func_total;
8369 for (i = 0; i < totalFunc; i++)
8370 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8372 MUTEX_EXIT(&rx_rpc_stats);
8377 rx_CopyProcessRPCStats(afs_uint64 op)
8379 rx_interface_stat_p rpc_stat;
8380 rx_function_entry_v1_p rpcop_stat =
8381 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8382 int currentFunc = (op & MAX_AFS_UINT32);
8383 afs_int32 rxInterface = (op >> 32);
8385 if (!rxi_monitor_processStats)
8388 if (rxInterface == -1)
8391 if (rpcop_stat == NULL)
8394 MUTEX_ENTER(&rx_rpc_stats);
8395 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8398 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8399 sizeof(rx_function_entry_v1_t));
8400 MUTEX_EXIT(&rx_rpc_stats);
8402 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8409 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8411 rx_interface_stat_p rpc_stat;
8412 rx_function_entry_v1_p rpcop_stat =
8413 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8414 int currentFunc = (op & MAX_AFS_UINT32);
8415 afs_int32 rxInterface = (op >> 32);
8416 struct rx_peer *peer;
8418 if (!rxi_monitor_peerStats)
8421 if (rxInterface == -1)
8424 if (rpcop_stat == NULL)
8427 peer = rxi_FindPeer(peerHost, peerPort, 0);
8431 MUTEX_ENTER(&rx_rpc_stats);
8432 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8435 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8436 sizeof(rx_function_entry_v1_t));
8437 MUTEX_EXIT(&rx_rpc_stats);
8439 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8446 rx_ReleaseRPCStats(void *stats)
8449 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8453 * Given all of the information for a particular rpc
8454 * call, create (if needed) and update the stat totals for the rpc.
8457 * the queue of stats that will be updated with the new value
8459 * @param rxInterface
8460 * a unique number that identifies the rpc interface
8462 * @param currentFunc
8463 * the index of the function being invoked
8466 * the total number of functions in this interface
8469 * the amount of time this function waited for a thread
8472 * the amount of time this function invocation took to execute
8475 * the number bytes sent by this invocation
8478 * the number bytes received by this invocation
8481 * if true, this invocation was made to a server
8484 * the ip address of the remote host
8487 * the port of the remote host
8489 * @param addToPeerList
8490 * if != 0, add newly created stat to the global peer list
8493 * if a new stats structure is allocated, the counter will
8494 * be updated with the new number of allocated stat structures
8499 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8500 afs_uint32 currentFunc, afs_uint32 totalFunc,
8501 struct clock *queueTime, struct clock *execTime,
8502 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8503 afs_uint32 remoteHost, afs_uint32 remotePort,
8504 int addToPeerList, unsigned int *counter)
8507 rx_interface_stat_p rpc_stat;
8509 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8510 remoteHost, remotePort, addToPeerList, counter,
8518 * Increment the stats for this function
8521 rpc_stat->stats[currentFunc].invocations++;
8522 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8523 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8524 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8525 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8526 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8527 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8529 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8530 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8532 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8533 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8535 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8536 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8538 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8539 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8547 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8548 afs_uint32 currentFunc, afs_uint32 totalFunc,
8549 struct clock *queueTime, struct clock *execTime,
8550 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8554 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8557 MUTEX_ENTER(&rx_rpc_stats);
8559 if (rxi_monitor_peerStats) {
8560 MUTEX_ENTER(&peer->peer_lock);
8561 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8562 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8563 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8564 MUTEX_EXIT(&peer->peer_lock);
8567 if (rxi_monitor_processStats) {
8568 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8569 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8570 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8573 MUTEX_EXIT(&rx_rpc_stats);
8577 * Increment the times and count for a particular rpc function.
8579 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8580 * call rx_RecordCallStatistics instead, so the public version of this
8581 * function is left purely for legacy callers.
