2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
75 #include <opr/queue.h>
79 #include "rx_atomic.h"
80 #include "rx_globals.h"
82 #include "rx_internal.h"
89 #include "rx_packet.h"
90 #include "rx_server.h"
92 #include <afs/rxgen_consts.h>
95 #ifdef AFS_PTHREAD_ENV
97 int (*registerProgram) (pid_t, char *) = 0;
98 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
101 int (*registerProgram) (PROCESS, char *) = 0;
102 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
106 /* Local static routines */
107 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
108 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
109 struct rx_call *, struct rx_peer *,
111 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
113 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
114 void *dummy, int dummy2);
115 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
116 void *dummy, int dummy2);
117 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
118 void *unused, int unused2);
119 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
120 void *unused2, int unused3);
121 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
122 struct rx_packet *packet,
123 int istack, int force);
124 static void rxi_AckAll(struct rx_call *call);
125 static struct rx_connection
126 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
127 u_short serviceId, afs_uint32 cid,
128 afs_uint32 epoch, int type, u_int securityIndex,
129 int *unknownService);
130 static struct rx_packet
131 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
132 int istack, osi_socket socket,
133 afs_uint32 host, u_short port, int *tnop,
134 struct rx_call **newcallp);
135 static struct rx_packet
136 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
138 static struct rx_packet
139 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
140 struct rx_packet *np, int istack);
141 static struct rx_packet
142 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
143 struct rx_packet *np, int istack);
144 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
145 int *tnop, struct rx_call **newcallp);
146 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
147 static void rxi_ClearReceiveQueue(struct rx_call *call);
148 static void rxi_ResetCall(struct rx_call *call, int newcall);
149 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
150 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
151 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
152 static void rxi_KeepAliveOn(struct rx_call *call);
153 static void rxi_GrowMTUOn(struct rx_call *call);
154 static void rxi_ChallengeOn(struct rx_connection *conn);
155 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
156 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
157 static void rxi_CancelKeepAliveEvent(struct rx_call *call);
158 static void rxi_CancelDelayedAbortEvent(struct rx_call *call);
159 static void rxi_CancelGrowMTUEvent(struct rx_call *call);
161 #ifdef RX_ENABLE_LOCKS
163 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
164 rx_atomic_t rxi_start_in_error;
166 #endif /* RX_ENABLE_LOCKS */
168 /* Constant delay time before sending an acknowledge of the last packet
169 * received. This is to avoid sending an extra acknowledge when the
170 * client is about to make another call, anyway, or the server is
173 * The lastAckDelay may not exceeed 400ms without causing peers to
174 * unecessarily timeout.
176 struct clock rx_lastAckDelay = {0, 400000};
178 /* Constant delay time before sending a soft ack when none was requested.
179 * This is to make sure we send soft acks before the sender times out,
180 * Normally we wait and send a hard ack when the receiver consumes the packet
182 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
183 * will require changes to the peer's RTT calculations.
185 struct clock rx_softAckDelay = {0, 100000};
188 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
189 * currently allocated within rx. This number is used to allocate the
190 * memory required to return the statistics when queried.
191 * Protected by the rx_rpc_stats mutex.
194 static unsigned int rxi_rpc_peer_stat_cnt;
197 * rxi_rpc_process_stat_cnt counts the total number of local process stat
198 * structures currently allocated within rx. The number is used to allocate
199 * the memory required to return the statistics when queried.
200 * Protected by the rx_rpc_stats mutex.
203 static unsigned int rxi_rpc_process_stat_cnt;
206 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
207 * errors should be reported to the application when a call channel appears busy
208 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
209 * and there are other call channels in the connection that are not busy.
210 * If 0, we do not return errors upon receiving busy packets; we just keep
211 * trying on the same call channel until we hit a timeout.
213 static afs_int32 rxi_busyChannelError = 0;
215 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
216 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
218 /* Incoming calls wait on this queue when there are no available
219 * server processes */
220 struct opr_queue rx_incomingCallQueue;
222 /* Server processes wait on this queue when there are no appropriate
223 * calls to process */
224 struct opr_queue rx_idleServerQueue;
226 #if !defined(offsetof)
227 #include <stddef.h> /* for definition of offsetof() */
230 #ifdef RX_ENABLE_LOCKS
231 afs_kmutex_t rx_atomic_mutex;
234 /* Forward prototypes */
235 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
238 putConnection (struct rx_connection *conn) {
239 MUTEX_ENTER(&rx_refcnt_mutex);
241 MUTEX_EXIT(&rx_refcnt_mutex);
244 #ifdef AFS_PTHREAD_ENV
247 * Use procedural initialization of mutexes/condition variables
251 extern afs_kmutex_t rx_quota_mutex;
252 extern afs_kmutex_t rx_pthread_mutex;
253 extern afs_kmutex_t rx_packets_mutex;
254 extern afs_kmutex_t rx_refcnt_mutex;
255 extern afs_kmutex_t des_init_mutex;
256 extern afs_kmutex_t des_random_mutex;
257 extern afs_kmutex_t rx_clock_mutex;
258 extern afs_kmutex_t rxi_connCacheMutex;
259 extern afs_kmutex_t event_handler_mutex;
260 extern afs_kmutex_t listener_mutex;
261 extern afs_kmutex_t rx_if_init_mutex;
262 extern afs_kmutex_t rx_if_mutex;
264 extern afs_kcondvar_t rx_event_handler_cond;
265 extern afs_kcondvar_t rx_listener_cond;
267 static afs_kmutex_t epoch_mutex;
268 static afs_kmutex_t rx_init_mutex;
269 static afs_kmutex_t rx_debug_mutex;
270 static afs_kmutex_t rx_rpc_stats;
273 rxi_InitPthread(void)
275 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
286 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
287 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
288 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
289 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
291 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
292 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
294 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
295 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
297 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
298 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
299 #ifdef RX_ENABLE_LOCKS
302 #endif /* RX_LOCKS_DB */
303 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
304 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
306 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
308 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
310 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
312 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
313 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
314 #endif /* RX_ENABLE_LOCKS */
317 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
318 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
320 * The rx_stats_mutex mutex protects the following global variables:
321 * rxi_lowConnRefCount
322 * rxi_lowPeerRefCount
331 * The rx_quota_mutex mutex protects the following global variables:
339 * The rx_freePktQ_lock protects the following global variables:
344 * The rx_packets_mutex mutex protects the following global variables:
352 * The rx_pthread_mutex mutex protects the following global variables:
353 * rxi_fcfs_thread_num
356 #define INIT_PTHREAD_LOCKS
360 /* Variables for handling the minProcs implementation. availProcs gives the
361 * number of threads available in the pool at this moment (not counting dudes
362 * executing right now). totalMin gives the total number of procs required
363 * for handling all minProcs requests. minDeficit is a dynamic variable
364 * tracking the # of procs required to satisfy all of the remaining minProcs
366 * For fine grain locking to work, the quota check and the reservation of
367 * a server thread has to come while rxi_availProcs and rxi_minDeficit
368 * are locked. To this end, the code has been modified under #ifdef
369 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
370 * same time. A new function, ReturnToServerPool() returns the allocation.
372 * A call can be on several queue's (but only one at a time). When
373 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
374 * that no one else is touching the queue. To this end, we store the address
375 * of the queue lock in the call structure (under the call lock) when we
376 * put the call on a queue, and we clear the call_queue_lock when the
377 * call is removed from a queue (once the call lock has been obtained).
378 * This allows rxi_ResetCall to safely synchronize with others wishing
379 * to manipulate the queue.
382 #if defined(RX_ENABLE_LOCKS)
383 static afs_kmutex_t rx_rpc_stats;
386 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
387 ** pretty good that the next packet coming in is from the same connection
388 ** as the last packet, since we're send multiple packets in a transmit window.
390 struct rx_connection *rxLastConn = 0;
392 #ifdef RX_ENABLE_LOCKS
393 /* The locking hierarchy for rx fine grain locking is composed of these
396 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
397 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
398 * call->lock - locks call data fields.
399 * These are independent of each other:
400 * rx_freeCallQueue_lock
405 * serverQueueEntry->lock
406 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
408 * peer->lock - locks peer data fields.
409 * conn_data_lock - that more than one thread is not updating a conn data
410 * field at the same time.
421 * Do we need a lock to protect the peer field in the conn structure?
422 * conn->peer was previously a constant for all intents and so has no
423 * lock protecting this field. The multihomed client delta introduced
424 * a RX code change : change the peer field in the connection structure
425 * to that remote interface from which the last packet for this
426 * connection was sent out. This may become an issue if further changes
429 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
430 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
432 /* rxdb_fileID is used to identify the lock location, along with line#. */
433 static int rxdb_fileID = RXDB_FILE_RX;
434 #endif /* RX_LOCKS_DB */
435 #else /* RX_ENABLE_LOCKS */
436 #define SET_CALL_QUEUE_LOCK(C, L)
437 #define CLEAR_CALL_QUEUE_LOCK(C)
438 #endif /* RX_ENABLE_LOCKS */
439 struct rx_serverQueueEntry *rx_waitForPacket = 0;
441 /* ------------Exported Interfaces------------- */
443 /* This function allows rxkad to set the epoch to a suitably random number
444 * which rx_NewConnection will use in the future. The principle purpose is to
445 * get rxnull connections to use the same epoch as the rxkad connections do, at
446 * least once the first rxkad connection is established. This is important now
447 * that the host/port addresses aren't used in FindConnection: the uniqueness
448 * of epoch/cid matters and the start time won't do. */
450 #ifdef AFS_PTHREAD_ENV
452 * This mutex protects the following global variables:
456 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
457 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
461 #endif /* AFS_PTHREAD_ENV */
464 rx_SetEpoch(afs_uint32 epoch)
471 /* Initialize rx. A port number may be mentioned, in which case this
472 * becomes the default port number for any service installed later.
473 * If 0 is provided for the port number, a random port will be chosen
474 * by the kernel. Whether this will ever overlap anything in
475 * /etc/services is anybody's guess... Returns 0 on success, -1 on
480 rx_atomic_t rxinit_status = RX_ATOMIC_INIT(1);
483 rx_InitHost(u_int host, u_int port)
490 char *htable, *ptable;
495 if (!rx_atomic_test_and_clear_bit(&rxinit_status, 0))
496 return 0; /* already started */
502 if (afs_winsockInit() < 0)
508 * Initialize anything necessary to provide a non-premptive threading
511 rxi_InitializeThreadSupport();
514 /* Allocate and initialize a socket for client and perhaps server
517 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
518 if (rx_socket == OSI_NULLSOCKET) {
521 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
524 #endif /* RX_LOCKS_DB */
525 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
526 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
527 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
528 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
529 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
530 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
531 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
532 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
533 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
534 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
536 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
538 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
540 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
542 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
543 #if defined(AFS_HPUX110_ENV)
545 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
546 #endif /* AFS_HPUX110_ENV */
547 #endif /* RX_ENABLE_LOCKS && KERNEL */
550 rx_connDeadTime = 12;
551 rx_tranquil = 0; /* reset flag */
552 rxi_ResetStatistics();
553 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
554 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
555 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
556 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
557 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
558 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
560 /* Malloc up a bunch of packets & buffers */
562 opr_queue_Init(&rx_freePacketQueue);
563 rxi_NeedMorePackets = FALSE;
564 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
566 /* enforce a minimum number of allocated packets */
567 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
568 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
570 /* allocate the initial free packet pool */
571 #ifdef RX_ENABLE_TSFPQ
572 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
573 #else /* RX_ENABLE_TSFPQ */
574 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
575 #endif /* RX_ENABLE_TSFPQ */
582 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
583 tv.tv_sec = clock_now.sec;
584 tv.tv_usec = clock_now.usec;
585 srand((unsigned int)tv.tv_usec);
592 #if defined(KERNEL) && !defined(UKERNEL)
593 /* Really, this should never happen in a real kernel */
596 struct sockaddr_in addr;
598 int addrlen = sizeof(addr);
600 socklen_t addrlen = sizeof(addr);
602 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
604 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
607 rx_port = addr.sin_port;
610 rx_stats.minRtt.sec = 9999999;
612 rx_SetEpoch(tv.tv_sec | 0x80000000);
614 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
615 * will provide a randomer value. */
617 MUTEX_ENTER(&rx_quota_mutex);
618 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
619 MUTEX_EXIT(&rx_quota_mutex);
620 /* *Slightly* random start time for the cid. This is just to help
621 * out with the hashing function at the peer */
622 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
623 rx_connHashTable = (struct rx_connection **)htable;
624 rx_peerHashTable = (struct rx_peer **)ptable;
626 rx_hardAckDelay.sec = 0;
627 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
629 rxevent_Init(20, rxi_ReScheduleEvents);
631 /* Initialize various global queues */
632 opr_queue_Init(&rx_idleServerQueue);
633 opr_queue_Init(&rx_incomingCallQueue);
634 opr_queue_Init(&rx_freeCallQueue);
636 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
637 /* Initialize our list of usable IP addresses. */
641 /* Start listener process (exact function is dependent on the
642 * implementation environment--kernel or user space) */
646 rx_atomic_clear_bit(&rxinit_status, 0);
653 return rx_InitHost(htonl(INADDR_ANY), port);
659 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
660 * maintaing the round trip timer.
665 * Start a new RTT timer for a given call and packet.
667 * There must be no resendEvent already listed for this call, otherwise this
668 * will leak events - intended for internal use within the RTO code only
671 * the RX call to start the timer for
672 * @param[in] lastPacket
673 * a flag indicating whether the last packet has been sent or not
675 * @pre call must be locked before calling this function
679 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
681 struct clock now, retryTime;
686 clock_Add(&retryTime, &call->rto);
688 /* If we're sending the last packet, and we're the client, then the server
689 * may wait for an additional 400ms before returning the ACK, wait for it
690 * rather than hitting a timeout */
691 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
692 clock_Addmsec(&retryTime, 400);
694 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
695 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
700 * Cancel an RTT timer for a given call.
704 * the RX call to cancel the timer for
706 * @pre call must be locked before calling this function
711 rxi_rto_cancel(struct rx_call *call)
713 rxevent_Cancel(&call->resendEvent);
714 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
718 * Tell the RTO timer that we have sent a packet.
720 * If the timer isn't already running, then start it. If the timer is running,
724 * the RX call that the packet has been sent on
725 * @param[in] lastPacket
726 * A flag which is true if this is the last packet for the call
728 * @pre The call must be locked before calling this function
733 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
735 if (call->resendEvent)
738 rxi_rto_startTimer(call, lastPacket, istack);
742 * Tell the RTO timer that we have received an new ACK message
744 * This function should be called whenever a call receives an ACK that
745 * acknowledges new packets. Whatever happens, we stop the current timer.
746 * If there are unacked packets in the queue which have been sent, then
747 * we restart the timer from now. Otherwise, we leave it stopped.
750 * the RX call that the ACK has been received on
754 rxi_rto_packet_acked(struct rx_call *call, int istack)
756 struct opr_queue *cursor;
758 rxi_rto_cancel(call);
760 if (opr_queue_IsEmpty(&call->tq))
763 for (opr_queue_Scan(&call->tq, cursor)) {
764 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
765 if (p->header.seq > call->tfirst + call->twind)
768 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
769 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
777 * Set an initial round trip timeout for a peer connection
779 * @param[in] secs The timeout to set in seconds
783 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
784 peer->rtt = secs * 8000;
788 * Enables or disables the busy call channel error (RX_CALL_BUSY).
790 * @param[in] onoff Non-zero to enable busy call channel errors.
792 * @pre Neither rx_Init nor rx_InitHost have been called yet
795 rx_SetBusyChannelError(afs_int32 onoff)
797 osi_Assert(rx_atomic_test_bit(&rxinit_status, 0));
798 rxi_busyChannelError = onoff ? 1 : 0;
802 * Set a delayed ack event on the specified call for the given time
804 * @param[in] call - the call on which to set the event
805 * @param[in] offset - the delay from now after which the event fires
808 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
810 struct clock now, when;
814 clock_Add(&when, offset);
816 if (call->delayedAckEvent && clock_Gt(&call->delayedAckTime, &when)) {
817 /* The event we're cancelling already has a reference, so we don't
819 rxevent_Cancel(&call->delayedAckEvent);
820 call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
823 call->delayedAckTime = when;
824 } else if (!call->delayedAckEvent) {
825 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
826 call->delayedAckEvent = rxevent_Post(&when, &now,
829 call->delayedAckTime = when;
834 rxi_CancelDelayedAckEvent(struct rx_call *call)
836 if (call->delayedAckEvent) {
837 rxevent_Cancel(&call->delayedAckEvent);
838 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
842 /* called with unincremented nRequestsRunning to see if it is OK to start
843 * a new thread in this service. Could be "no" for two reasons: over the
844 * max quota, or would prevent others from reaching their min quota.
846 #ifdef RX_ENABLE_LOCKS
847 /* This verion of QuotaOK reserves quota if it's ok while the
848 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
851 QuotaOK(struct rx_service *aservice)
853 /* check if over max quota */
854 if (aservice->nRequestsRunning >= aservice->maxProcs) {
858 /* under min quota, we're OK */
859 /* otherwise, can use only if there are enough to allow everyone
860 * to go to their min quota after this guy starts.
863 MUTEX_ENTER(&rx_quota_mutex);
864 if ((aservice->nRequestsRunning < aservice->minProcs)
865 || (rxi_availProcs > rxi_minDeficit)) {
866 aservice->nRequestsRunning++;
867 /* just started call in minProcs pool, need fewer to maintain
869 if (aservice->nRequestsRunning <= aservice->minProcs)
872 MUTEX_EXIT(&rx_quota_mutex);
875 MUTEX_EXIT(&rx_quota_mutex);
881 ReturnToServerPool(struct rx_service *aservice)
883 aservice->nRequestsRunning--;
884 MUTEX_ENTER(&rx_quota_mutex);
885 if (aservice->nRequestsRunning < aservice->minProcs)
888 MUTEX_EXIT(&rx_quota_mutex);
891 #else /* RX_ENABLE_LOCKS */
893 QuotaOK(struct rx_service *aservice)
896 /* under min quota, we're OK */
897 if (aservice->nRequestsRunning < aservice->minProcs)
900 /* check if over max quota */
901 if (aservice->nRequestsRunning >= aservice->maxProcs)
904 /* otherwise, can use only if there are enough to allow everyone
905 * to go to their min quota after this guy starts.
907 MUTEX_ENTER(&rx_quota_mutex);
908 if (rxi_availProcs > rxi_minDeficit)
910 MUTEX_EXIT(&rx_quota_mutex);
913 #endif /* RX_ENABLE_LOCKS */
916 /* Called by rx_StartServer to start up lwp's to service calls.
917 NExistingProcs gives the number of procs already existing, and which
918 therefore needn't be created. */
920 rxi_StartServerProcs(int nExistingProcs)
922 struct rx_service *service;
927 /* For each service, reserve N processes, where N is the "minimum"
928 * number of processes that MUST be able to execute a request in parallel,
929 * at any time, for that process. Also compute the maximum difference
930 * between any service's maximum number of processes that can run
931 * (i.e. the maximum number that ever will be run, and a guarantee
932 * that this number will run if other services aren't running), and its
933 * minimum number. The result is the extra number of processes that
934 * we need in order to provide the latter guarantee */
935 for (i = 0; i < RX_MAX_SERVICES; i++) {
937 service = rx_services[i];
938 if (service == (struct rx_service *)0)
940 nProcs += service->minProcs;
941 diff = service->maxProcs - service->minProcs;
945 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
946 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
947 for (i = 0; i < nProcs; i++) {
948 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
954 /* This routine is only required on Windows */
956 rx_StartClientThread(void)
958 #ifdef AFS_PTHREAD_ENV
960 pid = pthread_self();
961 #endif /* AFS_PTHREAD_ENV */
963 #endif /* AFS_NT40_ENV */
965 /* This routine must be called if any services are exported. If the
966 * donateMe flag is set, the calling process is donated to the server
969 rx_StartServer(int donateMe)
971 struct rx_service *service;
977 /* Start server processes, if necessary (exact function is dependent
978 * on the implementation environment--kernel or user space). DonateMe
979 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
980 * case, one less new proc will be created rx_StartServerProcs.
982 rxi_StartServerProcs(donateMe);
984 /* count up the # of threads in minProcs, and add set the min deficit to
985 * be that value, too.
987 for (i = 0; i < RX_MAX_SERVICES; i++) {
988 service = rx_services[i];
989 if (service == (struct rx_service *)0)
991 MUTEX_ENTER(&rx_quota_mutex);
992 rxi_totalMin += service->minProcs;
993 /* below works even if a thread is running, since minDeficit would
994 * still have been decremented and later re-incremented.
996 rxi_minDeficit += service->minProcs;
997 MUTEX_EXIT(&rx_quota_mutex);
1000 /* Turn on reaping of idle server connections */
1001 rxi_ReapConnections(NULL, NULL, NULL, 0);
1006 #ifndef AFS_NT40_ENV
1010 #ifdef AFS_PTHREAD_ENV
1012 pid = afs_pointer_to_int(pthread_self());
1013 #else /* AFS_PTHREAD_ENV */
1015 LWP_CurrentProcess(&pid);
1016 #endif /* AFS_PTHREAD_ENV */
1018 sprintf(name, "srv_%d", ++nProcs);
1019 if (registerProgram)
1020 (*registerProgram) (pid, name);
1022 #endif /* AFS_NT40_ENV */
1023 rx_ServerProc(NULL); /* Never returns */
1025 #ifdef RX_ENABLE_TSFPQ
1026 /* no use leaving packets around in this thread's local queue if
1027 * it isn't getting donated to the server thread pool.
1029 rxi_FlushLocalPacketsTSFPQ();
1030 #endif /* RX_ENABLE_TSFPQ */
1034 /* Create a new client connection to the specified service, using the
1035 * specified security object to implement the security model for this
1037 struct rx_connection *
1038 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1039 struct rx_securityClass *securityObject,
1040 int serviceSecurityIndex)
1044 struct rx_connection *conn;
1049 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1050 "serviceSecurityIndex %d)\n",
1051 ntohl(shost), ntohs(sport), sservice, securityObject,
1052 serviceSecurityIndex));
1054 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1055 * the case of kmem_alloc? */
1056 conn = rxi_AllocConnection();
1057 #ifdef RX_ENABLE_LOCKS
1058 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1059 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1060 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1063 MUTEX_ENTER(&rx_connHashTable_lock);
1064 cid = (rx_nextCid += RX_MAXCALLS);
1065 conn->type = RX_CLIENT_CONNECTION;
1067 conn->epoch = rx_epoch;
1068 conn->peer = rxi_FindPeer(shost, sport, 1);
1069 conn->serviceId = sservice;
1070 conn->securityObject = securityObject;
1071 conn->securityData = (void *) 0;
1072 conn->securityIndex = serviceSecurityIndex;
1073 rx_SetConnDeadTime(conn, rx_connDeadTime);
1074 rx_SetConnSecondsUntilNatPing(conn, 0);
1075 conn->ackRate = RX_FAST_ACK_RATE;
1076 conn->nSpecific = 0;
1077 conn->specific = NULL;
1078 conn->challengeEvent = NULL;
1079 conn->delayedAbortEvent = NULL;
1080 conn->abortCount = 0;
1082 for (i = 0; i < RX_MAXCALLS; i++) {
1083 conn->twind[i] = rx_initSendWindow;
1084 conn->rwind[i] = rx_initReceiveWindow;
1085 conn->lastBusy[i] = 0;
1088 RXS_NewConnection(securityObject, conn);
1090 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1092 conn->refCount++; /* no lock required since only this thread knows... */
1093 conn->next = rx_connHashTable[hashindex];
1094 rx_connHashTable[hashindex] = conn;
1095 if (rx_stats_active)
1096 rx_atomic_inc(&rx_stats.nClientConns);
1097 MUTEX_EXIT(&rx_connHashTable_lock);
1103 * Ensure a connection's timeout values are valid.
1105 * @param[in] conn The connection to check
1107 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1108 * unless idleDeadTime and/or hardDeadTime are not set
1112 rxi_CheckConnTimeouts(struct rx_connection *conn)
1114 /* a connection's timeouts must have the relationship
1115 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1116 * total loss of network to a peer may cause an idle timeout instead of a
1117 * dead timeout, simply because the idle timeout gets hit first. Also set
1118 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1119 /* this logic is slightly complicated by the fact that
1120 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1122 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1123 if (conn->idleDeadTime) {
1124 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1126 if (conn->hardDeadTime) {
1127 if (conn->idleDeadTime) {
1128 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1130 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1136 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1138 /* The idea is to set the dead time to a value that allows several
1139 * keepalives to be dropped without timing out the connection. */
1140 conn->secondsUntilDead = seconds;
1141 rxi_CheckConnTimeouts(conn);
1142 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1146 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1148 conn->hardDeadTime = seconds;
1149 rxi_CheckConnTimeouts(conn);
1153 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1155 conn->idleDeadTime = seconds;
1156 conn->idleDeadDetection = (seconds ? 1 : 0);
1157 rxi_CheckConnTimeouts(conn);
1160 int rxi_lowPeerRefCount = 0;
1161 int rxi_lowConnRefCount = 0;
1164 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1165 * NOTE: must not be called with rx_connHashTable_lock held.
1168 rxi_CleanupConnection(struct rx_connection *conn)
1170 /* Notify the service exporter, if requested, that this connection
1171 * is being destroyed */
1172 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1173 (*conn->service->destroyConnProc) (conn);
1175 /* Notify the security module that this connection is being destroyed */
1176 RXS_DestroyConnection(conn->securityObject, conn);
1178 /* If this is the last connection using the rx_peer struct, set its
1179 * idle time to now. rxi_ReapConnections will reap it if it's still
1180 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1182 MUTEX_ENTER(&rx_peerHashTable_lock);
1183 if (conn->peer->refCount < 2) {
1184 conn->peer->idleWhen = clock_Sec();
1185 if (conn->peer->refCount < 1) {
1186 conn->peer->refCount = 1;
1187 if (rx_stats_active) {
1188 MUTEX_ENTER(&rx_stats_mutex);
1189 rxi_lowPeerRefCount++;
1190 MUTEX_EXIT(&rx_stats_mutex);
1194 conn->peer->refCount--;
1195 MUTEX_EXIT(&rx_peerHashTable_lock);
1197 if (rx_stats_active)
1199 if (conn->type == RX_SERVER_CONNECTION)
1200 rx_atomic_dec(&rx_stats.nServerConns);
1202 rx_atomic_dec(&rx_stats.nClientConns);
1205 if (conn->specific) {
1207 for (i = 0; i < conn->nSpecific; i++) {
1208 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1209 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1210 conn->specific[i] = NULL;
1212 free(conn->specific);
1214 conn->specific = NULL;
1215 conn->nSpecific = 0;
1216 #endif /* !KERNEL */
1218 MUTEX_DESTROY(&conn->conn_call_lock);
1219 MUTEX_DESTROY(&conn->conn_data_lock);
1220 CV_DESTROY(&conn->conn_call_cv);
1222 rxi_FreeConnection(conn);
1225 /* Destroy the specified connection */
1227 rxi_DestroyConnection(struct rx_connection *conn)
1229 MUTEX_ENTER(&rx_connHashTable_lock);
1230 rxi_DestroyConnectionNoLock(conn);
1231 /* conn should be at the head of the cleanup list */
1232 if (conn == rx_connCleanup_list) {
1233 rx_connCleanup_list = rx_connCleanup_list->next;
1234 MUTEX_EXIT(&rx_connHashTable_lock);
1235 rxi_CleanupConnection(conn);
1237 #ifdef RX_ENABLE_LOCKS
1239 MUTEX_EXIT(&rx_connHashTable_lock);
1241 #endif /* RX_ENABLE_LOCKS */
1245 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1247 struct rx_connection **conn_ptr;
1249 struct rx_packet *packet;
1256 MUTEX_ENTER(&conn->conn_data_lock);
1257 MUTEX_ENTER(&rx_refcnt_mutex);
1258 if (conn->refCount > 0)
1261 if (rx_stats_active) {
1262 MUTEX_ENTER(&rx_stats_mutex);
1263 rxi_lowConnRefCount++;
1264 MUTEX_EXIT(&rx_stats_mutex);
1268 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1269 /* Busy; wait till the last guy before proceeding */
1270 MUTEX_EXIT(&rx_refcnt_mutex);
1271 MUTEX_EXIT(&conn->conn_data_lock);
1276 /* If the client previously called rx_NewCall, but it is still
1277 * waiting, treat this as a running call, and wait to destroy the
1278 * connection later when the call completes. */
1279 if ((conn->type == RX_CLIENT_CONNECTION)
1280 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1281 conn->flags |= RX_CONN_DESTROY_ME;
1282 MUTEX_EXIT(&conn->conn_data_lock);
1286 MUTEX_EXIT(&rx_refcnt_mutex);
1287 MUTEX_EXIT(&conn->conn_data_lock);
1289 /* Check for extant references to this connection */
1290 MUTEX_ENTER(&conn->conn_call_lock);
1291 for (i = 0; i < RX_MAXCALLS; i++) {
1292 struct rx_call *call = conn->call[i];
1295 if (conn->type == RX_CLIENT_CONNECTION) {
1296 MUTEX_ENTER(&call->lock);
1297 if (call->delayedAckEvent) {
1298 /* Push the final acknowledgment out now--there
1299 * won't be a subsequent call to acknowledge the
1300 * last reply packets */
1301 rxi_CancelDelayedAckEvent(call);
1302 if (call->state == RX_STATE_PRECALL
1303 || call->state == RX_STATE_ACTIVE) {
1304 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1309 MUTEX_EXIT(&call->lock);
1313 MUTEX_EXIT(&conn->conn_call_lock);
1315 #ifdef RX_ENABLE_LOCKS
1317 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1318 MUTEX_EXIT(&conn->conn_data_lock);
1320 /* Someone is accessing a packet right now. */
1324 #endif /* RX_ENABLE_LOCKS */
1327 /* Don't destroy the connection if there are any call
1328 * structures still in use */
1329 MUTEX_ENTER(&conn->conn_data_lock);
1330 conn->flags |= RX_CONN_DESTROY_ME;
1331 MUTEX_EXIT(&conn->conn_data_lock);
1336 if (conn->natKeepAliveEvent) {
1337 rxi_NatKeepAliveOff(conn);
1340 if (conn->delayedAbortEvent) {
1341 rxevent_Cancel(&conn->delayedAbortEvent);
1342 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1344 MUTEX_ENTER(&conn->conn_data_lock);
1345 rxi_SendConnectionAbort(conn, packet, 0, 1);
1346 MUTEX_EXIT(&conn->conn_data_lock);
1347 rxi_FreePacket(packet);
1351 /* Remove from connection hash table before proceeding */
1353 &rx_connHashTable[CONN_HASH
1354 (peer->host, peer->port, conn->cid, conn->epoch,
1356 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1357 if (*conn_ptr == conn) {
1358 *conn_ptr = conn->next;
1362 /* if the conn that we are destroying was the last connection, then we
1363 * clear rxLastConn as well */
1364 if (rxLastConn == conn)
1367 /* Make sure the connection is completely reset before deleting it. */
1368 /* get rid of pending events that could zap us later */
1369 rxevent_Cancel(&conn->challengeEvent);
1370 rxevent_Cancel(&conn->checkReachEvent);
1371 rxevent_Cancel(&conn->natKeepAliveEvent);
1373 /* Add the connection to the list of destroyed connections that
1374 * need to be cleaned up. This is necessary to avoid deadlocks
1375 * in the routines we call to inform others that this connection is
1376 * being destroyed. */
1377 conn->next = rx_connCleanup_list;
1378 rx_connCleanup_list = conn;
1381 /* Externally available version */
1383 rx_DestroyConnection(struct rx_connection *conn)
1388 rxi_DestroyConnection(conn);
1393 rx_GetConnection(struct rx_connection *conn)
1398 MUTEX_ENTER(&rx_refcnt_mutex);
1400 MUTEX_EXIT(&rx_refcnt_mutex);
1404 #ifdef RX_ENABLE_LOCKS
1405 /* Wait for the transmit queue to no longer be busy.
1406 * requires the call->lock to be held */
1408 rxi_WaitforTQBusy(struct rx_call *call) {
1409 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1410 call->flags |= RX_CALL_TQ_WAIT;
1412 MUTEX_ASSERT(&call->lock);
1413 CV_WAIT(&call->cv_tq, &call->lock);
1415 if (call->tqWaiters == 0) {
1416 call->flags &= ~RX_CALL_TQ_WAIT;
1423 rxi_WakeUpTransmitQueue(struct rx_call *call)
1425 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1426 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1427 call, call->tqWaiters, call->flags));
1428 #ifdef RX_ENABLE_LOCKS
1429 MUTEX_ASSERT(&call->lock);
1430 CV_BROADCAST(&call->cv_tq);
1431 #else /* RX_ENABLE_LOCKS */
1432 osi_rxWakeup(&call->tq);
1433 #endif /* RX_ENABLE_LOCKS */
1437 /* Start a new rx remote procedure call, on the specified connection.
1438 * If wait is set to 1, wait for a free call channel; otherwise return
1439 * 0. Maxtime gives the maximum number of seconds this call may take,
1440 * after rx_NewCall returns. After this time interval, a call to any
1441 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1442 * For fine grain locking, we hold the conn_call_lock in order to
1443 * to ensure that we don't get signalle after we found a call in an active
1444 * state and before we go to sleep.
1447 rx_NewCall(struct rx_connection *conn)
1449 int i, wait, ignoreBusy = 1;
1450 struct rx_call *call;
1451 struct clock queueTime;
1452 afs_uint32 leastBusy = 0;
1456 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1459 clock_GetTime(&queueTime);
1461 * Check if there are others waiting for a new call.
1462 * If so, let them go first to avoid starving them.
1463 * This is a fairly simple scheme, and might not be
1464 * a complete solution for large numbers of waiters.
1466 * makeCallWaiters keeps track of the number of
1467 * threads waiting to make calls and the
1468 * RX_CONN_MAKECALL_WAITING flag bit is used to
1469 * indicate that there are indeed calls waiting.
1470 * The flag is set when the waiter is incremented.
1471 * It is only cleared when makeCallWaiters is 0.
1472 * This prevents us from accidently destroying the
1473 * connection while it is potentially about to be used.
1475 MUTEX_ENTER(&conn->conn_call_lock);
1476 MUTEX_ENTER(&conn->conn_data_lock);
1477 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1478 conn->flags |= RX_CONN_MAKECALL_WAITING;
1479 conn->makeCallWaiters++;
1480 MUTEX_EXIT(&conn->conn_data_lock);
1482 #ifdef RX_ENABLE_LOCKS
1483 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1487 MUTEX_ENTER(&conn->conn_data_lock);
1488 conn->makeCallWaiters--;
1489 if (conn->makeCallWaiters == 0)
1490 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1493 /* We are now the active thread in rx_NewCall */
1494 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1495 MUTEX_EXIT(&conn->conn_data_lock);
1500 for (i = 0; i < RX_MAXCALLS; i++) {
1501 call = conn->call[i];
1503 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1504 /* we're not ignoring busy call slots; only look at the
1505 * call slot that is the "least" busy */
1509 if (call->state == RX_STATE_DALLY) {
1510 MUTEX_ENTER(&call->lock);
1511 if (call->state == RX_STATE_DALLY) {
1512 if (ignoreBusy && conn->lastBusy[i]) {
1513 /* if we're ignoring busy call slots, skip any ones that
1514 * have lastBusy set */
1515 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1516 leastBusy = conn->lastBusy[i];
1518 MUTEX_EXIT(&call->lock);
1523 * We are setting the state to RX_STATE_RESET to
1524 * ensure that no one else will attempt to use this
1525 * call once we drop the conn->conn_call_lock and
1526 * call->lock. We must drop the conn->conn_call_lock
1527 * before calling rxi_ResetCall because the process
1528 * of clearing the transmit queue can block for an
1529 * extended period of time. If we block while holding
1530 * the conn->conn_call_lock, then all rx_EndCall
1531 * processing will block as well. This has a detrimental
1532 * effect on overall system performance.
1534 call->state = RX_STATE_RESET;
1535 (*call->callNumber)++;
1536 MUTEX_EXIT(&conn->conn_call_lock);
1537 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1538 rxi_ResetCall(call, 0);
1539 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1543 * If we failed to be able to safely obtain the
1544 * conn->conn_call_lock we will have to drop the
1545 * call->lock to avoid a deadlock. When the call->lock
1546 * is released the state of the call can change. If it
1547 * is no longer RX_STATE_RESET then some other thread is
1550 MUTEX_EXIT(&call->lock);
1551 MUTEX_ENTER(&conn->conn_call_lock);
1552 MUTEX_ENTER(&call->lock);
1554 if (call->state == RX_STATE_RESET)
1558 * If we get here it means that after dropping
1559 * the conn->conn_call_lock and call->lock that
1560 * the call is no longer ours. If we can't find
1561 * a free call in the remaining slots we should
1562 * not go immediately to RX_CONN_MAKECALL_WAITING
1563 * because by dropping the conn->conn_call_lock
1564 * we have given up synchronization with rx_EndCall.
1565 * Instead, cycle through one more time to see if
1566 * we can find a call that can call our own.
1568 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1571 MUTEX_EXIT(&call->lock);
1574 if (ignoreBusy && conn->lastBusy[i]) {
1575 /* if we're ignoring busy call slots, skip any ones that
1576 * have lastBusy set */
1577 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1578 leastBusy = conn->lastBusy[i];
1583 /* rxi_NewCall returns with mutex locked */
1584 call = rxi_NewCall(conn, i);
1585 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1589 if (i < RX_MAXCALLS) {
1590 conn->lastBusy[i] = 0;
1591 call->flags &= ~RX_CALL_PEER_BUSY;
1596 if (leastBusy && ignoreBusy) {
1597 /* we didn't find a useable call slot, but we did see at least one
1598 * 'busy' slot; look again and only use a slot with the 'least
1604 MUTEX_ENTER(&conn->conn_data_lock);
1605 conn->flags |= RX_CONN_MAKECALL_WAITING;
1606 conn->makeCallWaiters++;
1607 MUTEX_EXIT(&conn->conn_data_lock);
1609 #ifdef RX_ENABLE_LOCKS
1610 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1614 MUTEX_ENTER(&conn->conn_data_lock);
1615 conn->makeCallWaiters--;
1616 if (conn->makeCallWaiters == 0)
1617 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1618 MUTEX_EXIT(&conn->conn_data_lock);
1620 /* Client is initially in send mode */
1621 call->state = RX_STATE_ACTIVE;
1622 call->error = conn->error;
1624 call->app.mode = RX_MODE_ERROR;
1626 call->app.mode = RX_MODE_SENDING;
1628 #ifdef AFS_RXERRQ_ENV
1629 /* remember how many network errors the peer has when we started, so if
1630 * more errors are encountered after the call starts, we know the other endpoint won't be
1631 * responding to us */
1632 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1635 /* remember start time for call in case we have hard dead time limit */
1636 call->queueTime = queueTime;
1637 clock_GetTime(&call->startTime);
1638 call->app.bytesSent = 0;
1639 call->app.bytesRcvd = 0;
1641 /* Turn on busy protocol. */
1642 rxi_KeepAliveOn(call);
1644 /* Attempt MTU discovery */
1645 rxi_GrowMTUOn(call);
1648 * We are no longer the active thread in rx_NewCall
1650 MUTEX_ENTER(&conn->conn_data_lock);
1651 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1652 MUTEX_EXIT(&conn->conn_data_lock);
1655 * Wake up anyone else who might be giving us a chance to
1656 * run (see code above that avoids resource starvation).