8584 * The peer who invoked the rpc
8586 * @param rxInterface
8587 * A unique number that identifies the rpc interface
8589 * @param currentFunc
8590 * The index of the function being invoked
8593 * The total number of functions in this interface
8596 * The amount of time this function waited for a thread
8599 * The amount of time this function invocation took to execute
8602 * The number bytes sent by this invocation
8605 * The number bytes received by this invocation
8608 * If true, this invocation was made to a server
8612 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8613 afs_uint32 currentFunc, afs_uint32 totalFunc,
8614 struct clock *queueTime, struct clock *execTime,
8615 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8621 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8622 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8624 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8625 queueTime, execTime, sent64, rcvd64,
8632 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8636 * IN callerVersion - the rpc stat version of the caller.
8638 * IN count - the number of entries to marshall.
8640 * IN stats - pointer to stats to be marshalled.
8642 * OUT ptr - Where to store the marshalled data.
8649 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8650 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8656 * We only support the first version
8658 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8659 *(ptr++) = stats->remote_peer;
8660 *(ptr++) = stats->remote_port;
8661 *(ptr++) = stats->remote_is_server;
8662 *(ptr++) = stats->interfaceId;
8663 *(ptr++) = stats->func_total;
8664 *(ptr++) = stats->func_index;
8665 *(ptr++) = stats->invocations >> 32;
8666 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8667 *(ptr++) = stats->bytes_sent >> 32;
8668 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8669 *(ptr++) = stats->bytes_rcvd >> 32;
8670 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8671 *(ptr++) = stats->queue_time_sum.sec;
8672 *(ptr++) = stats->queue_time_sum.usec;
8673 *(ptr++) = stats->queue_time_sum_sqr.sec;
8674 *(ptr++) = stats->queue_time_sum_sqr.usec;
8675 *(ptr++) = stats->queue_time_min.sec;
8676 *(ptr++) = stats->queue_time_min.usec;
8677 *(ptr++) = stats->queue_time_max.sec;
8678 *(ptr++) = stats->queue_time_max.usec;
8679 *(ptr++) = stats->execution_time_sum.sec;
8680 *(ptr++) = stats->execution_time_sum.usec;
8681 *(ptr++) = stats->execution_time_sum_sqr.sec;
8682 *(ptr++) = stats->execution_time_sum_sqr.usec;
8683 *(ptr++) = stats->execution_time_min.sec;
8684 *(ptr++) = stats->execution_time_min.usec;
8685 *(ptr++) = stats->execution_time_max.sec;
8686 *(ptr++) = stats->execution_time_max.usec;
8692 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8697 * IN callerVersion - the rpc stat version of the caller
8699 * OUT myVersion - the rpc stat version of this function
8701 * OUT clock_sec - local time seconds
8703 * OUT clock_usec - local time microseconds
8705 * OUT allocSize - the number of bytes allocated to contain stats
8707 * OUT statCount - the number stats retrieved from this process.
8709 * OUT stats - the actual stats retrieved from this process.
8713 * Returns void. If successful, stats will != NULL.
8717 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8718 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8719 size_t * allocSize, afs_uint32 * statCount,
8720 afs_uint32 ** stats)
8730 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8733 * Check to see if stats are enabled
8736 MUTEX_ENTER(&rx_rpc_stats);
8737 if (!rxi_monitor_processStats) {
8738 MUTEX_EXIT(&rx_rpc_stats);
8742 clock_GetTime(&now);
8743 *clock_sec = now.sec;
8744 *clock_usec = now.usec;
8747 * Allocate the space based upon the caller version
8749 * If the client is at an older version than we are,
8750 * we return the statistic data in the older data format, but
8751 * we still return our version number so the client knows we
8752 * are maintaining more data than it can retrieve.