1658 #ifdef RX_ENABLE_LOCKS
1659 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1660 osi_Panic("rx_NewCall call about to be used without an empty tq");
1663 CV_BROADCAST(&conn->conn_call_cv);
1667 MUTEX_EXIT(&conn->conn_call_lock);
1668 MUTEX_EXIT(&call->lock);
1671 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1676 rxi_HasActiveCalls(struct rx_connection *aconn)
1679 struct rx_call *tcall;
1683 for (i = 0; i < RX_MAXCALLS; i++) {
1684 if ((tcall = aconn->call[i])) {
1685 if ((tcall->state == RX_STATE_ACTIVE)
1686 || (tcall->state == RX_STATE_PRECALL)) {
1697 rxi_GetCallNumberVector(struct rx_connection *aconn,
1698 afs_int32 * aint32s)
1701 struct rx_call *tcall;
1705 MUTEX_ENTER(&aconn->conn_call_lock);
1706 for (i = 0; i < RX_MAXCALLS; i++) {
1707 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1708 aint32s[i] = aconn->callNumber[i] + 1;
1710 aint32s[i] = aconn->callNumber[i];
1712 MUTEX_EXIT(&aconn->conn_call_lock);
1718 rxi_SetCallNumberVector(struct rx_connection *aconn,
1719 afs_int32 * aint32s)
1722 struct rx_call *tcall;
1726 MUTEX_ENTER(&aconn->conn_call_lock);
1727 for (i = 0; i < RX_MAXCALLS; i++) {
1728 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1729 aconn->callNumber[i] = aint32s[i] - 1;
1731 aconn->callNumber[i] = aint32s[i];
1733 MUTEX_EXIT(&aconn->conn_call_lock);
1738 /* Advertise a new service. A service is named locally by a UDP port
1739 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1742 char *serviceName; Name for identification purposes (e.g. the
1743 service name might be used for probing for
1746 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1747 char *serviceName, struct rx_securityClass **securityObjects,
1748 int nSecurityObjects,
1749 afs_int32(*serviceProc) (struct rx_call * acall))
1751 osi_socket socket = OSI_NULLSOCKET;
1752 struct rx_service *tservice;
1758 if (serviceId == 0) {
1760 "rx_NewService: service id for service %s is not non-zero.\n",
1767 "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",
1775 tservice = rxi_AllocService();
1778 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1780 for (i = 0; i < RX_MAX_SERVICES; i++) {
1781 struct rx_service *service = rx_services[i];
1783 if (port == service->servicePort && host == service->serviceHost) {
1784 if (service->serviceId == serviceId) {
1785 /* The identical service has already been
1786 * installed; if the caller was intending to
1787 * change the security classes used by this
1788 * service, he/she loses. */
1790 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1791 serviceName, serviceId, service->serviceName);
1793 rxi_FreeService(tservice);
1796 /* Different service, same port: re-use the socket
1797 * which is bound to the same port */
1798 socket = service->socket;
1801 if (socket == OSI_NULLSOCKET) {
1802 /* If we don't already have a socket (from another
1803 * service on same port) get a new one */
1804 socket = rxi_GetHostUDPSocket(host, port);
1805 if (socket == OSI_NULLSOCKET) {
1807 rxi_FreeService(tservice);
1812 service->socket = socket;
1813 service->serviceHost = host;
1814 service->servicePort = port;
1815 service->serviceId = serviceId;
1816 service->serviceName = serviceName;
1817 service->nSecurityObjects = nSecurityObjects;
1818 service->securityObjects = securityObjects;
1819 service->minProcs = 0;
1820 service->maxProcs = 1;
1821 service->idleDeadTime = 60;
1822 service->idleDeadErr = 0;
1823 service->connDeadTime = rx_connDeadTime;
1824 service->executeRequestProc = serviceProc;
1825 service->checkReach = 0;
1826 service->nSpecific = 0;
1827 service->specific = NULL;
1828 rx_services[i] = service; /* not visible until now */
1834 rxi_FreeService(tservice);
1835 (osi_Msg "rx_NewService: cannot support > %d services\n",
1840 /* Set configuration options for all of a service's security objects */
1843 rx_SetSecurityConfiguration(struct rx_service *service,
1844 rx_securityConfigVariables type,
1848 for (i = 0; i<service->nSecurityObjects; i++) {
1849 if (service->securityObjects[i]) {
1850 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1858 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1859 struct rx_securityClass **securityObjects, int nSecurityObjects,
1860 afs_int32(*serviceProc) (struct rx_call * acall))
1862 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1865 /* Generic request processing loop. This routine should be called
1866 * by the implementation dependent rx_ServerProc. If socketp is
1867 * non-null, it will be set to the file descriptor that this thread
1868 * is now listening on. If socketp is null, this routine will never
1871 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1873 struct rx_call *call;
1875 struct rx_service *tservice = NULL;
1882 call = rx_GetCall(threadID, tservice, socketp);
1883 if (socketp && *socketp != OSI_NULLSOCKET) {
1884 /* We are now a listener thread */
1890 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1891 #ifdef RX_ENABLE_LOCKS
1893 #endif /* RX_ENABLE_LOCKS */
1894 afs_termState = AFSOP_STOP_AFS;
1895 afs_osi_Wakeup(&afs_termState);
1896 #ifdef RX_ENABLE_LOCKS
1898 #endif /* RX_ENABLE_LOCKS */
1903 /* if server is restarting( typically smooth shutdown) then do not
1904 * allow any new calls.
1907 if (rx_tranquil && (call != NULL)) {
1911 MUTEX_ENTER(&call->lock);
1913 rxi_CallError(call, RX_RESTARTING);
1914 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1916 MUTEX_EXIT(&call->lock);
1921 tservice = call->conn->service;
1923 if (tservice->beforeProc)
1924 (*tservice->beforeProc) (call);
1926 code = tservice->executeRequestProc(call);
1928 if (tservice->afterProc)
1929 (*tservice->afterProc) (call, code);
1931 rx_EndCall(call, code);
1933 if (tservice->postProc)
1934 (*tservice->postProc) (code);
1936 if (rx_stats_active) {
1937 MUTEX_ENTER(&rx_stats_mutex);
1939 MUTEX_EXIT(&rx_stats_mutex);
1946 rx_WakeupServerProcs(void)
1948 struct rx_serverQueueEntry *np, *tqp;
1949 struct opr_queue *cursor;
1953 MUTEX_ENTER(&rx_serverPool_lock);
1955 #ifdef RX_ENABLE_LOCKS
1956 if (rx_waitForPacket)
1957 CV_BROADCAST(&rx_waitForPacket->cv);
1958 #else /* RX_ENABLE_LOCKS */
1959 if (rx_waitForPacket)
1960 osi_rxWakeup(rx_waitForPacket);
1961 #endif /* RX_ENABLE_LOCKS */
1962 MUTEX_ENTER(&freeSQEList_lock);
1963 for (np = rx_FreeSQEList; np; np = tqp) {
1964 tqp = *(struct rx_serverQueueEntry **)np;
1965 #ifdef RX_ENABLE_LOCKS
1966 CV_BROADCAST(&np->cv);
1967 #else /* RX_ENABLE_LOCKS */
1969 #endif /* RX_ENABLE_LOCKS */
1971 MUTEX_EXIT(&freeSQEList_lock);
1972 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1973 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1974 #ifdef RX_ENABLE_LOCKS
1975 CV_BROADCAST(&np->cv);
1976 #else /* RX_ENABLE_LOCKS */
1978 #endif /* RX_ENABLE_LOCKS */
1980 MUTEX_EXIT(&rx_serverPool_lock);
1985 * One thing that seems to happen is that all the server threads get
1986 * tied up on some empty or slow call, and then a whole bunch of calls
1987 * arrive at once, using up the packet pool, so now there are more
1988 * empty calls. The most critical resources here are server threads
1989 * and the free packet pool. The "doreclaim" code seems to help in
1990 * general. I think that eventually we arrive in this state: there
1991 * are lots of pending calls which do have all their packets present,
1992 * so they won't be reclaimed, are multi-packet calls, so they won't
1993 * be scheduled until later, and thus are tying up most of the free
1994 * packet pool for a very long time.
1996 * 1. schedule multi-packet calls if all the packets are present.
1997 * Probably CPU-bound operation, useful to return packets to pool.
1998 * Do what if there is a full window, but the last packet isn't here?
1999 * 3. preserve one thread which *only* runs "best" calls, otherwise
2000 * it sleeps and waits for that type of call.
2001 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2002 * the current dataquota business is badly broken. The quota isn't adjusted
2003 * to reflect how many packets are presently queued for a running call.
2004 * So, when we schedule a queued call with a full window of packets queued
2005 * up for it, that *should* free up a window full of packets for other 2d-class
2006 * calls to be able to use from the packet pool. But it doesn't.
2008 * NB. Most of the time, this code doesn't run -- since idle server threads
2009 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2010 * as a new call arrives.
2012 /* Sleep until a call arrives. Returns a pointer to the call, ready
2013 * for an rx_Read. */
2014 #ifdef RX_ENABLE_LOCKS
2016 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2018 struct rx_serverQueueEntry *sq;
2019 struct rx_call *call = (struct rx_call *)0;
2020 struct rx_service *service = NULL;
2022 MUTEX_ENTER(&freeSQEList_lock);
2024 if ((sq = rx_FreeSQEList)) {
2025 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2026 MUTEX_EXIT(&freeSQEList_lock);
2027 } else { /* otherwise allocate a new one and return that */
2028 MUTEX_EXIT(&freeSQEList_lock);
2029 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2030 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2031 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2034 MUTEX_ENTER(&rx_serverPool_lock);
2035 if (cur_service != NULL) {
2036 ReturnToServerPool(cur_service);
2039 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2040 struct rx_call *tcall, *choice2 = NULL;
2041 struct opr_queue *cursor;
2043 /* Scan for eligible incoming calls. A call is not eligible
2044 * if the maximum number of calls for its service type are
2045 * already executing */
2046 /* One thread will process calls FCFS (to prevent starvation),
2047 * while the other threads may run ahead looking for calls which
2048 * have all their input data available immediately. This helps
2049 * keep threads from blocking, waiting for data from the client. */
2050 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2051 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2053 service = tcall->conn->service;
2054 if (!QuotaOK(service)) {
2057 MUTEX_ENTER(&rx_pthread_mutex);
2058 if (tno == rxi_fcfs_thread_num
2059 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2060 MUTEX_EXIT(&rx_pthread_mutex);
2061 /* If we're the fcfs thread , then we'll just use
2062 * this call. If we haven't been able to find an optimal
2063 * choice, and we're at the end of the list, then use a
2064 * 2d choice if one has been identified. Otherwise... */
2065 call = (choice2 ? choice2 : tcall);
2066 service = call->conn->service;
2068 MUTEX_EXIT(&rx_pthread_mutex);
2069 if (!opr_queue_IsEmpty(&tcall->rq)) {
2070 struct rx_packet *rp;
2071 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2073 if (rp->header.seq == 1) {
2075 || (rp->header.flags & RX_LAST_PACKET)) {
2077 } else if (rxi_2dchoice && !choice2
2078 && !(tcall->flags & RX_CALL_CLEARED)
2079 && (tcall->rprev > rxi_HardAckRate)) {
2089 ReturnToServerPool(service);
2095 opr_queue_Remove(&call->entry);
2096 MUTEX_EXIT(&rx_serverPool_lock);
2097 MUTEX_ENTER(&call->lock);
2099 if (call->flags & RX_CALL_WAIT_PROC) {
2100 call->flags &= ~RX_CALL_WAIT_PROC;
2101 rx_atomic_dec(&rx_nWaiting);
2104 if (call->state != RX_STATE_PRECALL || call->error) {
2105 MUTEX_EXIT(&call->lock);
2106 MUTEX_ENTER(&rx_serverPool_lock);
2107 ReturnToServerPool(service);
2112 if (opr_queue_IsEmpty(&call->rq)
2113 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2114 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2116 CLEAR_CALL_QUEUE_LOCK(call);
2119 /* If there are no eligible incoming calls, add this process
2120 * to the idle server queue, to wait for one */
2124 *socketp = OSI_NULLSOCKET;
2126 sq->socketp = socketp;
2127 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2128 #ifndef AFS_AIX41_ENV
2129 rx_waitForPacket = sq;
2130 #endif /* AFS_AIX41_ENV */
2132 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2134 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2135 MUTEX_EXIT(&rx_serverPool_lock);
2136 return (struct rx_call *)0;
2139 } while (!(call = sq->newcall)
2140 && !(socketp && *socketp != OSI_NULLSOCKET));
2141 MUTEX_EXIT(&rx_serverPool_lock);
2143 MUTEX_ENTER(&call->lock);
2149 MUTEX_ENTER(&freeSQEList_lock);
2150 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2151 rx_FreeSQEList = sq;
2152 MUTEX_EXIT(&freeSQEList_lock);
2155 clock_GetTime(&call->startTime);
2156 call->state = RX_STATE_ACTIVE;
2157 call->app.mode = RX_MODE_RECEIVING;
2158 #ifdef RX_KERNEL_TRACE
2159 if (ICL_SETACTIVE(afs_iclSetp)) {
2160 int glockOwner = ISAFS_GLOCK();
2163 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2164 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2171 rxi_calltrace(RX_CALL_START, call);
2172 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2173 call->conn->service->servicePort, call->conn->service->serviceId,
2176 MUTEX_EXIT(&call->lock);
2177 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2179 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2184 #else /* RX_ENABLE_LOCKS */
2186 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2188 struct rx_serverQueueEntry *sq;
2189 struct rx_call *call = (struct rx_call *)0, *choice2;
2190 struct rx_service *service = NULL;
2194 MUTEX_ENTER(&freeSQEList_lock);
2196 if ((sq = rx_FreeSQEList)) {
2197 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2198 MUTEX_EXIT(&freeSQEList_lock);
2199 } else { /* otherwise allocate a new one and return that */
2200 MUTEX_EXIT(&freeSQEList_lock);
2201 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2202 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2203 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2205 MUTEX_ENTER(&sq->lock);
2207 if (cur_service != NULL) {
2208 cur_service->nRequestsRunning--;
2209 MUTEX_ENTER(&rx_quota_mutex);
2210 if (cur_service->nRequestsRunning < cur_service->minProcs)
2213 MUTEX_EXIT(&rx_quota_mutex);
2215 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2216 struct rx_call *tcall;
2217 struct opr_queue *cursor;
2218 /* Scan for eligible incoming calls. A call is not eligible
2219 * if the maximum number of calls for its service type are
2220 * already executing */
2221 /* One thread will process calls FCFS (to prevent starvation),
2222 * while the other threads may run ahead looking for calls which
2223 * have all their input data available immediately. This helps
2224 * keep threads from blocking, waiting for data from the client. */
2225 choice2 = (struct rx_call *)0;
2226 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2227 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2228 service = tcall->conn->service;
2229 if (QuotaOK(service)) {
2230 MUTEX_ENTER(&rx_pthread_mutex);
2231 /* XXX - If tcall->entry.next is NULL, then we're no longer
2232 * on a queue at all. This shouldn't happen. */
2233 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2234 MUTEX_EXIT(&rx_pthread_mutex);
2235 /* If we're the fcfs thread, then we'll just use
2236 * this call. If we haven't been able to find an optimal
2237 * choice, and we're at the end of the list, then use a
2238 * 2d choice if one has been identified. Otherwise... */
2239 call = (choice2 ? choice2 : tcall);
2240 service = call->conn->service;
2242 MUTEX_EXIT(&rx_pthread_mutex);
2243 if (!opr_queue_IsEmpty(&tcall->rq)) {
2244 struct rx_packet *rp;
2245 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2247 if (rp->header.seq == 1
2249 || (rp->header.flags & RX_LAST_PACKET))) {
2251 } else if (rxi_2dchoice && !choice2
2252 && !(tcall->flags & RX_CALL_CLEARED)
2253 && (tcall->rprev > rxi_HardAckRate)) {
2266 opr_queue_Remove(&call->entry);
2267 /* we can't schedule a call if there's no data!!! */
2268 /* send an ack if there's no data, if we're missing the
2269 * first packet, or we're missing something between first
2270 * and last -- there's a "hole" in the incoming data. */
2271 if (opr_queue_IsEmpty(&call->rq)
2272 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2273 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2274 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2276 call->flags &= (~RX_CALL_WAIT_PROC);
2277 service->nRequestsRunning++;
2278 /* just started call in minProcs pool, need fewer to maintain
2280 MUTEX_ENTER(&rx_quota_mutex);
2281 if (service->nRequestsRunning <= service->minProcs)
2284 MUTEX_EXIT(&rx_quota_mutex);
2285 rx_atomic_dec(&rx_nWaiting);
2286 /* MUTEX_EXIT(&call->lock); */
2288 /* If there are no eligible incoming calls, add this process
2289 * to the idle server queue, to wait for one */
2292 *socketp = OSI_NULLSOCKET;
2294 sq->socketp = socketp;
2295 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2299 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2301 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2302 return (struct rx_call *)0;
2305 } while (!(call = sq->newcall)
2306 && !(socketp && *socketp != OSI_NULLSOCKET));
2308 MUTEX_EXIT(&sq->lock);
2310 MUTEX_ENTER(&freeSQEList_lock);
2311 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2312 rx_FreeSQEList = sq;
2313 MUTEX_EXIT(&freeSQEList_lock);
2316 clock_GetTime(&call->startTime);
2317 call->state = RX_STATE_ACTIVE;
2318 call->app.mode = RX_MODE_RECEIVING;
2319 #ifdef RX_KERNEL_TRACE
2320 if (ICL_SETACTIVE(afs_iclSetp)) {
2321 int glockOwner = ISAFS_GLOCK();
2324 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2325 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2332 rxi_calltrace(RX_CALL_START, call);
2333 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2334 call->conn->service->servicePort, call->conn->service->serviceId,
2337 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2344 #endif /* RX_ENABLE_LOCKS */
2348 /* Establish a procedure to be called when a packet arrives for a
2349 * call. This routine will be called at most once after each call,
2350 * and will also be called if there is an error condition on the or
2351 * the call is complete. Used by multi rx to build a selection
2352 * function which determines which of several calls is likely to be a
2353 * good one to read from.
2354 * NOTE: the way this is currently implemented it is probably only a
2355 * good idea to (1) use it immediately after a newcall (clients only)
2356 * and (2) only use it once. Other uses currently void your warranty
2359 rx_SetArrivalProc(struct rx_call *call,
2360 void (*proc) (struct rx_call * call,
2363 void * handle, int arg)
2365 call->arrivalProc = proc;
2366 call->arrivalProcHandle = handle;
2367 call->arrivalProcArg = arg;
2370 /* Call is finished (possibly prematurely). Return rc to the peer, if
2371 * appropriate, and return the final error code from the conversation
2375 rx_EndCall(struct rx_call *call, afs_int32 rc)
2377 struct rx_connection *conn = call->conn;
2381 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2382 call, rc, call->error, call->abortCode));
2385 MUTEX_ENTER(&call->lock);
2387 if (rc == 0 && call->error == 0) {
2388 call->abortCode = 0;
2389 call->abortCount = 0;
2392 call->arrivalProc = (void (*)())0;
2393 if (rc && call->error == 0) {
2394 rxi_CallError(call, rc);
2395 call->app.mode = RX_MODE_ERROR;
2396 /* Send an abort message to the peer if this error code has
2397 * only just been set. If it was set previously, assume the
2398 * peer has already been sent the error code or will request it
2400 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2402 if (conn->type == RX_SERVER_CONNECTION) {
2403 /* Make sure reply or at least dummy reply is sent */
2404 if (call->app.mode == RX_MODE_RECEIVING) {
2405 MUTEX_EXIT(&call->lock);
2406 rxi_WriteProc(call, 0, 0);
2407 MUTEX_ENTER(&call->lock);
2409 if (call->app.mode == RX_MODE_SENDING) {
2410 MUTEX_EXIT(&call->lock);
2411 rxi_FlushWrite(call);
2412 MUTEX_ENTER(&call->lock);
2414 rxi_calltrace(RX_CALL_END, call);
2415 /* Call goes to hold state until reply packets are acknowledged */
2416 if (call->tfirst + call->nSoftAcked < call->tnext) {
2417 call->state = RX_STATE_HOLD;
2419 call->state = RX_STATE_DALLY;
2420 rxi_ClearTransmitQueue(call, 0);
2421 rxi_rto_cancel(call);
2422 rxi_CancelKeepAliveEvent(call);
2424 } else { /* Client connection */
2426 /* Make sure server receives input packets, in the case where
2427 * no reply arguments are expected */
2429 if ((call->app.mode == RX_MODE_SENDING)
2430 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2431 MUTEX_EXIT(&call->lock);
2432 (void)rxi_ReadProc(call, &dummy, 1);
2433 MUTEX_ENTER(&call->lock);
2436 /* If we had an outstanding delayed ack, be nice to the server
2437 * and force-send it now.
2439 if (call->delayedAckEvent) {
2440 rxi_CancelDelayedAckEvent(call);
2441 rxi_SendDelayedAck(NULL, call, NULL, 0);
2444 /* We need to release the call lock since it's lower than the
2445 * conn_call_lock and we don't want to hold the conn_call_lock
2446 * over the rx_ReadProc call. The conn_call_lock needs to be held
2447 * here for the case where rx_NewCall is perusing the calls on
2448 * the connection structure. We don't want to signal until
2449 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2450 * have checked this call, found it active and by the time it
2451 * goes to sleep, will have missed the signal.
2453 MUTEX_EXIT(&call->lock);
2454 MUTEX_ENTER(&conn->conn_call_lock);
2455 MUTEX_ENTER(&call->lock);
2457 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2458 conn->lastBusy[call->channel] = 0;
2461 MUTEX_ENTER(&conn->conn_data_lock);
2462 conn->flags |= RX_CONN_BUSY;
2463 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2464 MUTEX_EXIT(&conn->conn_data_lock);
2465 #ifdef RX_ENABLE_LOCKS
2466 CV_BROADCAST(&conn->conn_call_cv);
2471 #ifdef RX_ENABLE_LOCKS
2473 MUTEX_EXIT(&conn->conn_data_lock);
2475 #endif /* RX_ENABLE_LOCKS */
2476 call->state = RX_STATE_DALLY;
2478 error = call->error;
2480 /* currentPacket, nLeft, and NFree must be zeroed here, because
2481 * ResetCall cannot: ResetCall may be called at splnet(), in the
2482 * kernel version, and may interrupt the macros rx_Read or
2483 * rx_Write, which run at normal priority for efficiency. */
2484 if (call->app.currentPacket) {
2485 #ifdef RX_TRACK_PACKETS
2486 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2488 rxi_FreePacket(call->app.currentPacket);
2489 call->app.currentPacket = (struct rx_packet *)0;
2492 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2494 /* Free any packets from the last call to ReadvProc/WritevProc */
2495 #ifdef RXDEBUG_PACKET
2497 #endif /* RXDEBUG_PACKET */
2498 rxi_FreePackets(0, &call->app.iovq);
2499 MUTEX_EXIT(&call->lock);
2501 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2502 if (conn->type == RX_CLIENT_CONNECTION) {
2503 MUTEX_ENTER(&conn->conn_data_lock);
2504 conn->flags &= ~RX_CONN_BUSY;
2505 MUTEX_EXIT(&conn->conn_data_lock);
2506 MUTEX_EXIT(&conn->conn_call_lock);
2510 * Map errors to the local host's errno.h format.
2512 error = ntoh_syserr_conv(error);
2516 #if !defined(KERNEL)
2518 /* Call this routine when shutting down a server or client (especially
2519 * clients). This will allow Rx to gracefully garbage collect server
2520 * connections, and reduce the number of retries that a server might
2521 * make to a dead client.
2522 * This is not quite right, since some calls may still be ongoing and
2523 * we can't lock them to destroy them. */
2527 struct rx_connection **conn_ptr, **conn_end;
2530 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
2531 return; /* Already shutdown. */
2533 rxi_DeleteCachedConnections();
2534 if (rx_connHashTable) {
2535 MUTEX_ENTER(&rx_connHashTable_lock);
2536 for (conn_ptr = &rx_connHashTable[0], conn_end =
2537 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2539 struct rx_connection *conn, *next;
2540 for (conn = *conn_ptr; conn; conn = next) {
2542 if (conn->type == RX_CLIENT_CONNECTION) {
2543 MUTEX_ENTER(&rx_refcnt_mutex);
2545 MUTEX_EXIT(&rx_refcnt_mutex);
2546 #ifdef RX_ENABLE_LOCKS
2547 rxi_DestroyConnectionNoLock(conn);
2548 #else /* RX_ENABLE_LOCKS */
2549 rxi_DestroyConnection(conn);
2550 #endif /* RX_ENABLE_LOCKS */
2554 #ifdef RX_ENABLE_LOCKS
2555 while (rx_connCleanup_list) {
2556 struct rx_connection *conn;
2557 conn = rx_connCleanup_list;
2558 rx_connCleanup_list = rx_connCleanup_list->next;
2559 MUTEX_EXIT(&rx_connHashTable_lock);
2560 rxi_CleanupConnection(conn);
2561 MUTEX_ENTER(&rx_connHashTable_lock);
2563 MUTEX_EXIT(&rx_connHashTable_lock);
2564 #endif /* RX_ENABLE_LOCKS */
2569 afs_winsockCleanup();
2575 /* if we wakeup packet waiter too often, can get in loop with two
2576 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2578 rxi_PacketsUnWait(void)
2580 if (!rx_waitingForPackets) {
2584 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2585 return; /* still over quota */
2588 rx_waitingForPackets = 0;
2589 #ifdef RX_ENABLE_LOCKS
2590 CV_BROADCAST(&rx_waitingForPackets_cv);
2592 osi_rxWakeup(&rx_waitingForPackets);
2598 /* ------------------Internal interfaces------------------------- */
2600 /* Return this process's service structure for the
2601 * specified socket and service */
2602 static struct rx_service *
2603 rxi_FindService(osi_socket socket, u_short serviceId)
2605 struct rx_service **sp;
2606 for (sp = &rx_services[0]; *sp; sp++) {
2607 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2613 #ifdef RXDEBUG_PACKET
2614 #ifdef KDUMP_RX_LOCK
2615 static struct rx_call_rx_lock *rx_allCallsp = 0;
2617 static struct rx_call *rx_allCallsp = 0;
2619 #endif /* RXDEBUG_PACKET */
2621 /* Allocate a call structure, for the indicated channel of the
2622 * supplied connection. The mode and state of the call must be set by
2623 * the caller. Returns the call with mutex locked. */
2624 static struct rx_call *
2625 rxi_NewCall(struct rx_connection *conn, int channel)
2627 struct rx_call *call;
2628 #ifdef RX_ENABLE_LOCKS
2629 struct rx_call *cp; /* Call pointer temp */
2630 struct opr_queue *cursor;
2633 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2635 /* Grab an existing call structure, or allocate a new one.
2636 * Existing call structures are assumed to have been left reset by
2638 MUTEX_ENTER(&rx_freeCallQueue_lock);
2640 #ifdef RX_ENABLE_LOCKS
2642 * EXCEPT that the TQ might not yet be cleared out.
2643 * Skip over those with in-use TQs.
2646 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2647 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2648 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2654 #else /* RX_ENABLE_LOCKS */
2655 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2656 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2657 #endif /* RX_ENABLE_LOCKS */
2658 opr_queue_Remove(&call->entry);
2659 if (rx_stats_active)
2660 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2661 MUTEX_EXIT(&rx_freeCallQueue_lock);
2662 MUTEX_ENTER(&call->lock);
2663 CLEAR_CALL_QUEUE_LOCK(call);
2664 #ifdef RX_ENABLE_LOCKS
2665 /* Now, if TQ wasn't cleared earlier, do it now. */
2666 rxi_WaitforTQBusy(call);
2667 if (call->flags & RX_CALL_TQ_CLEARME) {
2668 rxi_ClearTransmitQueue(call, 1);
2669 /*queue_Init(&call->tq);*/
2671 #endif /* RX_ENABLE_LOCKS */
2672 /* Bind the call to its connection structure */
2674 rxi_ResetCall(call, 1);
2677 call = rxi_Alloc(sizeof(struct rx_call));
2678 #ifdef RXDEBUG_PACKET
2679 call->allNextp = rx_allCallsp;
2680 rx_allCallsp = call;
2682 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2683 #else /* RXDEBUG_PACKET */
2684 rx_atomic_inc(&rx_stats.nCallStructs);
2685 #endif /* RXDEBUG_PACKET */
2687 MUTEX_EXIT(&rx_freeCallQueue_lock);
2688 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2689 MUTEX_ENTER(&call->lock);
2690 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2691 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2692 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2694 /* Initialize once-only items */
2695 opr_queue_Init(&call->tq);
2696 opr_queue_Init(&call->rq);
2697 opr_queue_Init(&call->app.iovq);
2698 #ifdef RXDEBUG_PACKET
2699 call->rqc = call->tqc = call->iovqc = 0;
2700 #endif /* RXDEBUG_PACKET */
2701 /* Bind the call to its connection structure (prereq for reset) */
2703 rxi_ResetCall(call, 1);
2705 call->channel = channel;
2706 call->callNumber = &conn->callNumber[channel];
2707 call->rwind = conn->rwind[channel];
2708 call->twind = conn->twind[channel];
2709 /* Note that the next expected call number is retained (in
2710 * conn->callNumber[i]), even if we reallocate the call structure
2712 conn->call[channel] = call;
2713 /* if the channel's never been used (== 0), we should start at 1, otherwise
2714 * the call number is valid from the last time this channel was used */
2715 if (*call->callNumber == 0)
2716 *call->callNumber = 1;
2721 /* A call has been inactive long enough that so we can throw away
2722 * state, including the call structure, which is placed on the call
2725 * call->lock amd rx_refcnt_mutex are held upon entry.
2726 * haveCTLock is set when called from rxi_ReapConnections.
2728 * return 1 if the call is freed, 0 if not.
2731 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2733 int channel = call->channel;
2734 struct rx_connection *conn = call->conn;
2735 u_char state = call->state;
2738 * We are setting the state to RX_STATE_RESET to
2739 * ensure that no one else will attempt to use this
2740 * call once we drop the refcnt lock. We must drop
2741 * the refcnt lock before calling rxi_ResetCall
2742 * because it cannot be held across acquiring the
2743 * freepktQ lock. NewCall does the same.
2745 call->state = RX_STATE_RESET;
2746 MUTEX_EXIT(&rx_refcnt_mutex);
2747 rxi_ResetCall(call, 0);
2749 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2751 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2752 (*call->callNumber)++;
2754 if (call->conn->call[channel] == call)
2755 call->conn->call[channel] = 0;
2756 MUTEX_EXIT(&conn->conn_call_lock);
2759 * We couldn't obtain the conn_call_lock so we can't
2760 * disconnect the call from the connection. Set the
2761 * call state to dally so that the call can be reused.
2763 MUTEX_ENTER(&rx_refcnt_mutex);
2764 call->state = RX_STATE_DALLY;
2768 MUTEX_ENTER(&rx_freeCallQueue_lock);
2769 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2770 #ifdef RX_ENABLE_LOCKS
2771 /* A call may be free even though its transmit queue is still in use.
2772 * Since we search the call list from head to tail, put busy calls at
2773 * the head of the list, and idle calls at the tail.
2775 if (call->flags & RX_CALL_TQ_BUSY)
2776 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2778 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2779 #else /* RX_ENABLE_LOCKS */
2780 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2781 #endif /* RX_ENABLE_LOCKS */
2782 if (rx_stats_active)
2783 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2784 MUTEX_EXIT(&rx_freeCallQueue_lock);
2786 /* Destroy the connection if it was previously slated for
2787 * destruction, i.e. the Rx client code previously called
2788 * rx_DestroyConnection (client connections), or
2789 * rxi_ReapConnections called the same routine (server
2790 * connections). Only do this, however, if there are no
2791 * outstanding calls. Note that for fine grain locking, there appears
2792 * to be a deadlock in that rxi_FreeCall has a call locked and
2793 * DestroyConnectionNoLock locks each call in the conn. But note a
2794 * few lines up where we have removed this call from the conn.
2795 * If someone else destroys a connection, they either have no
2796 * call lock held or are going through this section of code.
2798 MUTEX_ENTER(&conn->conn_data_lock);
2799 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2800 MUTEX_ENTER(&rx_refcnt_mutex);
2802 MUTEX_EXIT(&rx_refcnt_mutex);
2803 MUTEX_EXIT(&conn->conn_data_lock);
2804 #ifdef RX_ENABLE_LOCKS
2806 rxi_DestroyConnectionNoLock(conn);
2808 rxi_DestroyConnection(conn);
2809 #else /* RX_ENABLE_LOCKS */
2810 rxi_DestroyConnection(conn);
2811 #endif /* RX_ENABLE_LOCKS */
2813 MUTEX_EXIT(&conn->conn_data_lock);
2815 MUTEX_ENTER(&rx_refcnt_mutex);
2819 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2820 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2823 rxi_Alloc(size_t size)
2827 if (rx_stats_active) {
2828 rx_atomic_add(&rxi_Allocsize, (int) size);
2829 rx_atomic_inc(&rxi_Alloccnt);
2833 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2834 afs_osi_Alloc_NoSleep(size);
2839 osi_Panic("rxi_Alloc error");
2845 rxi_Free(void *addr, size_t size)
2847 if (rx_stats_active) {
2848 rx_atomic_sub(&rxi_Allocsize, (int) size);
2849 rx_atomic_dec(&rxi_Alloccnt);
2851 osi_Free(addr, size);
2855 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2857 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2858 struct rx_peer *next = NULL;
2862 MUTEX_ENTER(&rx_peerHashTable_lock);
2864 peer_ptr = &rx_peerHashTable[0];
2865 peer_end = &rx_peerHashTable[rx_hashTableSize];
2868 for ( ; peer_ptr < peer_end; peer_ptr++) {
2871 for ( ; peer; peer = next) {
2873 if (host == peer->host)
2878 hashIndex = PEER_HASH(host, port);
2879 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2880 if ((peer->host == host) && (peer->port == port))
2885 MUTEX_ENTER(&rx_peerHashTable_lock);
2890 MUTEX_EXIT(&rx_peerHashTable_lock);
2892 MUTEX_ENTER(&peer->peer_lock);
2893 /* We don't handle dropping below min, so don't */
2894 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2895 peer->ifMTU=MIN(mtu, peer->ifMTU);
2896 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2897 /* if we tweaked this down, need to tune our peer MTU too */
2898 peer->MTU = MIN(peer->MTU, peer->natMTU);
2899 /* if we discovered a sub-1500 mtu, degrade */
2900 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2901 peer->maxDgramPackets = 1;
2902 /* We no longer have valid peer packet information */
2903 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2904 peer->maxPacketSize = 0;
2905 MUTEX_EXIT(&peer->peer_lock);
2907 MUTEX_ENTER(&rx_peerHashTable_lock);
2909 if (host && !port) {
2911 /* pick up where we left off */
2915 MUTEX_EXIT(&rx_peerHashTable_lock);
2918 #ifdef AFS_RXERRQ_ENV
2920 rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
2922 int hashIndex = PEER_HASH(host, port);
2923 struct rx_peer *peer;
2925 MUTEX_ENTER(&rx_peerHashTable_lock);
2927 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2928 if (peer->host == host && peer->port == port) {
2934 MUTEX_EXIT(&rx_peerHashTable_lock);
2937 rx_atomic_inc(&peer->neterrs);
2938 MUTEX_ENTER(&peer->peer_lock);
2939 peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
2940 peer->last_err_type = err->ee_type;
2941 peer->last_err_code = err->ee_code;
2942 MUTEX_EXIT(&peer->peer_lock);
2944 MUTEX_ENTER(&rx_peerHashTable_lock);
2946 MUTEX_EXIT(&rx_peerHashTable_lock);
2951 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2953 # ifdef AFS_ADAPT_PMTU
2954 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2955 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2959 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2960 switch (err->ee_code) {
2961 case ICMP_NET_UNREACH:
2962 case ICMP_HOST_UNREACH:
2963 case ICMP_PORT_UNREACH:
2966 rxi_SetPeerDead(err, addr, port);
2973 rxi_TranslateICMP(int type, int code)
2976 case ICMP_DEST_UNREACH:
2978 case ICMP_NET_UNREACH:
2979 return "Destination Net Unreachable";
2980 case ICMP_HOST_UNREACH:
2981 return "Destination Host Unreachable";
2982 case ICMP_PROT_UNREACH:
2983 return "Destination Protocol Unreachable";
2984 case ICMP_PORT_UNREACH:
2985 return "Destination Port Unreachable";
2987 return "Destination Net Prohibited";
2989 return "Destination Host Prohibited";
2995 #endif /* AFS_RXERRQ_ENV */
2998 * Get the last network error for a connection
3000 * A "network error" here means an error retrieved from ICMP, or some other
3001 * mechanism outside of Rx that informs us of errors in network reachability.
3003 * If a peer associated with the given Rx connection has received a network
3004 * error recently, this function allows the caller to know what error
3005 * specifically occurred. This can be useful to know, since e.g. ICMP errors
3006 * can cause calls to that peer to be quickly aborted. So, this function can
3007 * help see why a call was aborted due to network errors.
3009 * If we have received traffic from a peer since the last network error, we
3010 * treat that peer as if we had not received an network error for it.
3012 * @param[in] conn The Rx connection to examine
3013 * @param[out] err_origin The origin of the last network error (e.g. ICMP);
3014 * one of the RX_NETWORK_ERROR_ORIGIN_* constants
3015 * @param[out] err_type The type of the last error
3016 * @param[out] err_code The code of the last error
3017 * @param[out] msg Human-readable error message, if applicable; NULL otherwise
3019 * @return If we have an error
3020 * @retval -1 No error to get; 'out' params are undefined
3021 * @retval 0 We have an error; 'out' params contain the last error
3024 rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
3025 int *err_code, const char **msg)
3027 #ifdef AFS_RXERRQ_ENV
3028 struct rx_peer *peer = conn->peer;
3029 if (rx_atomic_read(&peer->neterrs)) {
3030 MUTEX_ENTER(&peer->peer_lock);
3031 *err_origin = peer->last_err_origin;
3032 *err_type = peer->last_err_type;
3033 *err_code = peer->last_err_code;
3034 MUTEX_EXIT(&peer->peer_lock);
3037 if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
3038 *msg = rxi_TranslateICMP(*err_type, *err_code);
3047 /* Find the peer process represented by the supplied (host,port)
3048 * combination. If there is no appropriate active peer structure, a
3049 * new one will be allocated and initialized
3052 rxi_FindPeer(afs_uint32 host, u_short port, int create)
3056 hashIndex = PEER_HASH(host, port);
3057 MUTEX_ENTER(&rx_peerHashTable_lock);
3058 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
3059 if ((pp->host == host) && (pp->port == port))
3064 pp = rxi_AllocPeer(); /* This bzero's *pp */
3065 pp->host = host; /* set here or in InitPeerParams is zero */
3067 #ifdef AFS_RXERRQ_ENV
3068 rx_atomic_set(&pp->neterrs, 0);
3070 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3071 opr_queue_Init(&pp->rpcStats);
3072 pp->next = rx_peerHashTable[hashIndex];
3073 rx_peerHashTable[hashIndex] = pp;
3074 rxi_InitPeerParams(pp);
3075 if (rx_stats_active)
3076 rx_atomic_inc(&rx_stats.nPeerStructs);
3082 MUTEX_EXIT(&rx_peerHashTable_lock);
3087 /* Find the connection at (host, port) started at epoch, and with the
3088 * given connection id. Creates the server connection if necessary.
3089 * The type specifies whether a client connection or a server
3090 * connection is desired. In both cases, (host, port) specify the
3091 * peer's (host, pair) pair. Client connections are not made
3092 * automatically by this routine. The parameter socket gives the
3093 * socket descriptor on which the packet was received. This is used,
3094 * in the case of server connections, to check that *new* connections
3095 * come via a valid (port, serviceId). Finally, the securityIndex
3096 * parameter must match the existing index for the connection. If a
3097 * server connection is created, it will be created using the supplied
3098 * index, if the index is valid for this service */
3099 static struct rx_connection *
3100 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3101 u_short port, u_short serviceId, afs_uint32 cid,
3102 afs_uint32 epoch, int type, u_int securityIndex,
3103 int *unknownService)
3105 int hashindex, flag, i;
3106 struct rx_connection *conn;
3107 *unknownService = 0;
3108 hashindex = CONN_HASH(host, port, cid, epoch, type);
3109 MUTEX_ENTER(&rx_connHashTable_lock);
3110 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3111 rx_connHashTable[hashindex],
3114 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3115 && (epoch == conn->epoch)) {
3116 struct rx_peer *pp = conn->peer;
3117 if (securityIndex != conn->securityIndex) {
3118 /* this isn't supposed to happen, but someone could forge a packet
3119 * like this, and there seems to be some CM bug that makes this
3120 * happen from time to time -- in which case, the fileserver
3122 MUTEX_EXIT(&rx_connHashTable_lock);
3123 return (struct rx_connection *)0;
3125 if (pp->host == host && pp->port == port)
3127 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3129 /* So what happens when it's a callback connection? */
3130 if ( /*type == RX_CLIENT_CONNECTION && */
3131 (conn->epoch & 0x80000000))
3135 /* the connection rxLastConn that was used the last time is not the
3136 ** one we are looking for now. Hence, start searching in the hash */
3138 conn = rx_connHashTable[hashindex];
3143 struct rx_service *service;
3144 if (type == RX_CLIENT_CONNECTION) {
3145 MUTEX_EXIT(&rx_connHashTable_lock);
3146 return (struct rx_connection *)0;
3148 service = rxi_FindService(socket, serviceId);
3149 if (!service || (securityIndex >= service->nSecurityObjects)
3150 || (service->securityObjects[securityIndex] == 0)) {
3151 MUTEX_EXIT(&rx_connHashTable_lock);
3152 *unknownService = 1;
3153 return (struct rx_connection *)0;
3155 conn = rxi_AllocConnection(); /* This bzero's the connection */
3156 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3157 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3158 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3159 conn->next = rx_connHashTable[hashindex];
3160 rx_connHashTable[hashindex] = conn;
3161 conn->peer = rxi_FindPeer(host, port, 1);
3162 conn->type = RX_SERVER_CONNECTION;
3163 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3164 conn->epoch = epoch;
3165 conn->cid = cid & RX_CIDMASK;
3166 conn->ackRate = RX_FAST_ACK_RATE;
3167 conn->service = service;
3168 conn->serviceId = serviceId;
3169 conn->securityIndex = securityIndex;
3170 conn->securityObject = service->securityObjects[securityIndex];
3171 conn->nSpecific = 0;
3172 conn->specific = NULL;
3173 rx_SetConnDeadTime(conn, service->connDeadTime);
3174 conn->idleDeadTime = service->idleDeadTime;
3175 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3176 for (i = 0; i < RX_MAXCALLS; i++) {
3177 conn->twind[i] = rx_initSendWindow;
3178 conn->rwind[i] = rx_initReceiveWindow;
3180 /* Notify security object of the new connection */
3181 RXS_NewConnection(conn->securityObject, conn);
3182 /* XXXX Connection timeout? */
3183 if (service->newConnProc)
3184 (*service->newConnProc) (conn);
3185 if (rx_stats_active)
3186 rx_atomic_inc(&rx_stats.nServerConns);
3189 MUTEX_ENTER(&rx_refcnt_mutex);
3191 MUTEX_EXIT(&rx_refcnt_mutex);
3193 rxLastConn = conn; /* store this connection as the last conn used */
3194 MUTEX_EXIT(&rx_connHashTable_lock);
3199 * Timeout a call on a busy call channel if appropriate.
3201 * @param[in] call The busy call.