8755 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8756 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8757 *statCount = rxi_rpc_process_stat_cnt;
8760 * This can't happen yet, but in the future version changes
8761 * can be handled by adding additional code here
8765 if (space > (size_t) 0) {
8767 ptr = *stats = rxi_Alloc(space);
8770 struct opr_queue *cursor;
8772 for (opr_queue_Scan(&processStats, cursor)) {
8773 struct rx_interface_stat *rpc_stat =
8774 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8776 * Copy the data based upon the caller version
8778 rx_MarshallProcessRPCStats(callerVersion,
8779 rpc_stat->stats[0].func_total,
8780 rpc_stat->stats, &ptr);
8786 MUTEX_EXIT(&rx_rpc_stats);
8791 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8795 * IN callerVersion - the rpc stat version of the caller
8797 * OUT myVersion - the rpc stat version of this function
8799 * OUT clock_sec - local time seconds
8801 * OUT clock_usec - local time microseconds
8803 * OUT allocSize - the number of bytes allocated to contain stats
8805 * OUT statCount - the number of stats retrieved from the individual
8808 * OUT stats - the actual stats retrieved from the individual peer structures.
8812 * Returns void. If successful, stats will != NULL.
8816 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8817 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8818 size_t * allocSize, afs_uint32 * statCount,
8819 afs_uint32 ** stats)
8829 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8832 * Check to see if stats are enabled
8835 MUTEX_ENTER(&rx_rpc_stats);
8836 if (!rxi_monitor_peerStats) {
8837 MUTEX_EXIT(&rx_rpc_stats);
8841 clock_GetTime(&now);
8842 *clock_sec = now.sec;
8843 *clock_usec = now.usec;
8846 * Allocate the space based upon the caller version
8848 * If the client is at an older version than we are,
8849 * we return the statistic data in the older data format, but
8850 * we still return our version number so the client knows we
8851 * are maintaining more data than it can retrieve.
8854 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8855 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8856 *statCount = rxi_rpc_peer_stat_cnt;
8859 * This can't happen yet, but in the future version changes
8860 * can be handled by adding additional code here
8864 if (space > (size_t) 0) {
8866 ptr = *stats = rxi_Alloc(space);
8869 struct opr_queue *cursor;
8871 for (opr_queue_Scan(&peerStats, cursor)) {
8872 struct rx_interface_stat *rpc_stat
8873 = opr_queue_Entry(cursor, struct rx_interface_stat,
8877 * Copy the data based upon the caller version
8879 rx_MarshallProcessRPCStats(callerVersion,
8880 rpc_stat->stats[0].func_total,
8881 rpc_stat->stats, &ptr);
8887 MUTEX_EXIT(&rx_rpc_stats);
8892 * rx_FreeRPCStats - free memory allocated by
8893 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8897 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8898 * rx_RetrievePeerRPCStats
8900 * IN allocSize - the number of bytes in stats.
8908 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8910 rxi_Free(stats, allocSize);
8914 * rx_queryProcessRPCStats - see if process rpc stat collection is
8915 * currently enabled.
8921 * Returns 0 if stats are not enabled != 0 otherwise
8925 rx_queryProcessRPCStats(void)
8928 MUTEX_ENTER(&rx_rpc_stats);
8929 rc = rxi_monitor_processStats;
8930 MUTEX_EXIT(&rx_rpc_stats);
8935 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8941 * Returns 0 if stats are not enabled != 0 otherwise
8945 rx_queryPeerRPCStats(void)
8948 MUTEX_ENTER(&rx_rpc_stats);
8949 rc = rxi_monitor_peerStats;
8950 MUTEX_EXIT(&rx_rpc_stats);
8955 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8965 rx_enableProcessRPCStats(void)
8967 MUTEX_ENTER(&rx_rpc_stats);
8968 rx_enable_stats = 1;
8969 rxi_monitor_processStats = 1;
8970 MUTEX_EXIT(&rx_rpc_stats);
8974 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8984 rx_enablePeerRPCStats(void)
8986 MUTEX_ENTER(&rx_rpc_stats);
8987 rx_enable_stats = 1;
8988 rxi_monitor_peerStats = 1;
8989 MUTEX_EXIT(&rx_rpc_stats);
8993 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
9003 rx_disableProcessRPCStats(void)
9005 struct opr_queue *cursor, *store;
9008 MUTEX_ENTER(&rx_rpc_stats);
9011 * Turn off process statistics and if peer stats is also off, turn
9015 rxi_monitor_processStats = 0;
9016 if (rxi_monitor_peerStats == 0) {
9017 rx_enable_stats = 0;
9020 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
9021 unsigned int num_funcs = 0;
9022 struct rx_interface_stat *rpc_stat
9023 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9025 opr_queue_Remove(&rpc_stat->entry);
9027 num_funcs = rpc_stat->stats[0].func_total;
9029 sizeof(rx_interface_stat_t) +
9030 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
9032 rxi_Free(rpc_stat, space);
9033 rxi_rpc_process_stat_cnt -= num_funcs;
9035 MUTEX_EXIT(&rx_rpc_stats);
9039 * rx_disablePeerRPCStats - stop rpc stat collection for peers
9049 rx_disablePeerRPCStats(void)
9051 struct rx_peer **peer_ptr, **peer_end;
9055 * Turn off peer statistics and if process stats is also off, turn
9059 rxi_monitor_peerStats = 0;
9060 if (rxi_monitor_processStats == 0) {
9061 rx_enable_stats = 0;
9064 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9065 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9067 struct rx_peer *peer, *next, *prev;
9069 MUTEX_ENTER(&rx_peerHashTable_lock);
9070 MUTEX_ENTER(&rx_rpc_stats);
9071 for (prev = peer = *peer_ptr; peer; peer = next) {
9073 code = MUTEX_TRYENTER(&peer->peer_lock);
9076 struct opr_queue *cursor, *store;
9078 if (prev == *peer_ptr) {
9089 MUTEX_EXIT(&rx_peerHashTable_lock);
9091 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9092 unsigned int num_funcs = 0;
9093 struct rx_interface_stat *rpc_stat
9094 = opr_queue_Entry(cursor, struct rx_interface_stat,
9097 opr_queue_Remove(&rpc_stat->entry);
9098 opr_queue_Remove(&rpc_stat->entryPeers);
9099 num_funcs = rpc_stat->stats[0].func_total;
9101 sizeof(rx_interface_stat_t) +
9102 rpc_stat->stats[0].func_total *
9103 sizeof(rx_function_entry_v1_t);
9105 rxi_Free(rpc_stat, space);
9106 rxi_rpc_peer_stat_cnt -= num_funcs;
9108 MUTEX_EXIT(&peer->peer_lock);
9110 MUTEX_ENTER(&rx_peerHashTable_lock);
9120 MUTEX_EXIT(&rx_rpc_stats);
9121 MUTEX_EXIT(&rx_peerHashTable_lock);
9126 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9131 * IN clearFlag - flag indicating which stats to clear
9139 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9141 struct opr_queue *cursor;
9143 MUTEX_ENTER(&rx_rpc_stats);
9145 for (opr_queue_Scan(&processStats, cursor)) {
9146 unsigned int num_funcs = 0, i;
9147 struct rx_interface_stat *rpc_stat
9148 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9150 num_funcs = rpc_stat->stats[0].func_total;
9151 for (i = 0; i < num_funcs; i++) {
9152 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9153 rpc_stat->stats[i].invocations = 0;
9155 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9156 rpc_stat->stats[i].bytes_sent = 0;
9158 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9159 rpc_stat->stats[i].bytes_rcvd = 0;
9161 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9162 rpc_stat->stats[i].queue_time_sum.sec = 0;
9163 rpc_stat->stats[i].queue_time_sum.usec = 0;
9165 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9166 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9167 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9169 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9170 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9171 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9173 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9174 rpc_stat->stats[i].queue_time_max.sec = 0;
9175 rpc_stat->stats[i].queue_time_max.usec = 0;
9177 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9178 rpc_stat->stats[i].execution_time_sum.sec = 0;
9179 rpc_stat->stats[i].execution_time_sum.usec = 0;
9181 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9182 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9183 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9185 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9186 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9187 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9189 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9190 rpc_stat->stats[i].execution_time_max.sec = 0;
9191 rpc_stat->stats[i].execution_time_max.usec = 0;
9196 MUTEX_EXIT(&rx_rpc_stats);
9200 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9205 * IN clearFlag - flag indicating which stats to clear
9213 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9215 struct opr_queue *cursor;
9217 MUTEX_ENTER(&rx_rpc_stats);
9219 for (opr_queue_Scan(&peerStats, cursor)) {
9220 unsigned int num_funcs, i;
9221 struct rx_interface_stat *rpc_stat