3203 * @pre 'call' is marked as busy (namely,
3204 * call->conn->lastBusy[call->channel] != 0)
3206 * @pre call->lock is held
3207 * @pre rxi_busyChannelError is nonzero
3209 * @note call->lock is dropped and reacquired
3212 rxi_CheckBusy(struct rx_call *call)
3214 struct rx_connection *conn = call->conn;
3215 int channel = call->channel;
3216 int freechannel = 0;
3219 MUTEX_EXIT(&call->lock);
3221 MUTEX_ENTER(&conn->conn_call_lock);
3223 /* Are there any other call slots on this conn that we should try? Look for
3224 * slots that are empty and are either non-busy, or were marked as busy
3225 * longer than conn->secondsUntilDead seconds before this call started. */
3227 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3229 /* only look at channels that aren't us */
3233 if (conn->lastBusy[i]) {
3234 /* if this channel looked busy too recently, don't look at it */
3235 if (conn->lastBusy[i] >= call->startTime.sec) {
3238 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3243 if (conn->call[i]) {
3244 struct rx_call *tcall = conn->call[i];
3245 MUTEX_ENTER(&tcall->lock);
3246 if (tcall->state == RX_STATE_DALLY) {
3249 MUTEX_EXIT(&tcall->lock);
3255 MUTEX_ENTER(&call->lock);
3257 /* Since the call->lock has been released it is possible that the call may
3258 * no longer be busy (the call channel cannot have been reallocated as we
3259 * haven't dropped the conn_call_lock) Therefore, we must confirm
3260 * that the call state has not changed when deciding whether or not to
3261 * force this application thread to retry by forcing a Timeout error. */
3263 if (freechannel && (call->flags & RX_CALL_PEER_BUSY)) {
3264 /* Since 'freechannel' is set, there exists another channel in this
3265 * rx_conn that the application thread might be able to use. We know
3266 * that we have the correct call since callNumber is unchanged, and we
3267 * know that the call is still busy. So, set the call error state to
3268 * rxi_busyChannelError so the application can retry the request,
3269 * presumably on a less-busy call channel. */
3271 rxi_CallError(call, RX_CALL_BUSY);
3273 MUTEX_EXIT(&conn->conn_call_lock);
3277 * Abort the call if the server is over the busy threshold. This
3278 * can be used without requiring a call structure be initialised,
3279 * or connected to a particular channel
3282 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3283 struct rx_packet *np)
3285 if ((rx_BusyThreshold > 0) &&
3286 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3287 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3288 rx_BusyError, np, 0);
3289 if (rx_stats_active)
3290 rx_atomic_inc(&rx_stats.nBusies);
3297 static_inline struct rx_call *
3298 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3301 struct rx_call *call;
3303 channel = np->header.cid & RX_CHANNELMASK;
3304 MUTEX_ENTER(&conn->conn_call_lock);
3305 call = conn->call[channel];
3306 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3307 MUTEX_EXIT(&conn->conn_call_lock);
3308 if (rx_stats_active)
3309 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3313 MUTEX_ENTER(&call->lock);
3314 MUTEX_EXIT(&conn->conn_call_lock);
3316 if ((call->state == RX_STATE_DALLY)
3317 && np->header.type == RX_PACKET_TYPE_ACK) {
3318 if (rx_stats_active)
3319 rx_atomic_inc(&rx_stats.ignorePacketDally);
3320 MUTEX_EXIT(&call->lock);
3327 static_inline struct rx_call *
3328 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3329 struct rx_connection *conn)
3332 struct rx_call *call;
3334 channel = np->header.cid & RX_CHANNELMASK;
3335 MUTEX_ENTER(&conn->conn_call_lock);
3336 call = conn->call[channel];
3339 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3340 MUTEX_EXIT(&conn->conn_call_lock);
3344 call = rxi_NewCall(conn, channel); /* returns locked call */
3345 *call->callNumber = np->header.callNumber;
3346 MUTEX_EXIT(&conn->conn_call_lock);
3348 call->state = RX_STATE_PRECALL;
3349 clock_GetTime(&call->queueTime);
3350 call->app.bytesSent = 0;
3351 call->app.bytesRcvd = 0;
3352 rxi_KeepAliveOn(call);
3357 if (np->header.callNumber == conn->callNumber[channel]) {
3358 MUTEX_ENTER(&call->lock);
3359 MUTEX_EXIT(&conn->conn_call_lock);
3363 if (np->header.callNumber < conn->callNumber[channel]) {
3364 MUTEX_EXIT(&conn->conn_call_lock);
3365 if (rx_stats_active)
3366 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3370 MUTEX_ENTER(&call->lock);
3371 MUTEX_EXIT(&conn->conn_call_lock);
3373 /* Wait until the transmit queue is idle before deciding
3374 * whether to reset the current call. Chances are that the
3375 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3378 #ifdef RX_ENABLE_LOCKS
3379 if (call->state == RX_STATE_ACTIVE && !call->error) {
3380 rxi_WaitforTQBusy(call);
3381 /* If we entered error state while waiting,
3382 * must call rxi_CallError to permit rxi_ResetCall
3383 * to processed when the tqWaiter count hits zero.
3386 rxi_CallError(call, call->error);
3387 MUTEX_EXIT(&call->lock);
3391 #endif /* RX_ENABLE_LOCKS */
3392 /* If the new call cannot be taken right now send a busy and set
3393 * the error condition in this call, so that it terminates as
3394 * quickly as possible */
3395 if (call->state == RX_STATE_ACTIVE) {
3396 rxi_CallError(call, RX_CALL_DEAD);
3397 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3399 MUTEX_EXIT(&call->lock);
3403 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3404 MUTEX_EXIT(&call->lock);
3408 rxi_ResetCall(call, 0);
3409 /* The conn_call_lock is not held but no one else should be
3410 * using this call channel while we are processing this incoming
3411 * packet. This assignment should be safe.
3413 *call->callNumber = np->header.callNumber;
3414 call->state = RX_STATE_PRECALL;
3415 clock_GetTime(&call->queueTime);
3416 call->app.bytesSent = 0;
3417 call->app.bytesRcvd = 0;
3418 rxi_KeepAliveOn(call);
3424 /* There are two packet tracing routines available for testing and monitoring
3425 * Rx. One is called just after every packet is received and the other is
3426 * called just before every packet is sent. Received packets, have had their
3427 * headers decoded, and packets to be sent have not yet had their headers
3428 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3429 * containing the network address. Both can be modified. The return value, if
3430 * non-zero, indicates that the packet should be dropped. */
3432 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3433 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3435 /* A packet has been received off the interface. Np is the packet, socket is
3436 * the socket number it was received from (useful in determining which service
3437 * this packet corresponds to), and (host, port) reflect the host,port of the
3438 * sender. This call returns the packet to the caller if it is finished with
3439 * it, rather than de-allocating it, just as a small performance hack */
3442 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3443 afs_uint32 host, u_short port, int *tnop,
3444 struct rx_call **newcallp)
3446 struct rx_call *call;
3447 struct rx_connection *conn;
3449 int unknownService = 0;
3453 struct rx_packet *tnp;
3456 /* We don't print out the packet until now because (1) the time may not be
3457 * accurate enough until now in the lwp implementation (rx_Listener only gets
3458 * the time after the packet is read) and (2) from a protocol point of view,
3459 * this is the first time the packet has been seen */
3460 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3461 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3462 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3463 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3464 np->header.epoch, np->header.cid, np->header.callNumber,
3465 np->header.seq, np->header.flags, np));
3468 /* Account for connectionless packets */
3469 if (rx_stats_active &&
3470 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3471 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3472 struct rx_peer *peer;
3474 /* Try to look up the peer structure, but don't create one */
3475 peer = rxi_FindPeer(host, port, 0);
3477 /* Since this may not be associated with a connection, it may have
3478 * no refCount, meaning we could race with ReapConnections
3481 if (peer && (peer->refCount > 0)) {
3482 #ifdef AFS_RXERRQ_ENV
3483 if (rx_atomic_read(&peer->neterrs)) {
3484 rx_atomic_set(&peer->neterrs, 0);
3487 MUTEX_ENTER(&peer->peer_lock);
3488 peer->bytesReceived += np->length;
3489 MUTEX_EXIT(&peer->peer_lock);
3493 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3494 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3497 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3498 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3501 /* If an input tracer function is defined, call it with the packet and
3502 * network address. Note this function may modify its arguments. */
3503 if (rx_justReceived) {
3504 struct sockaddr_in addr;
3506 addr.sin_family = AF_INET;
3507 addr.sin_port = port;
3508 addr.sin_addr.s_addr = host;
3509 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3510 addr.sin_len = sizeof(addr);
3511 #endif /* AFS_OSF_ENV */
3512 drop = (*rx_justReceived) (np, &addr);
3513 /* drop packet if return value is non-zero */
3516 port = addr.sin_port; /* in case fcn changed addr */
3517 host = addr.sin_addr.s_addr;
3521 /* If packet was not sent by the client, then *we* must be the client */
3522 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3523 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3525 /* Find the connection (or fabricate one, if we're the server & if
3526 * necessary) associated with this packet */
3528 rxi_FindConnection(socket, host, port, np->header.serviceId,
3529 np->header.cid, np->header.epoch, type,
3530 np->header.securityIndex, &unknownService);
3532 /* To avoid having 2 connections just abort at each other,
3533 don't abort an abort. */
3535 if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
3536 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3541 #ifdef AFS_RXERRQ_ENV
3542 if (rx_atomic_read(&conn->peer->neterrs)) {
3543 rx_atomic_set(&conn->peer->neterrs, 0);
3547 /* If we're doing statistics, then account for the incoming packet */
3548 if (rx_stats_active) {
3549 MUTEX_ENTER(&conn->peer->peer_lock);
3550 conn->peer->bytesReceived += np->length;
3551 MUTEX_EXIT(&conn->peer->peer_lock);
3554 /* If the connection is in an error state, send an abort packet and ignore
3555 * the incoming packet */
3557 /* Don't respond to an abort packet--we don't want loops! */
3558 MUTEX_ENTER(&conn->conn_data_lock);
3559 if (np->header.type != RX_PACKET_TYPE_ABORT)
3560 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3561 putConnection(conn);
3562 MUTEX_EXIT(&conn->conn_data_lock);
3566 /* Check for connection-only requests (i.e. not call specific). */
3567 if (np->header.callNumber == 0) {
3568 switch (np->header.type) {
3569 case RX_PACKET_TYPE_ABORT: {
3570 /* What if the supplied error is zero? */
3571 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3572 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3573 rxi_ConnectionError(conn, errcode);
3574 putConnection(conn);
3577 case RX_PACKET_TYPE_CHALLENGE:
3578 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3579 putConnection(conn);
3581 case RX_PACKET_TYPE_RESPONSE:
3582 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3583 putConnection(conn);
3585 case RX_PACKET_TYPE_PARAMS:
3586 case RX_PACKET_TYPE_PARAMS + 1:
3587 case RX_PACKET_TYPE_PARAMS + 2:
3588 /* ignore these packet types for now */
3589 putConnection(conn);
3593 /* Should not reach here, unless the peer is broken: send an
3595 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3596 MUTEX_ENTER(&conn->conn_data_lock);
3597 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3598 putConnection(conn);
3599 MUTEX_EXIT(&conn->conn_data_lock);
3604 if (type == RX_SERVER_CONNECTION)
3605 call = rxi_ReceiveServerCall(socket, np, conn);
3607 call = rxi_ReceiveClientCall(np, conn);
3610 putConnection(conn);
3614 MUTEX_ASSERT(&call->lock);
3615 /* Set remote user defined status from packet */
3616 call->remoteStatus = np->header.userStatus;
3618 /* Now do packet type-specific processing */
3619 switch (np->header.type) {
3620 case RX_PACKET_TYPE_DATA:
3621 /* If we're a client, and receiving a response, then all the packets
3622 * we transmitted packets are implicitly acknowledged. */
3623 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3624 rxi_AckAllInTransmitQueue(call);
3626 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3629 case RX_PACKET_TYPE_ACK:
3630 /* Respond immediately to ack packets requesting acknowledgement
3632 if (np->header.flags & RX_REQUEST_ACK) {
3634 (void)rxi_SendCallAbort(call, 0, 1, 0);
3636 (void)rxi_SendAck(call, 0, np->header.serial,
3637 RX_ACK_PING_RESPONSE, 1);
3639 np = rxi_ReceiveAckPacket(call, np, 1);
3641 case RX_PACKET_TYPE_ABORT: {
3642 /* An abort packet: reset the call, passing the error up to the user. */
3643 /* What if error is zero? */
3644 /* What if the error is -1? the application will treat it as a timeout. */
3645 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3646 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3647 rxi_CallError(call, errdata);
3648 MUTEX_EXIT(&call->lock);
3649 putConnection(conn);
3650 return np; /* xmitting; drop packet */
3652 case RX_PACKET_TYPE_BUSY: {
3653 struct clock busyTime;
3655 clock_GetTime(&busyTime);
3657 MUTEX_EXIT(&call->lock);
3659 MUTEX_ENTER(&conn->conn_call_lock);
3660 MUTEX_ENTER(&call->lock);
3661 conn->lastBusy[call->channel] = busyTime.sec;
3662 call->flags |= RX_CALL_PEER_BUSY;
3663 MUTEX_EXIT(&call->lock);
3664 MUTEX_EXIT(&conn->conn_call_lock);
3666 putConnection(conn);
3670 case RX_PACKET_TYPE_ACKALL:
3671 /* All packets acknowledged, so we can drop all packets previously
3672 * readied for sending */
3673 rxi_AckAllInTransmitQueue(call);
3676 /* Should not reach here, unless the peer is broken: send an abort
3678 rxi_CallError(call, RX_PROTOCOL_ERROR);
3679 np = rxi_SendCallAbort(call, np, 1, 0);
3682 /* Note when this last legitimate packet was received, for keep-alive
3683 * processing. Note, we delay getting the time until now in the hope that
3684 * the packet will be delivered to the user before any get time is required
3685 * (if not, then the time won't actually be re-evaluated here). */
3686 call->lastReceiveTime = clock_Sec();
3687 /* we've received a legit packet, so the channel is not busy */
3688 call->flags &= ~RX_CALL_PEER_BUSY;
3689 MUTEX_EXIT(&call->lock);
3690 putConnection(conn);
3694 /* return true if this is an "interesting" connection from the point of view
3695 of someone trying to debug the system */
3697 rxi_IsConnInteresting(struct rx_connection *aconn)
3700 struct rx_call *tcall;
3702 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3705 for (i = 0; i < RX_MAXCALLS; i++) {
3706 tcall = aconn->call[i];
3708 if ((tcall->state == RX_STATE_PRECALL)
3709 || (tcall->state == RX_STATE_ACTIVE))
3711 if ((tcall->app.mode == RX_MODE_SENDING)
3712 || (tcall->app.mode == RX_MODE_RECEIVING))
3720 /* if this is one of the last few packets AND it wouldn't be used by the
3721 receiving call to immediately satisfy a read request, then drop it on
3722 the floor, since accepting it might prevent a lock-holding thread from
3723 making progress in its reading. If a call has been cleared while in
3724 the precall state then ignore all subsequent packets until the call
3725 is assigned to a thread. */
3728 TooLow(struct rx_packet *ap, struct rx_call *acall)
3732 MUTEX_ENTER(&rx_quota_mutex);
3733 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3734 && (acall->state == RX_STATE_PRECALL))
3735 || ((rx_nFreePackets < rxi_dataQuota + 2)
3736 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3737 && (acall->flags & RX_CALL_READER_WAIT)))) {
3740 MUTEX_EXIT(&rx_quota_mutex);
3746 * Clear the attach wait flag on a connection and proceed.
3748 * Any processing waiting for a connection to be attached should be
3749 * unblocked. We clear the flag and do any other needed tasks.
3752 * the conn to unmark waiting for attach
3754 * @pre conn's conn_data_lock must be locked before calling this function
3758 rxi_ConnClearAttachWait(struct rx_connection *conn)
3760 /* Indicate that rxi_CheckReachEvent is no longer running by
3761 * clearing the flag. Must be atomic under conn_data_lock to
3762 * avoid a new call slipping by: rxi_CheckConnReach holds
3763 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3765 conn->flags &= ~RX_CONN_ATTACHWAIT;
3766 if (conn->flags & RX_CONN_NAT_PING) {
3767 conn->flags &= ~RX_CONN_NAT_PING;
3768 rxi_ScheduleNatKeepAliveEvent(conn);
3773 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3775 struct rx_connection *conn = arg1;
3776 struct rx_call *acall = arg2;
3777 struct rx_call *call = acall;
3778 struct clock when, now;
3781 MUTEX_ENTER(&conn->conn_data_lock);
3784 rxevent_Put(&conn->checkReachEvent);
3786 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3788 putConnection(conn);
3790 MUTEX_EXIT(&conn->conn_data_lock);
3794 MUTEX_ENTER(&conn->conn_call_lock);
3795 MUTEX_ENTER(&conn->conn_data_lock);
3796 for (i = 0; i < RX_MAXCALLS; i++) {
3797 struct rx_call *tc = conn->call[i];
3798 if (tc && tc->state == RX_STATE_PRECALL) {
3804 rxi_ConnClearAttachWait(conn);
3805 MUTEX_EXIT(&conn->conn_data_lock);
3806 MUTEX_EXIT(&conn->conn_call_lock);
3811 MUTEX_ENTER(&call->lock);
3812 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3814 MUTEX_EXIT(&call->lock);
3816 clock_GetTime(&now);
3818 when.sec += RX_CHECKREACH_TIMEOUT;
3819 MUTEX_ENTER(&conn->conn_data_lock);
3820 if (!conn->checkReachEvent) {
3821 MUTEX_ENTER(&rx_refcnt_mutex);
3823 MUTEX_EXIT(&rx_refcnt_mutex);
3824 conn->checkReachEvent = rxevent_Post(&when, &now,
3825 rxi_CheckReachEvent, conn,
3828 MUTEX_EXIT(&conn->conn_data_lock);
3834 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3836 struct rx_service *service = conn->service;
3837 struct rx_peer *peer = conn->peer;
3838 afs_uint32 now, lastReach;
3840 if (service->checkReach == 0)
3844 MUTEX_ENTER(&peer->peer_lock);
3845 lastReach = peer->lastReachTime;
3846 MUTEX_EXIT(&peer->peer_lock);
3847 if (now - lastReach < RX_CHECKREACH_TTL)
3850 MUTEX_ENTER(&conn->conn_data_lock);
3851 if (conn->flags & RX_CONN_ATTACHWAIT) {
3852 MUTEX_EXIT(&conn->conn_data_lock);
3855 conn->flags |= RX_CONN_ATTACHWAIT;
3856 MUTEX_EXIT(&conn->conn_data_lock);
3857 if (!conn->checkReachEvent)
3858 rxi_CheckReachEvent(NULL, conn, call, 0);
3863 /* try to attach call, if authentication is complete */
3865 TryAttach(struct rx_call *acall, osi_socket socket,
3866 int *tnop, struct rx_call **newcallp,
3869 struct rx_connection *conn = acall->conn;
3871 if (conn->type == RX_SERVER_CONNECTION
3872 && acall->state == RX_STATE_PRECALL) {
3873 /* Don't attach until we have any req'd. authentication. */
3874 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3875 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3876 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3877 /* Note: this does not necessarily succeed; there
3878 * may not any proc available
3881 rxi_ChallengeOn(acall->conn);
3886 /* A data packet has been received off the interface. This packet is
3887 * appropriate to the call (the call is in the right state, etc.). This
3888 * routine can return a packet to the caller, for re-use */
3890 static struct rx_packet *
3891 rxi_ReceiveDataPacket(struct rx_call *call,
3892 struct rx_packet *np, int istack,
3893 osi_socket socket, afs_uint32 host, u_short port,
3894 int *tnop, struct rx_call **newcallp)
3896 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3901 afs_uint32 serial=0, flags=0;
3903 struct rx_packet *tnp;
3904 if (rx_stats_active)
3905 rx_atomic_inc(&rx_stats.dataPacketsRead);
3908 /* If there are no packet buffers, drop this new packet, unless we can find
3909 * packet buffers from inactive calls */
3911 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3912 MUTEX_ENTER(&rx_freePktQ_lock);
3913 rxi_NeedMorePackets = TRUE;
3914 MUTEX_EXIT(&rx_freePktQ_lock);
3915 if (rx_stats_active)
3916 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3917 rxi_calltrace(RX_TRACE_DROP, call);
3918 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3919 /* We used to clear the receive queue here, in an attempt to free
3920 * packets. However this is unsafe if the queue has received a
3921 * soft ACK for the final packet */
3922 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3928 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3929 * packet is one of several packets transmitted as a single
3930 * datagram. Do not send any soft or hard acks until all packets
3931 * in a jumbogram have been processed. Send negative acks right away.
3933 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3934 /* tnp is non-null when there are more packets in the
3935 * current jumbo gram */
3942 seq = np->header.seq;
3943 serial = np->header.serial;
3944 flags = np->header.flags;
3946 /* If the call is in an error state, send an abort message */
3948 return rxi_SendCallAbort(call, np, istack, 0);
3950 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3951 * AFS 3.5 jumbogram. */
3952 if (flags & RX_JUMBO_PACKET) {
3953 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3958 if (np->header.spare != 0) {
3959 MUTEX_ENTER(&call->conn->conn_data_lock);
3960 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3961 MUTEX_EXIT(&call->conn->conn_data_lock);
3964 /* The usual case is that this is the expected next packet */
3965 if (seq == call->rnext) {
3967 /* Check to make sure it is not a duplicate of one already queued */
3968 if (!opr_queue_IsEmpty(&call->rq)
3969 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3970 if (rx_stats_active)
3971 rx_atomic_inc(&rx_stats.dupPacketsRead);
3972 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3973 rxi_CancelDelayedAckEvent(call);
3974 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3980 /* It's the next packet. Stick it on the receive queue
3981 * for this call. Set newPackets to make sure we wake
3982 * the reader once all packets have been processed */
3983 #ifdef RX_TRACK_PACKETS
3984 np->flags |= RX_PKTFLAG_RQ;
3986 opr_queue_Prepend(&call->rq, &np->entry);
3987 #ifdef RXDEBUG_PACKET
3989 #endif /* RXDEBUG_PACKET */
3991 np = NULL; /* We can't use this anymore */
3994 /* If an ack is requested then set a flag to make sure we
3995 * send an acknowledgement for this packet */
3996 if (flags & RX_REQUEST_ACK) {
3997 ackNeeded = RX_ACK_REQUESTED;
4000 /* Keep track of whether we have received the last packet */
4001 if (flags & RX_LAST_PACKET) {
4002 call->flags |= RX_CALL_HAVE_LAST;
4006 /* Check whether we have all of the packets for this call */
4007 if (call->flags & RX_CALL_HAVE_LAST) {
4008 afs_uint32 tseq; /* temporary sequence number */
4009 struct opr_queue *cursor;
4011 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
4012 struct rx_packet *tp;
4014 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
4015 if (tseq != tp->header.seq)
4017 if (tp->header.flags & RX_LAST_PACKET) {
4018 call->flags |= RX_CALL_RECEIVE_DONE;
4025 /* Provide asynchronous notification for those who want it
4026 * (e.g. multi rx) */
4027 if (call->arrivalProc) {
4028 (*call->arrivalProc) (call, call->arrivalProcHandle,
4029 call->arrivalProcArg);
4030 call->arrivalProc = (void (*)())0;
4033 /* Update last packet received */
4036 /* If there is no server process serving this call, grab
4037 * one, if available. We only need to do this once. If a
4038 * server thread is available, this thread becomes a server
4039 * thread and the server thread becomes a listener thread. */
4041 TryAttach(call, socket, tnop, newcallp, 0);
4044 /* This is not the expected next packet. */
4046 /* Determine whether this is a new or old packet, and if it's
4047 * a new one, whether it fits into the current receive window.
4048 * Also figure out whether the packet was delivered in sequence.
4049 * We use the prev variable to determine whether the new packet
4050 * is the successor of its immediate predecessor in the
4051 * receive queue, and the missing flag to determine whether
4052 * any of this packets predecessors are missing. */
4054 afs_uint32 prev; /* "Previous packet" sequence number */
4055 struct opr_queue *cursor;
4056 int missing; /* Are any predecessors missing? */
4058 /* If the new packet's sequence number has been sent to the
4059 * application already, then this is a duplicate */
4060 if (seq < call->rnext) {
4061 if (rx_stats_active)
4062 rx_atomic_inc(&rx_stats.dupPacketsRead);
4063 rxi_CancelDelayedAckEvent(call);
4064 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4070 /* If the sequence number is greater than what can be
4071 * accomodated by the current window, then send a negative
4072 * acknowledge and drop the packet */
4073 if ((call->rnext + call->rwind) <= seq) {
4074 rxi_CancelDelayedAckEvent(call);
4075 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4082 /* Look for the packet in the queue of old received packets */
4083 prev = call->rnext - 1;
4085 for (opr_queue_Scan(&call->rq, cursor)) {
4086 struct rx_packet *tp
4087 = opr_queue_Entry(cursor, struct rx_packet, entry);
4089 /*Check for duplicate packet */
4090 if (seq == tp->header.seq) {
4091 if (rx_stats_active)
4092 rx_atomic_inc(&rx_stats.dupPacketsRead);
4093 rxi_CancelDelayedAckEvent(call);
4094 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4100 /* If we find a higher sequence packet, break out and
4101 * insert the new packet here. */
4102 if (seq < tp->header.seq)
4104 /* Check for missing packet */
4105 if (tp->header.seq != prev + 1) {
4109 prev = tp->header.seq;
4112 /* Keep track of whether we have received the last packet. */
4113 if (flags & RX_LAST_PACKET) {
4114 call->flags |= RX_CALL_HAVE_LAST;
4117 /* It's within the window: add it to the the receive queue.
4118 * tp is left by the previous loop either pointing at the
4119 * packet before which to insert the new packet, or at the
4120 * queue head if the queue is empty or the packet should be
4122 #ifdef RX_TRACK_PACKETS
4123 np->flags |= RX_PKTFLAG_RQ;
4125 #ifdef RXDEBUG_PACKET
4127 #endif /* RXDEBUG_PACKET */
4128 opr_queue_InsertBefore(cursor, &np->entry);
4132 /* Check whether we have all of the packets for this call */
4133 if ((call->flags & RX_CALL_HAVE_LAST)
4134 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4135 afs_uint32 tseq; /* temporary sequence number */
4138 for (opr_queue_Scan(&call->rq, cursor)) {
4139 struct rx_packet *tp
4140 = opr_queue_Entry(cursor, struct rx_packet, entry);
4141 if (tseq != tp->header.seq)
4143 if (tp->header.flags & RX_LAST_PACKET) {
4144 call->flags |= RX_CALL_RECEIVE_DONE;
4151 /* We need to send an ack of the packet is out of sequence,
4152 * or if an ack was requested by the peer. */
4153 if (seq != prev + 1 || missing) {
4154 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4155 } else if (flags & RX_REQUEST_ACK) {
4156 ackNeeded = RX_ACK_REQUESTED;
4159 /* Acknowledge the last packet for each call */
4160 if (flags & RX_LAST_PACKET) {
4171 * If the receiver is waiting for an iovec, fill the iovec
4172 * using the data from the receive queue */
4173 if (call->flags & RX_CALL_IOVEC_WAIT) {
4174 didHardAck = rxi_FillReadVec(call, serial);
4175 /* the call may have been aborted */
4184 /* Wakeup the reader if any */
4185 if ((call->flags & RX_CALL_READER_WAIT)
4186 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4187 || (call->iovNext >= call->iovMax)
4188 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4189 call->flags &= ~RX_CALL_READER_WAIT;
4190 #ifdef RX_ENABLE_LOCKS
4191 CV_BROADCAST(&call->cv_rq);
4193 osi_rxWakeup(&call->rq);
4199 * Send an ack when requested by the peer, or once every
4200 * rxi_SoftAckRate packets until the last packet has been
4201 * received. Always send a soft ack for the last packet in
4202 * the server's reply. */
4204 rxi_CancelDelayedAckEvent(call);
4205 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4206 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4207 rxi_CancelDelayedAckEvent(call);
4208 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4209 } else if (call->nSoftAcks) {
4210 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4211 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4213 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4214 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4215 rxi_CancelDelayedAckEvent(call);
4222 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4224 struct rx_peer *peer = conn->peer;
4226 MUTEX_ENTER(&peer->peer_lock);
4227 peer->lastReachTime = clock_Sec();
4228 MUTEX_EXIT(&peer->peer_lock);
4230 MUTEX_ENTER(&conn->conn_data_lock);
4231 if (conn->flags & RX_CONN_ATTACHWAIT) {
4234 rxi_ConnClearAttachWait(conn);
4235 MUTEX_EXIT(&conn->conn_data_lock);
4237 for (i = 0; i < RX_MAXCALLS; i++) {
4238 struct rx_call *call = conn->call[i];
4241 MUTEX_ENTER(&call->lock);
4242 /* tnop can be null if newcallp is null */
4243 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4245 MUTEX_EXIT(&call->lock);
4249 MUTEX_EXIT(&conn->conn_data_lock);
4252 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4254 rx_ack_reason(int reason)
4257 case RX_ACK_REQUESTED:
4259 case RX_ACK_DUPLICATE:
4261 case RX_ACK_OUT_OF_SEQUENCE:
4263 case RX_ACK_EXCEEDS_WINDOW:
4265 case RX_ACK_NOSPACE:
4269 case RX_ACK_PING_RESPONSE:
4282 /* The real smarts of the whole thing. */
4283 static struct rx_packet *
4284 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4287 struct rx_ackPacket *ap;
4289 struct rx_packet *tp;
4290 struct rx_connection *conn = call->conn;
4291 struct rx_peer *peer = conn->peer;
4292 struct opr_queue *cursor;
4293 struct clock now; /* Current time, for RTT calculations */
4301 int newAckCount = 0;
4302 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4303 int pktsize = 0; /* Set if we need to update the peer mtu */
4304 int conn_data_locked = 0;
4306 if (rx_stats_active)
4307 rx_atomic_inc(&rx_stats.ackPacketsRead);
4308 ap = (struct rx_ackPacket *)rx_DataOf(np);
4309 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4311 return np; /* truncated ack packet */
4313 /* depends on ack packet struct */
4314 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4315 first = ntohl(ap->firstPacket);
4316 prev = ntohl(ap->previousPacket);
4317 serial = ntohl(ap->serial);
4320 * Ignore ack packets received out of order while protecting
4321 * against peers that set the previousPacket field to a packet
4322 * serial number instead of a sequence number.
4324 if (first < call->tfirst ||
4325 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4332 if (np->header.flags & RX_SLOW_START_OK) {
4333 call->flags |= RX_CALL_SLOW_START_OK;
4336 if (ap->reason == RX_ACK_PING_RESPONSE)
4337 rxi_UpdatePeerReach(conn, call);
4339 if (conn->lastPacketSizeSeq) {
4340 MUTEX_ENTER(&conn->conn_data_lock);
4341 conn_data_locked = 1;
4342 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4343 pktsize = conn->lastPacketSize;
4344 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4347 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4348 if (!conn_data_locked) {
4349 MUTEX_ENTER(&conn->conn_data_lock);
4350 conn_data_locked = 1;
4352 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4353 /* process mtu ping ack */
4354 pktsize = conn->lastPingSize;
4355 conn->lastPingSizeSer = conn->lastPingSize = 0;
4359 if (conn_data_locked) {
4360 MUTEX_EXIT(&conn->conn_data_lock);
4361 conn_data_locked = 0;
4365 if (rxdebug_active) {
4369 len = _snprintf(msg, sizeof(msg),
4370 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4371 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4372 ntohl(ap->serial), ntohl(ap->previousPacket),
4373 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4374 ap->nAcks, ntohs(ap->bufferSpace) );
4378 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4379 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4383 OutputDebugString(msg);
4385 #else /* AFS_NT40_ENV */
4388 "RACK: reason %x previous %u seq %u serial %u first %u",
4389 ap->reason, ntohl(ap->previousPacket),
4390 (unsigned int)np->header.seq, (unsigned int)serial,
4391 ntohl(ap->firstPacket));
4394 for (offset = 0; offset < nAcks; offset++)
4395 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4400 #endif /* AFS_NT40_ENV */
4403 MUTEX_ENTER(&peer->peer_lock);
4406 * Start somewhere. Can't assume we can send what we can receive,
4407 * but we are clearly receiving.
4409 if (!peer->maxPacketSize)
4410 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4412 if (pktsize > peer->maxPacketSize) {
4413 peer->maxPacketSize = pktsize;
4414 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4415 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4416 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4417 rxi_ScheduleGrowMTUEvent(call, 1);
4422 clock_GetTime(&now);
4424 /* The transmit queue splits into 4 sections.
4426 * The first section is packets which have now been acknowledged
4427 * by a window size change in the ack. These have reached the
4428 * application layer, and may be discarded. These are packets
4429 * with sequence numbers < ap->firstPacket.
4431 * The second section is packets which have sequence numbers in
4432 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4433 * contents of the packet's ack array determines whether these
4434 * packets are acknowledged or not.
4436 * The third section is packets which fall above the range
4437 * addressed in the ack packet. These have not yet been received
4440 * The four section is packets which have not yet been transmitted.
4441 * These packets will have a header.serial of 0.
4444 /* First section - implicitly acknowledged packets that can be
4448 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4449 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4450 struct rx_packet *next;
4452 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4453 call->tfirst = tp->header.seq + 1;
4455 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4457 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4460 #ifdef RX_ENABLE_LOCKS
4461 /* XXX Hack. Because we have to release the global call lock when sending
4462 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4463 * in rxi_Start sending packets out because packets may move to the
4464 * freePacketQueue as result of being here! So we drop these packets until
4465 * we're safely out of the traversing. Really ugly!
4466 * To make it even uglier, if we're using fine grain locking, we can
4467 * set the ack bits in the packets and have rxi_Start remove the packets
4468 * when it's done transmitting.
4470 if (call->flags & RX_CALL_TQ_BUSY) {
4471 tp->flags |= RX_PKTFLAG_ACKED;
4472 call->flags |= RX_CALL_TQ_SOME_ACKED;
4474 #endif /* RX_ENABLE_LOCKS */
4476 opr_queue_Remove(&tp->entry);
4477 #ifdef RX_TRACK_PACKETS
4478 tp->flags &= ~RX_PKTFLAG_TQ;
4480 #ifdef RXDEBUG_PACKET
4482 #endif /* RXDEBUG_PACKET */
4483 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4488 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4490 /* Second section of the queue - packets for which we are receiving
4493 * Go through the explicit acks/nacks and record the results in
4494 * the waiting packets. These are packets that can't be released
4495 * yet, even with a positive acknowledge. This positive
4496 * acknowledge only means the packet has been received by the
4497 * peer, not that it will be retained long enough to be sent to
4498 * the peer's upper level. In addition, reset the transmit timers
4499 * of any missing packets (those packets that must be missing
4500 * because this packet was out of sequence) */
4502 call->nSoftAcked = 0;
4504 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4505 && tp->header.seq < first + nAcks) {
4506 /* Set the acknowledge flag per packet based on the
4507 * information in the ack packet. An acknowlegded packet can
4508 * be downgraded when the server has discarded a packet it
4509 * soacked previously, or when an ack packet is received
4510 * out of sequence. */
4511 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4512 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4514 tp->flags |= RX_PKTFLAG_ACKED;
4515 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4522 } else /* RX_ACK_TYPE_NACK */ {
4523 tp->flags &= ~RX_PKTFLAG_ACKED;
4527 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4530 /* We don't need to take any action with the 3rd or 4th section in the
4531 * queue - they're not addressed by the contents of this ACK packet.
4534 /* If the window has been extended by this acknowledge packet,
4535 * then wakeup a sender waiting in alloc for window space, or try
4536 * sending packets now, if he's been sitting on packets due to
4537 * lack of window space */
4538 if (call->tnext < (call->tfirst + call->twind)) {
4539 #ifdef RX_ENABLE_LOCKS
4540 CV_SIGNAL(&call->cv_twind);
4542 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4543 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4544 osi_rxWakeup(&call->twind);
4547 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4548 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4552 /* if the ack packet has a receivelen field hanging off it,
4553 * update our state */
4554 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4557 /* If the ack packet has a "recommended" size that is less than
4558 * what I am using now, reduce my size to match */
4559 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4560 (int)sizeof(afs_int32), &tSize);
4561 tSize = (afs_uint32) ntohl(tSize);
4562 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4564 /* Get the maximum packet size to send to this peer */
4565 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4567 tSize = (afs_uint32) ntohl(tSize);
4568 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4569 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4571 /* sanity check - peer might have restarted with different params.
4572 * If peer says "send less", dammit, send less... Peer should never
4573 * be unable to accept packets of the size that prior AFS versions would
4574 * send without asking. */
4575 if (peer->maxMTU != tSize) {
4576 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4578 peer->maxMTU = tSize;
4579 peer->MTU = MIN(tSize, peer->MTU);
4580 call->MTU = MIN(call->MTU, tSize);
4583 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4586 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4587 (int)sizeof(afs_int32), &tSize);
4588 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4589 if (tSize < call->twind) { /* smaller than our send */
4590 call->twind = tSize; /* window, we must send less... */
4591 call->ssthresh = MIN(call->twind, call->ssthresh);
4592 call->conn->twind[call->channel] = call->twind;
4595 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4596 * network MTU confused with the loopback MTU. Calculate the
4597 * maximum MTU here for use in the slow start code below.
4599 /* Did peer restart with older RX version? */
4600 if (peer->maxDgramPackets > 1) {
4601 peer->maxDgramPackets = 1;
4603 } else if (np->length >=
4604 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4607 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4608 sizeof(afs_int32), &tSize);
4609 tSize = (afs_uint32) ntohl(tSize);
4611 * As of AFS 3.5 we set the send window to match the receive window.
4613 if (tSize < call->twind) {
4614 call->twind = tSize;
4615 call->conn->twind[call->channel] = call->twind;
4616 call->ssthresh = MIN(call->twind, call->ssthresh);
4617 } else if (tSize > call->twind) {
4618 call->twind = tSize;
4619 call->conn->twind[call->channel] = call->twind;
4623 * As of AFS 3.5, a jumbogram is more than one fixed size
4624 * packet transmitted in a single UDP datagram. If the remote
4625 * MTU is smaller than our local MTU then never send a datagram
4626 * larger than the natural MTU.
4629 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4630 (int)sizeof(afs_int32), &tSize);
4631 maxDgramPackets = (afs_uint32) ntohl(tSize);
4632 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4634 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4635 if (maxDgramPackets > 1) {
4636 peer->maxDgramPackets = maxDgramPackets;
4637 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4639 peer->maxDgramPackets = 1;
4640 call->MTU = peer->natMTU;
4642 } else if (peer->maxDgramPackets > 1) {
4643 /* Restarted with lower version of RX */
4644 peer->maxDgramPackets = 1;
4646 } else if (peer->maxDgramPackets > 1
4647 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4648 /* Restarted with lower version of RX */
4649 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4650 peer->natMTU = OLD_MAX_PACKET_SIZE;
4651 peer->MTU = OLD_MAX_PACKET_SIZE;
4652 peer->maxDgramPackets = 1;
4653 peer->nDgramPackets = 1;
4655 call->MTU = OLD_MAX_PACKET_SIZE;
4660 * Calculate how many datagrams were successfully received after
4661 * the first missing packet and adjust the negative ack counter
4666 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4667 if (call->nNacks < nNacked) {
4668 call->nNacks = nNacked;
4671 call->nAcks += newAckCount;
4675 /* If the packet contained new acknowledgements, rather than just
4676 * being a duplicate of one we have previously seen, then we can restart
4679 if (newAckCount > 0)
4680 rxi_rto_packet_acked(call, istack);
4682 if (call->flags & RX_CALL_FAST_RECOVER) {
4683 if (newAckCount == 0) {
4684 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4686 call->flags &= ~RX_CALL_FAST_RECOVER;
4687 call->cwind = call->nextCwind;
4688 call->nextCwind = 0;
4691 call->nCwindAcks = 0;
4692 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4693 /* Three negative acks in a row trigger congestion recovery */
4694 call->flags |= RX_CALL_FAST_RECOVER;
4695 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4697 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4698 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4699 call->nextCwind = call->ssthresh;
4702 peer->MTU = call->MTU;
4703 peer->cwind = call->nextCwind;
4704 peer->nDgramPackets = call->nDgramPackets;
4706 call->congestSeq = peer->congestSeq;
4708 /* Reset the resend times on the packets that were nacked
4709 * so we will retransmit as soon as the window permits
4713 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4714 struct rx_packet *tp =
4715 opr_queue_Entry(cursor, struct rx_packet, entry);
4717 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4718 tp->flags &= ~RX_PKTFLAG_SENT;
4720 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4725 /* If cwind is smaller than ssthresh, then increase
4726 * the window one packet for each ack we receive (exponential
4728 * If cwind is greater than or equal to ssthresh then increase
4729 * the congestion window by one packet for each cwind acks we
4730 * receive (linear growth). */
4731 if (call->cwind < call->ssthresh) {
4733 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4734 call->nCwindAcks = 0;
4736 call->nCwindAcks += newAckCount;
4737 if (call->nCwindAcks >= call->cwind) {
4738 call->nCwindAcks = 0;
4739 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4743 * If we have received several acknowledgements in a row then
4744 * it is time to increase the size of our datagrams
4746 if ((int)call->nAcks > rx_nDgramThreshold) {
4747 if (peer->maxDgramPackets > 1) {
4748 if (call->nDgramPackets < peer->maxDgramPackets) {
4749 call->nDgramPackets++;
4751 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4752 } else if (call->MTU < peer->maxMTU) {
4753 /* don't upgrade if we can't handle it */
4754 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4755 call->MTU = peer->ifMTU;
4757 call->MTU += peer->natMTU;
4758 call->MTU = MIN(call->MTU, peer->maxMTU);
4765 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4767 /* Servers need to hold the call until all response packets have
4768 * been acknowledged. Soft acks are good enough since clients
4769 * are not allowed to clear their receive queues. */
4770 if (call->state == RX_STATE_HOLD
4771 && call->tfirst + call->nSoftAcked >= call->tnext) {
4772 call->state = RX_STATE_DALLY;
4773 rxi_ClearTransmitQueue(call, 0);
4774 rxi_CancelKeepAliveEvent(call);
4775 } else if (!opr_queue_IsEmpty(&call->tq)) {
4776 rxi_Start(call, istack);
4781 /* Received a response to a challenge packet */
4782 static struct rx_packet *
4783 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4784 struct rx_packet *np, int istack)
4788 /* Ignore the packet if we're the client */
4789 if (conn->type == RX_CLIENT_CONNECTION)
4792 /* If already authenticated, ignore the packet (it's probably a retry) */
4793 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4796 /* Otherwise, have the security object evaluate the response packet */
4797 error = RXS_CheckResponse(conn->securityObject, conn, np);
4799 /* If the response is invalid, reset the connection, sending
4800 * an abort to the peer */
4804 rxi_ConnectionError(conn, error);
4805 MUTEX_ENTER(&conn->conn_data_lock);
4806 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4807 MUTEX_EXIT(&conn->conn_data_lock);
4810 /* If the response is valid, any calls waiting to attach
4811 * servers can now do so */
4814 for (i = 0; i < RX_MAXCALLS; i++) {
4815 struct rx_call *call = conn->call[i];
4817 MUTEX_ENTER(&call->lock);
4818 if (call->state == RX_STATE_PRECALL)
4819 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4820 /* tnop can be null if newcallp is null */
4821 MUTEX_EXIT(&call->lock);
4825 /* Update the peer reachability information, just in case
4826 * some calls went into attach-wait while we were waiting
4827 * for authentication..