9222 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9224 num_funcs = rpc_stat->stats[0].func_total;
9225 for (i = 0; i < num_funcs; i++) {
9226 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9227 rpc_stat->stats[i].invocations = 0;
9229 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9230 rpc_stat->stats[i].bytes_sent = 0;
9232 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9233 rpc_stat->stats[i].bytes_rcvd = 0;
9235 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9236 rpc_stat->stats[i].queue_time_sum.sec = 0;
9237 rpc_stat->stats[i].queue_time_sum.usec = 0;
9239 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9240 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9241 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9243 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9244 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9245 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9247 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9248 rpc_stat->stats[i].queue_time_max.sec = 0;
9249 rpc_stat->stats[i].queue_time_max.usec = 0;
9251 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9252 rpc_stat->stats[i].execution_time_sum.sec = 0;
9253 rpc_stat->stats[i].execution_time_sum.usec = 0;
9255 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9256 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9257 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9259 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9260 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9261 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9263 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9264 rpc_stat->stats[i].execution_time_max.sec = 0;
9265 rpc_stat->stats[i].execution_time_max.usec = 0;
9270 MUTEX_EXIT(&rx_rpc_stats);
9274 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9275 * is authorized to enable/disable/clear RX statistics.
9277 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9280 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9282 rxi_rxstat_userok = proc;
9286 rx_RxStatUserOk(struct rx_call *call)
9288 if (!rxi_rxstat_userok)
9290 return rxi_rxstat_userok(call);
9295 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9296 * function in the MSVC runtime DLL (msvcrt.dll).
9298 * Note: the system serializes calls to this function.
9301 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9302 DWORD reason, /* reason function is being called */
9303 LPVOID reserved) /* reserved for future use */
9306 case DLL_PROCESS_ATTACH:
9307 /* library is being attached to a process */
9311 case DLL_PROCESS_DETACH:
9318 #endif /* AFS_NT40_ENV */
9321 int rx_DumpCalls(FILE *outputFile, char *cookie)
9323 #ifdef RXDEBUG_PACKET
9324 #ifdef KDUMP_RX_LOCK
9325 struct rx_call_rx_lock *c;
9332 #define RXDPRINTF sprintf
9333 #define RXDPRINTOUT output
9335 #define RXDPRINTF fprintf
9336 #define RXDPRINTOUT outputFile
9339 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9341 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9344 for (c = rx_allCallsp; c; c = c->allNextp) {
9345 u_short rqc, tqc, iovqc;
9347 MUTEX_ENTER(&c->lock);
9348 rqc = opr_queue_Count(&c->rq);
9349 tqc = opr_queue_Count(&c->tq);
9350 iovqc = opr_queue_Count(&c->app.iovq);
9352 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, "
9353 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9354 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9355 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9356 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9357 #ifdef RX_ENABLE_LOCKS
9360 #ifdef RX_REFCOUNT_CHECK
9361 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9362 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9365 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,
9366 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9367 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9368 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9369 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9370 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9371 #ifdef RX_ENABLE_LOCKS
9372 , (afs_uint32)c->refCount
9374 #ifdef RX_REFCOUNT_CHECK
9375 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9378 MUTEX_EXIT(&c->lock);
9381 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9384 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9386 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9388 #endif /* RXDEBUG_PACKET */