4829 rxi_UpdatePeerReach(conn, NULL);
4834 /* A client has received an authentication challenge: the security
4835 * object is asked to cough up a respectable response packet to send
4836 * back to the server. The server is responsible for retrying the
4837 * challenge if it fails to get a response. */
4839 static struct rx_packet *
4840 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4841 struct rx_packet *np, int istack)
4845 /* Ignore the challenge if we're the server */
4846 if (conn->type == RX_SERVER_CONNECTION)
4849 /* Ignore the challenge if the connection is otherwise idle; someone's
4850 * trying to use us as an oracle. */
4851 if (!rxi_HasActiveCalls(conn))
4854 /* Send the security object the challenge packet. It is expected to fill
4855 * in the response. */
4856 error = RXS_GetResponse(conn->securityObject, conn, np);
4858 /* If the security object is unable to return a valid response, reset the
4859 * connection and send an abort to the peer. Otherwise send the response
4860 * packet to the peer connection. */
4862 rxi_ConnectionError(conn, error);
4863 MUTEX_ENTER(&conn->conn_data_lock);
4864 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4865 MUTEX_EXIT(&conn->conn_data_lock);
4867 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4868 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4874 /* Find an available server process to service the current request in
4875 * the given call structure. If one isn't available, queue up this
4876 * call so it eventually gets one */
4878 rxi_AttachServerProc(struct rx_call *call,
4879 osi_socket socket, int *tnop,
4880 struct rx_call **newcallp)
4882 struct rx_serverQueueEntry *sq;
4883 struct rx_service *service = call->conn->service;
4886 /* May already be attached */
4887 if (call->state == RX_STATE_ACTIVE)
4890 MUTEX_ENTER(&rx_serverPool_lock);
4892 haveQuota = QuotaOK(service);
4893 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4894 /* If there are no processes available to service this call,
4895 * put the call on the incoming call queue (unless it's
4896 * already on the queue).
4898 #ifdef RX_ENABLE_LOCKS
4900 ReturnToServerPool(service);
4901 #endif /* RX_ENABLE_LOCKS */
4903 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4904 call->flags |= RX_CALL_WAIT_PROC;
4905 rx_atomic_inc(&rx_nWaiting);
4906 rx_atomic_inc(&rx_nWaited);
4907 rxi_calltrace(RX_CALL_ARRIVAL, call);
4908 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4909 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4912 sq = opr_queue_Last(&rx_idleServerQueue,
4913 struct rx_serverQueueEntry, entry);
4915 /* If hot threads are enabled, and both newcallp and sq->socketp
4916 * are non-null, then this thread will process the call, and the
4917 * idle server thread will start listening on this threads socket.
4919 opr_queue_Remove(&sq->entry);
4921 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4924 *sq->socketp = socket;
4925 clock_GetTime(&call->startTime);
4926 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4930 if (call->flags & RX_CALL_WAIT_PROC) {
4931 /* Conservative: I don't think this should happen */
4932 call->flags &= ~RX_CALL_WAIT_PROC;
4933 rx_atomic_dec(&rx_nWaiting);
4934 if (opr_queue_IsOnQueue(&call->entry)) {
4935 opr_queue_Remove(&call->entry);
4938 call->state = RX_STATE_ACTIVE;
4939 call->app.mode = RX_MODE_RECEIVING;
4940 #ifdef RX_KERNEL_TRACE
4942 int glockOwner = ISAFS_GLOCK();
4945 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4946 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4952 if (call->flags & RX_CALL_CLEARED) {
4953 /* send an ack now to start the packet flow up again */
4954 call->flags &= ~RX_CALL_CLEARED;
4955 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4957 #ifdef RX_ENABLE_LOCKS
4960 service->nRequestsRunning++;
4961 MUTEX_ENTER(&rx_quota_mutex);
4962 if (service->nRequestsRunning <= service->minProcs)
4965 MUTEX_EXIT(&rx_quota_mutex);
4969 MUTEX_EXIT(&rx_serverPool_lock);
4972 /* Delay the sending of an acknowledge event for a short while, while
4973 * a new call is being prepared (in the case of a client) or a reply
4974 * is being prepared (in the case of a server). Rather than sending
4975 * an ack packet, an ACKALL packet is sent. */
4977 rxi_AckAll(struct rx_call *call)
4979 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4981 call->flags |= RX_CALL_ACKALL_SENT;
4985 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4988 struct rx_call *call = arg1;
4989 #ifdef RX_ENABLE_LOCKS
4991 MUTEX_ENTER(&call->lock);
4992 if (event == call->delayedAckEvent)
4993 rxevent_Put(&call->delayedAckEvent);
4994 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4996 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4998 MUTEX_EXIT(&call->lock);
4999 #else /* RX_ENABLE_LOCKS */
5001 rxevent_Put(&call->delayedAckEvent);
5002 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
5003 #endif /* RX_ENABLE_LOCKS */
5006 #ifdef RX_ENABLE_LOCKS
5007 /* Set ack in all packets in transmit queue. rxi_Start will deal with
5008 * clearing them out.
5011 rxi_SetAcksInTransmitQueue(struct rx_call *call)
5013 struct opr_queue *cursor;
5016 for (opr_queue_Scan(&call->tq, cursor)) {
5018 = opr_queue_Entry(cursor, struct rx_packet, entry);
5020 p->flags |= RX_PKTFLAG_ACKED;
5025 call->flags |= RX_CALL_TQ_CLEARME;
5026 call->flags |= RX_CALL_TQ_SOME_ACKED;
5029 rxi_rto_cancel(call);
5031 call->tfirst = call->tnext;
5032 call->nSoftAcked = 0;
5034 if (call->flags & RX_CALL_FAST_RECOVER) {
5035 call->flags &= ~RX_CALL_FAST_RECOVER;
5036 call->cwind = call->nextCwind;
5037 call->nextCwind = 0;
5040 CV_SIGNAL(&call->cv_twind);
5042 #endif /* RX_ENABLE_LOCKS */
5045 * Acknowledge the whole transmit queue.
5047 * If we're running without locks, or the transmit queue isn't busy, then
5048 * we can just clear the queue now. Otherwise, we have to mark all of the
5049 * packets as acknowledged, and let rxi_Start clear it later on
5052 rxi_AckAllInTransmitQueue(struct rx_call *call)
5054 #ifdef RX_ENABLE_LOCKS
5055 if (call->flags & RX_CALL_TQ_BUSY) {
5056 rxi_SetAcksInTransmitQueue(call);
5060 rxi_ClearTransmitQueue(call, 0);
5062 /* Clear out the transmit queue for the current call (all packets have
5063 * been received by peer) */
5065 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5067 #ifdef RX_ENABLE_LOCKS
5068 struct opr_queue *cursor;
5069 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5071 for (opr_queue_Scan(&call->tq, cursor)) {
5073 = opr_queue_Entry(cursor, struct rx_packet, entry);
5075 p->flags |= RX_PKTFLAG_ACKED;
5079 call->flags |= RX_CALL_TQ_CLEARME;
5080 call->flags |= RX_CALL_TQ_SOME_ACKED;
5083 #endif /* RX_ENABLE_LOCKS */
5084 #ifdef RXDEBUG_PACKET
5086 #endif /* RXDEBUG_PACKET */
5087 rxi_FreePackets(0, &call->tq);
5088 rxi_WakeUpTransmitQueue(call);
5089 #ifdef RX_ENABLE_LOCKS
5090 call->flags &= ~RX_CALL_TQ_CLEARME;
5094 rxi_rto_cancel(call);
5095 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5096 call->nSoftAcked = 0;
5098 if (call->flags & RX_CALL_FAST_RECOVER) {
5099 call->flags &= ~RX_CALL_FAST_RECOVER;
5100 call->cwind = call->nextCwind;
5102 #ifdef RX_ENABLE_LOCKS
5103 CV_SIGNAL(&call->cv_twind);
5105 osi_rxWakeup(&call->twind);
5110 rxi_ClearReceiveQueue(struct rx_call *call)
5112 if (!opr_queue_IsEmpty(&call->rq)) {
5115 count = rxi_FreePackets(0, &call->rq);
5116 rx_packetReclaims += count;
5117 #ifdef RXDEBUG_PACKET
5119 if ( call->rqc != 0 )
5120 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5122 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5124 if (call->state == RX_STATE_PRECALL) {
5125 call->flags |= RX_CALL_CLEARED;
5129 /* Send an abort packet for the specified call */
5130 static struct rx_packet *
5131 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5132 int istack, int force)
5134 afs_int32 error, cerror;
5135 struct clock when, now;
5140 switch (call->error) {
5143 cerror = RX_CALL_TIMEOUT;
5146 cerror = call->error;
5149 /* Clients should never delay abort messages */
5150 if (rx_IsClientConn(call->conn))
5153 if (call->abortCode != cerror) {
5154 call->abortCode = cerror;
5155 call->abortCount = 0;
5158 if (force || rxi_callAbortThreshhold == 0
5159 || call->abortCount < rxi_callAbortThreshhold) {
5160 rxi_CancelDelayedAbortEvent(call);
5161 error = htonl(cerror);
5164 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5165 (char *)&error, sizeof(error), istack);
5166 } else if (!call->delayedAbortEvent) {
5167 clock_GetTime(&now);
5169 clock_Addmsec(&when, rxi_callAbortDelay);
5170 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5171 call->delayedAbortEvent =
5172 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5178 rxi_CancelDelayedAbortEvent(struct rx_call *call)
5180 if (call->delayedAbortEvent) {
5181 rxevent_Cancel(&call->delayedAbortEvent);
5182 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5186 /* Send an abort packet for the specified connection. Packet is an
5187 * optional pointer to a packet that can be used to send the abort.
5188 * Once the number of abort messages reaches the threshhold, an
5189 * event is scheduled to send the abort. Setting the force flag
5190 * overrides sending delayed abort messages.
5192 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5193 * to send the abort packet.
5196 rxi_SendConnectionAbort(struct rx_connection *conn,
5197 struct rx_packet *packet, int istack, int force)
5200 struct clock when, now;
5205 /* Clients should never delay abort messages */
5206 if (rx_IsClientConn(conn))
5209 if (force || rxi_connAbortThreshhold == 0
5210 || conn->abortCount < rxi_connAbortThreshhold) {
5212 rxevent_Cancel(&conn->delayedAbortEvent);
5213 error = htonl(conn->error);
5215 MUTEX_EXIT(&conn->conn_data_lock);
5217 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5218 RX_PACKET_TYPE_ABORT, (char *)&error,
5219 sizeof(error), istack);
5220 MUTEX_ENTER(&conn->conn_data_lock);
5221 } else if (!conn->delayedAbortEvent) {
5222 clock_GetTime(&now);
5224 clock_Addmsec(&when, rxi_connAbortDelay);
5225 conn->delayedAbortEvent =
5226 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5231 /* Associate an error all of the calls owned by a connection. Called
5232 * with error non-zero. This is only for really fatal things, like
5233 * bad authentication responses. The connection itself is set in
5234 * error at this point, so that future packets received will be
5237 rxi_ConnectionError(struct rx_connection *conn,
5243 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5245 MUTEX_ENTER(&conn->conn_data_lock);
5246 rxevent_Cancel(&conn->challengeEvent);
5247 rxevent_Cancel(&conn->natKeepAliveEvent);
5248 if (conn->checkReachEvent) {
5249 rxevent_Cancel(&conn->checkReachEvent);
5250 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5251 putConnection(conn);
5253 MUTEX_EXIT(&conn->conn_data_lock);
5254 for (i = 0; i < RX_MAXCALLS; i++) {
5255 struct rx_call *call = conn->call[i];
5257 MUTEX_ENTER(&call->lock);
5258 rxi_CallError(call, error);
5259 MUTEX_EXIT(&call->lock);
5262 conn->error = error;
5263 if (rx_stats_active)
5264 rx_atomic_inc(&rx_stats.fatalErrors);
5269 * Interrupt an in-progress call with the specified error and wakeup waiters.
5271 * @param[in] call The call to interrupt
5272 * @param[in] error The error code to send to the peer
5275 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5277 MUTEX_ENTER(&call->lock);
5278 rxi_CallError(call, error);
5279 rxi_SendCallAbort(call, NULL, 0, 1);
5280 MUTEX_EXIT(&call->lock);
5284 rxi_CallError(struct rx_call *call, afs_int32 error)
5286 MUTEX_ASSERT(&call->lock);
5287 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5289 error = call->error;
5291 #ifdef RX_ENABLE_LOCKS
5292 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5293 rxi_ResetCall(call, 0);
5296 rxi_ResetCall(call, 0);
5298 call->error = error;
5301 /* Reset various fields in a call structure, and wakeup waiting
5302 * processes. Some fields aren't changed: state & mode are not
5303 * touched (these must be set by the caller), and bufptr, nLeft, and
5304 * nFree are not reset, since these fields are manipulated by
5305 * unprotected macros, and may only be reset by non-interrupting code.
5309 rxi_ResetCall(struct rx_call *call, int newcall)
5312 struct rx_peer *peer;
5313 struct rx_packet *packet;
5315 MUTEX_ASSERT(&call->lock);
5316 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5318 /* Notify anyone who is waiting for asynchronous packet arrival */
5319 if (call->arrivalProc) {
5320 (*call->arrivalProc) (call, call->arrivalProcHandle,
5321 call->arrivalProcArg);
5322 call->arrivalProc = (void (*)())0;
5326 rxi_CancelGrowMTUEvent(call);
5328 if (call->delayedAbortEvent) {
5329 rxi_CancelDelayedAbortEvent(call);
5330 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5332 rxi_SendCallAbort(call, packet, 0, 1);
5333 rxi_FreePacket(packet);
5338 * Update the peer with the congestion information in this call
5339 * so other calls on this connection can pick up where this call
5340 * left off. If the congestion sequence numbers don't match then
5341 * another call experienced a retransmission.
5343 peer = call->conn->peer;
5344 MUTEX_ENTER(&peer->peer_lock);
5346 if (call->congestSeq == peer->congestSeq) {
5347 peer->cwind = MAX(peer->cwind, call->cwind);
5348 peer->MTU = MAX(peer->MTU, call->MTU);
5349 peer->nDgramPackets =
5350 MAX(peer->nDgramPackets, call->nDgramPackets);
5353 call->abortCode = 0;
5354 call->abortCount = 0;
5356 if (peer->maxDgramPackets > 1) {
5357 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5359 call->MTU = peer->MTU;
5361 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5362 call->ssthresh = rx_maxSendWindow;
5363 call->nDgramPackets = peer->nDgramPackets;
5364 call->congestSeq = peer->congestSeq;
5365 call->rtt = peer->rtt;
5366 call->rtt_dev = peer->rtt_dev;
5367 clock_Zero(&call->rto);
5368 clock_Addmsec(&call->rto,
5369 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5370 MUTEX_EXIT(&peer->peer_lock);
5372 flags = call->flags;
5373 rxi_WaitforTQBusy(call);
5375 rxi_ClearTransmitQueue(call, 1);
5376 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5377 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5381 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5382 /* The call channel is still busy; resetting the call doesn't change
5383 * that. However, if 'newcall' is set, we are processing a call
5384 * structure that has either been recycled from the free list, or has
5385 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5386 * 'newcall' is set, since it describes a completely different call
5387 * channel which we do not care about. */
5388 call->flags |= RX_CALL_PEER_BUSY;
5391 rxi_ClearReceiveQueue(call);
5392 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5396 call->twind = call->conn->twind[call->channel];
5397 call->rwind = call->conn->rwind[call->channel];
5398 call->nSoftAcked = 0;
5399 call->nextCwind = 0;
5402 call->nCwindAcks = 0;
5403 call->nSoftAcks = 0;
5404 call->nHardAcks = 0;
5406 call->tfirst = call->rnext = call->tnext = 1;
5409 call->lastAcked = 0;
5410 call->localStatus = call->remoteStatus = 0;
5412 if (flags & RX_CALL_READER_WAIT) {
5413 #ifdef RX_ENABLE_LOCKS
5414 CV_BROADCAST(&call->cv_rq);
5416 osi_rxWakeup(&call->rq);
5419 if (flags & RX_CALL_WAIT_PACKETS) {
5420 MUTEX_ENTER(&rx_freePktQ_lock);
5421 rxi_PacketsUnWait(); /* XXX */
5422 MUTEX_EXIT(&rx_freePktQ_lock);
5424 #ifdef RX_ENABLE_LOCKS
5425 CV_SIGNAL(&call->cv_twind);
5427 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5428 osi_rxWakeup(&call->twind);
5431 if (flags & RX_CALL_WAIT_PROC) {
5432 rx_atomic_dec(&rx_nWaiting);
5434 #ifdef RX_ENABLE_LOCKS
5435 /* The following ensures that we don't mess with any queue while some
5436 * other thread might also be doing so. The call_queue_lock field is
5437 * is only modified under the call lock. If the call is in the process
5438 * of being removed from a queue, the call is not locked until the
5439 * the queue lock is dropped and only then is the call_queue_lock field
5440 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5441 * Note that any other routine which removes a call from a queue has to
5442 * obtain the queue lock before examing the queue and removing the call.
5444 if (call->call_queue_lock) {
5445 MUTEX_ENTER(call->call_queue_lock);
5446 if (opr_queue_IsOnQueue(&call->entry)) {
5447 opr_queue_Remove(&call->entry);
5449 MUTEX_EXIT(call->call_queue_lock);
5450 CLEAR_CALL_QUEUE_LOCK(call);
5452 #else /* RX_ENABLE_LOCKS */
5453 if (opr_queue_IsOnQueue(&call->entry)) {
5454 opr_queue_Remove(&call->entry);
5456 #endif /* RX_ENABLE_LOCKS */
5458 rxi_CancelKeepAliveEvent(call);
5459 rxi_CancelDelayedAckEvent(call);
5462 /* Send an acknowledge for the indicated packet (seq,serial) of the
5463 * indicated call, for the indicated reason (reason). This
5464 * acknowledge will specifically acknowledge receiving the packet, and
5465 * will also specify which other packets for this call have been
5466 * received. This routine returns the packet that was used to the
5467 * caller. The caller is responsible for freeing it or re-using it.
5468 * This acknowledgement also returns the highest sequence number
5469 * actually read out by the higher level to the sender; the sender
5470 * promises to keep around packets that have not been read by the
5471 * higher level yet (unless, of course, the sender decides to abort
5472 * the call altogether). Any of p, seq, serial, pflags, or reason may
5473 * be set to zero without ill effect. That is, if they are zero, they
5474 * will not convey any information.
5475 * NOW there is a trailer field, after the ack where it will safely be
5476 * ignored by mundanes, which indicates the maximum size packet this
5477 * host can swallow. */
5479 struct rx_packet *optionalPacket; use to send ack (or null)
5480 int seq; Sequence number of the packet we are acking
5481 int serial; Serial number of the packet
5482 int pflags; Flags field from packet header
5483 int reason; Reason an acknowledge was prompted
5487 rxi_SendAck(struct rx_call *call,
5488 struct rx_packet *optionalPacket, int serial, int reason,
5491 struct rx_ackPacket *ap;
5492 struct rx_packet *p;
5493 struct opr_queue *cursor;
5496 afs_uint32 padbytes = 0;
5497 #ifdef RX_ENABLE_TSFPQ
5498 struct rx_ts_info_t * rx_ts_info;
5502 * Open the receive window once a thread starts reading packets
5504 if (call->rnext > 1) {
5505 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5508 /* Don't attempt to grow MTU if this is a critical ping */
5509 if (reason == RX_ACK_MTU) {
5510 /* keep track of per-call attempts, if we're over max, do in small
5511 * otherwise in larger? set a size to increment by, decrease
5514 if (call->conn->peer->maxPacketSize &&
5515 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5517 padbytes = call->conn->peer->maxPacketSize+16;
5519 padbytes = call->conn->peer->maxMTU + 128;
5521 /* do always try a minimum size ping */
5522 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5524 /* subtract the ack payload */
5525 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5526 reason = RX_ACK_PING;
5529 call->nHardAcks = 0;
5530 call->nSoftAcks = 0;
5531 if (call->rnext > call->lastAcked)
5532 call->lastAcked = call->rnext;
5536 rx_computelen(p, p->length); /* reset length, you never know */
5537 } /* where that's been... */
5538 #ifdef RX_ENABLE_TSFPQ
5540 RX_TS_INFO_GET(rx_ts_info);
5541 if ((p = rx_ts_info->local_special_packet)) {
5542 rx_computelen(p, p->length);
5543 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5544 rx_ts_info->local_special_packet = p;
5545 } else { /* We won't send the ack, but don't panic. */
5546 return optionalPacket;
5550 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5551 /* We won't send the ack, but don't panic. */
5552 return optionalPacket;
5557 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5560 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5561 #ifndef RX_ENABLE_TSFPQ
5562 if (!optionalPacket)
5565 return optionalPacket;
5567 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5568 if (rx_Contiguous(p) < templ) {
5569 #ifndef RX_ENABLE_TSFPQ
5570 if (!optionalPacket)
5573 return optionalPacket;
5578 /* MTUXXX failing to send an ack is very serious. We should */
5579 /* try as hard as possible to send even a partial ack; it's */
5580 /* better than nothing. */
5581 ap = (struct rx_ackPacket *)rx_DataOf(p);
5582 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5583 ap->reason = reason;
5585 /* The skew computation used to be bogus, I think it's better now. */
5586 /* We should start paying attention to skew. XXX */
5587 ap->serial = htonl(serial);
5588 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5591 * First packet not yet forwarded to reader. When ACKALL has been
5592 * sent the peer has been told that all received packets will be
5593 * delivered to the reader. The value 'rnext' is used internally
5594 * to refer to the next packet in the receive queue that must be
5595 * delivered to the reader. From the perspective of the peer it
5596 * already has so report the last sequence number plus one if there
5597 * are packets in the receive queue awaiting processing.
5599 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5600 !opr_queue_IsEmpty(&call->rq)) {
5601 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5603 ap->firstPacket = htonl(call->rnext);
5605 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5607 /* No fear of running out of ack packet here because there can only
5608 * be at most one window full of unacknowledged packets. The window
5609 * size must be constrained to be less than the maximum ack size,
5610 * of course. Also, an ack should always fit into a single packet
5611 * -- it should not ever be fragmented. */
5613 for (opr_queue_Scan(&call->rq, cursor)) {
5614 struct rx_packet *rqp
5615 = opr_queue_Entry(cursor, struct rx_packet, entry);
5617 if (!rqp || !call->rq.next
5618 || (rqp->header.seq > (call->rnext + call->rwind))) {
5619 #ifndef RX_ENABLE_TSFPQ
5620 if (!optionalPacket)
5623 rxi_CallError(call, RX_CALL_DEAD);
5624 return optionalPacket;
5627 while (rqp->header.seq > call->rnext + offset)
5628 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5629 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5631 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5632 #ifndef RX_ENABLE_TSFPQ
5633 if (!optionalPacket)
5636 rxi_CallError(call, RX_CALL_DEAD);
5637 return optionalPacket;
5643 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5645 /* these are new for AFS 3.3 */
5646 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5647 templ = htonl(templ);
5648 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5649 templ = htonl(call->conn->peer->ifMTU);
5650 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5651 sizeof(afs_int32), &templ);
5653 /* new for AFS 3.4 */
5654 templ = htonl(call->rwind);
5655 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5656 sizeof(afs_int32), &templ);
5658 /* new for AFS 3.5 */
5659 templ = htonl(call->conn->peer->ifDgramPackets);
5660 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5661 sizeof(afs_int32), &templ);
5663 p->header.serviceId = call->conn->serviceId;
5664 p->header.cid = (call->conn->cid | call->channel);
5665 p->header.callNumber = *call->callNumber;
5667 p->header.securityIndex = call->conn->securityIndex;
5668 p->header.epoch = call->conn->epoch;
5669 p->header.type = RX_PACKET_TYPE_ACK;
5670 p->header.flags = RX_SLOW_START_OK;
5671 if (reason == RX_ACK_PING) {
5672 p->header.flags |= RX_REQUEST_ACK;
5674 p->length = padbytes +
5675 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5678 /* not fast but we can potentially use this if truncated
5679 * fragments are delivered to figure out the mtu.
5681 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5682 sizeof(afs_int32), sizeof(afs_int32),
5686 if (call->conn->type == RX_CLIENT_CONNECTION)
5687 p->header.flags |= RX_CLIENT_INITIATED;
5691 if (rxdebug_active) {
5695 len = _snprintf(msg, sizeof(msg),
5696 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5697 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5698 ntohl(ap->serial), ntohl(ap->previousPacket),
5699 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5700 ap->nAcks, ntohs(ap->bufferSpace) );
5704 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5705 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5709 OutputDebugString(msg);
5711 #else /* AFS_NT40_ENV */
5713 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5714 ap->reason, ntohl(ap->previousPacket),
5715 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5717 for (offset = 0; offset < ap->nAcks; offset++)
5718 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5723 #endif /* AFS_NT40_ENV */
5726 int i, nbytes = p->length;
5728 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5729 if (nbytes <= p->wirevec[i].iov_len) {
5732 savelen = p->wirevec[i].iov_len;
5734 p->wirevec[i].iov_len = nbytes;
5736 rxi_Send(call, p, istack);
5737 p->wirevec[i].iov_len = savelen;
5741 nbytes -= p->wirevec[i].iov_len;
5744 if (rx_stats_active)
5745 rx_atomic_inc(&rx_stats.ackPacketsSent);
5746 #ifndef RX_ENABLE_TSFPQ
5747 if (!optionalPacket)
5750 return optionalPacket; /* Return packet for re-use by caller */
5754 struct rx_packet **list;
5759 /* Send all of the packets in the list in single datagram */
5761 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5762 int istack, int moreFlag)
5768 struct rx_connection *conn = call->conn;
5769 struct rx_peer *peer = conn->peer;
5771 MUTEX_ENTER(&peer->peer_lock);
5772 peer->nSent += xmit->len;
5773 if (xmit->resending)
5774 peer->reSends += xmit->len;
5775 MUTEX_EXIT(&peer->peer_lock);
5777 if (rx_stats_active) {
5778 if (xmit->resending)
5779 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5781 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5784 clock_GetTime(&now);
5786 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5790 /* Set the packet flags and schedule the resend events */
5791 /* Only request an ack for the last packet in the list */
5792 for (i = 0; i < xmit->len; i++) {
5793 struct rx_packet *packet = xmit->list[i];
5795 /* Record the time sent */
5796 packet->timeSent = now;
5797 packet->flags |= RX_PKTFLAG_SENT;
5799 /* Ask for an ack on retransmitted packets, on every other packet
5800 * if the peer doesn't support slow start. Ask for an ack on every
5801 * packet until the congestion window reaches the ack rate. */
5802 if (packet->header.serial) {
5805 packet->firstSent = now;
5806 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5807 || (!(call->flags & RX_CALL_SLOW_START_OK)
5808 && (packet->header.seq & 1)))) {
5813 /* Tag this packet as not being the last in this group,
5814 * for the receiver's benefit */
5815 if (i < xmit->len - 1 || moreFlag) {
5816 packet->header.flags |= RX_MORE_PACKETS;
5821 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5824 /* Since we're about to send a data packet to the peer, it's
5825 * safe to nuke any scheduled end-of-packets ack */
5826 rxi_CancelDelayedAckEvent(call);
5828 MUTEX_EXIT(&call->lock);
5829 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5830 if (xmit->len > 1) {
5831 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5833 rxi_SendPacket(call, conn, xmit->list[0], istack);
5835 MUTEX_ENTER(&call->lock);
5836 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5838 /* Tell the RTO calculation engine that we have sent a packet, and
5839 * if it was the last one */
5840 rxi_rto_packet_sent(call, lastPacket, istack);
5842 /* Update last send time for this call (for keep-alive
5843 * processing), and for the connection (so that we can discover
5844 * idle connections) */
5845 conn->lastSendTime = call->lastSendTime = clock_Sec();
5846 /* Let a set of retransmits trigger an idle timeout */
5847 if (!xmit->resending)
5848 call->lastSendData = call->lastSendTime;
5851 /* When sending packets we need to follow these rules:
5852 * 1. Never send more than maxDgramPackets in a jumbogram.
5853 * 2. Never send a packet with more than two iovecs in a jumbogram.
5854 * 3. Never send a retransmitted packet in a jumbogram.
5855 * 4. Never send more than cwind/4 packets in a jumbogram
5856 * We always keep the last list we should have sent so we
5857 * can set the RX_MORE_PACKETS flags correctly.
5861 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5866 struct xmitlist working;
5867 struct xmitlist last;
5869 struct rx_peer *peer = call->conn->peer;
5870 int morePackets = 0;
5872 memset(&last, 0, sizeof(struct xmitlist));
5873 working.list = &list[0];
5875 working.resending = 0;
5877 recovery = call->flags & RX_CALL_FAST_RECOVER;
5879 for (i = 0; i < len; i++) {
5880 /* Does the current packet force us to flush the current list? */
5882 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5883 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5885 /* This sends the 'last' list and then rolls the current working
5886 * set into the 'last' one, and resets the working set */
5889 rxi_SendList(call, &last, istack, 1);
5890 /* If the call enters an error state stop sending, or if
5891 * we entered congestion recovery mode, stop sending */
5893 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5898 working.resending = 0;
5899 working.list = &list[i];
5901 /* Add the current packet to the list if it hasn't been acked.
5902 * Otherwise adjust the list pointer to skip the current packet. */
5903 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5906 if (list[i]->header.serial)
5907 working.resending = 1;
5909 /* Do we need to flush the list? */
5910 if (working.len >= (int)peer->maxDgramPackets
5911 || working.len >= (int)call->nDgramPackets
5912 || working.len >= (int)call->cwind
5913 || list[i]->header.serial
5914 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5916 rxi_SendList(call, &last, istack, 1);
5917 /* If the call enters an error state stop sending, or if
5918 * we entered congestion recovery mode, stop sending */
5920 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5925 working.resending = 0;
5926 working.list = &list[i + 1];
5929 if (working.len != 0) {
5930 osi_Panic("rxi_SendList error");
5932 working.list = &list[i + 1];
5936 /* Send the whole list when the call is in receive mode, when
5937 * the call is in eof mode, when we are in fast recovery mode,
5938 * and when we have the last packet */
5939 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5940 * the listener or event threads
5942 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5943 || (call->flags & RX_CALL_FLUSH)
5944 || (call->flags & RX_CALL_FAST_RECOVER)) {
5945 /* Check for the case where the current list contains
5946 * an acked packet. Since we always send retransmissions
5947 * in a separate packet, we only need to check the first
5948 * packet in the list */
5949 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5953 rxi_SendList(call, &last, istack, morePackets);
5954 /* If the call enters an error state stop sending, or if
5955 * we entered congestion recovery mode, stop sending */
5957 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5961 rxi_SendList(call, &working, istack, 0);
5963 } else if (last.len > 0) {
5964 rxi_SendList(call, &last, istack, 0);
5965 /* Packets which are in 'working' are not sent by this call */
5970 * Check if the peer for the given call is known to be dead
5972 * If the call's peer appears dead (it has encountered fatal network errors
5973 * since the call started) the call is killed with RX_CALL_DEAD if the call
5974 * is active. Otherwise, we do nothing.
5976 * @param[in] call The call to check
5979 * @retval 0 The call is fine, and we haven't done anything to the call
5980 * @retval nonzero The call's peer appears dead, and the call has been
5981 * terminated if it was active
5983 * @pre call->lock must be locked
5986 rxi_CheckPeerDead(struct rx_call *call)
5988 #ifdef AFS_RXERRQ_ENV
5991 if (call->state == RX_STATE_DALLY) {
5995 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5996 if (call->neterr_gen < peererrs) {
5997 /* we have received network errors since this call started; kill
5999 if (call->state == RX_STATE_ACTIVE) {
6000 rxi_CallError(call, RX_CALL_DEAD);
6004 if (call->neterr_gen > peererrs) {
6005 /* someone has reset the number of peer errors; set the call error gen
6006 * so we can detect if more errors are encountered */
6007 call->neterr_gen = peererrs;
6014 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
6016 struct rx_call *call = arg0;
6017 struct rx_peer *peer;
6018 struct opr_queue *cursor;
6019 struct clock maxTimeout = { 60, 0 };
6021 MUTEX_ENTER(&call->lock);
6023 peer = call->conn->peer;
6025 /* Make sure that the event pointer is removed from the call
6026 * structure, since there is no longer a per-call retransmission
6028 if (event == call->resendEvent) {
6029 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
6030 rxevent_Put(&call->resendEvent);
6033 rxi_CheckPeerDead(call);
6035 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
6036 rxi_CheckBusy(call);
6039 if (opr_queue_IsEmpty(&call->tq)) {
6040 /* Nothing to do. This means that we've been raced, and that an
6041 * ACK has come in between when we were triggered, and when we
6042 * actually got to run. */
6046 /* We're in loss recovery */
6047 call->flags |= RX_CALL_FAST_RECOVER;
6049 /* Mark all of the pending packets in the queue as being lost */
6050 for (opr_queue_Scan(&call->tq, cursor)) {
6051 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
6052 if (!(p->flags & RX_PKTFLAG_ACKED))
6053 p->flags &= ~RX_PKTFLAG_SENT;
6056 /* We're resending, so we double the timeout of the call. This will be
6057 * dropped back down by the first successful ACK that we receive.
6059 * We apply a maximum value here of 60 seconds
6061 clock_Add(&call->rto, &call->rto);
6062 if (clock_Gt(&call->rto, &maxTimeout))
6063 call->rto = maxTimeout;
6065 /* Packet loss is most likely due to congestion, so drop our window size
6066 * and start again from the beginning */
6067 if (peer->maxDgramPackets >1) {
6068 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6069 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6071 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6072 call->nDgramPackets = 1;
6074 call->nextCwind = 1;
6077 MUTEX_ENTER(&peer->peer_lock);
6078 peer->MTU = call->MTU;
6079 peer->cwind = call->cwind;
6080 peer->nDgramPackets = 1;
6082 call->congestSeq = peer->congestSeq;
6083 MUTEX_EXIT(&peer->peer_lock);
6085 rxi_Start(call, istack);
6088 MUTEX_EXIT(&call->lock);
6091 /* This routine is called when new packets are readied for
6092 * transmission and when retransmission may be necessary, or when the
6093 * transmission window or burst count are favourable. This should be
6094 * better optimized for new packets, the usual case, now that we've
6095 * got rid of queues of send packets. XXXXXXXXXXX */
6097 rxi_Start(struct rx_call *call, int istack)
6099 struct opr_queue *cursor;
6100 #ifdef RX_ENABLE_LOCKS
6101 struct opr_queue *store;
6107 #ifdef RX_ENABLE_LOCKS
6108 if (rx_stats_active)
6109 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6114 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6115 /* Send (or resend) any packets that need it, subject to
6116 * window restrictions and congestion burst control
6117 * restrictions. Ask for an ack on the last packet sent in
6118 * this burst. For now, we're relying upon the window being
6119 * considerably bigger than the largest number of packets that
6120 * are typically sent at once by one initial call to
6121 * rxi_Start. This is probably bogus (perhaps we should ask
6122 * for an ack when we're half way through the current
6123 * window?). Also, for non file transfer applications, this
6124 * may end up asking for an ack for every packet. Bogus. XXXX
6127 * But check whether we're here recursively, and let the other guy
6130 #ifdef RX_ENABLE_LOCKS
6131 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6132 call->flags |= RX_CALL_TQ_BUSY;
6134 #endif /* RX_ENABLE_LOCKS */
6136 #ifdef RX_ENABLE_LOCKS
6137 call->flags &= ~RX_CALL_NEED_START;
6138 #endif /* RX_ENABLE_LOCKS */
6140 maxXmitPackets = MIN(call->twind, call->cwind);
6141 for (opr_queue_Scan(&call->tq, cursor)) {
6143 = opr_queue_Entry(cursor, struct rx_packet, entry);
6145 if (p->flags & RX_PKTFLAG_ACKED) {
6146 /* Since we may block, don't trust this */
6147 if (rx_stats_active)
6148 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6149 continue; /* Ignore this packet if it has been acknowledged */
6152 /* Turn off all flags except these ones, which are the same
6153 * on each transmission */
6154 p->header.flags &= RX_PRESET_FLAGS;
6156 if (p->header.seq >=
6157 call->tfirst + MIN((int)call->twind,
6158 (int)(call->nSoftAcked +
6160 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6161 /* Note: if we're waiting for more window space, we can
6162 * still send retransmits; hence we don't return here, but
6163 * break out to schedule a retransmit event */
6164 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6165 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6170 /* Transmit the packet if it needs to be sent. */
6171 if (!(p->flags & RX_PKTFLAG_SENT)) {
6172 if (nXmitPackets == maxXmitPackets) {
6173 rxi_SendXmitList(call, call->xmitList,
6174 nXmitPackets, istack);
6177 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6178 *(call->callNumber), p));
6179 call->xmitList[nXmitPackets++] = p;
6181 } /* end of the queue_Scan */
6183 /* xmitList now hold pointers to all of the packets that are
6184 * ready to send. Now we loop to send the packets */
6185 if (nXmitPackets > 0) {
6186 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6190 #ifdef RX_ENABLE_LOCKS
6192 /* We went into the error state while sending packets. Now is
6193 * the time to reset the call. This will also inform the using
6194 * process that the call is in an error state.
6196 if (rx_stats_active)
6197 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6198 call->flags &= ~RX_CALL_TQ_BUSY;
6199 rxi_WakeUpTransmitQueue(call);
6200 rxi_CallError(call, call->error);
6204 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6206 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6207 /* Some packets have received acks. If they all have, we can clear
6208 * the transmit queue.
6211 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6213 = opr_queue_Entry(cursor, struct rx_packet, entry);
6215 if (p->header.seq < call->tfirst
6216 && (p->flags & RX_PKTFLAG_ACKED)) {
6217 opr_queue_Remove(&p->entry);
6218 #ifdef RX_TRACK_PACKETS
6219 p->flags &= ~RX_PKTFLAG_TQ;
6221 #ifdef RXDEBUG_PACKET
6229 call->flags |= RX_CALL_TQ_CLEARME;
6231 if (call->flags & RX_CALL_TQ_CLEARME)
6232 rxi_ClearTransmitQueue(call, 1);
6233 } while (call->flags & RX_CALL_NEED_START);
6235 * TQ references no longer protected by this flag; they must remain
6236 * protected by the call lock.
6238 call->flags &= ~RX_CALL_TQ_BUSY;
6239 rxi_WakeUpTransmitQueue(call);
6241 call->flags |= RX_CALL_NEED_START;
6243 #endif /* RX_ENABLE_LOCKS */
6245 rxi_rto_cancel(call);
6249 /* Also adjusts the keep alive parameters for the call, to reflect
6250 * that we have just sent a packet (so keep alives aren't sent
6253 rxi_Send(struct rx_call *call, struct rx_packet *p,
6256 struct rx_connection *conn = call->conn;
6258 /* Stamp each packet with the user supplied status */
6259 p->header.userStatus = call->localStatus;
6261 /* Allow the security object controlling this call's security to
6262 * make any last-minute changes to the packet */
6263 RXS_SendPacket(conn->securityObject, call, p);
6265 /* Since we're about to send SOME sort of packet to the peer, it's
6266 * safe to nuke any scheduled end-of-packets ack */
6267 rxi_CancelDelayedAckEvent(call);
6269 /* Actually send the packet, filling in more connection-specific fields */
6270 MUTEX_EXIT(&call->lock);
6271 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6272 rxi_SendPacket(call, conn, p, istack);
6273 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6274 MUTEX_ENTER(&call->lock);
6276 /* Update last send time for this call (for keep-alive
6277 * processing), and for the connection (so that we can discover
6278 * idle connections) */
6279 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6280 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6281 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6283 conn->lastSendTime = call->lastSendTime = clock_Sec();
6284 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6285 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6286 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6287 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6288 RX_ACK_PING_RESPONSE)))
6289 call->lastSendData = call->lastSendTime;
6293 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6294 * that things are fine. Also called periodically to guarantee that nothing
6295 * falls through the cracks (e.g. (error + dally) connections have keepalive
6296 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6298 * haveCTLock Set if calling from rxi_ReapConnections
6301 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6303 struct rx_connection *conn = call->conn;
6305 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6306 afs_uint32 fudgeFactor;
6309 int idle_timeout = 0;
6310 afs_int32 clock_diff = 0;
6312 if (rxi_CheckPeerDead(call)) {
6318 /* Large swings in the clock can have a significant impact on
6319 * the performance of RX call processing. Forward clock shifts
6320 * will result in premature event triggering or timeouts.
6321 * Backward shifts can result in calls not completing until
6322 * the clock catches up with the original start clock value.
6324 * If a backward clock shift of more than five minutes is noticed,
6325 * just fail the call.
6327 if (now < call->lastSendTime)
6328 clock_diff = call->lastSendTime - now;
6329 if (now < call->startWait)
6330 clock_diff = MAX(clock_diff, call->startWait - now);
6331 if (now < call->lastReceiveTime)
6332 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6333 if (clock_diff > 5 * 60)
6335 if (call->state == RX_STATE_ACTIVE)
6336 rxi_CallError(call, RX_CALL_TIMEOUT);
6340 #ifdef RX_ENABLE_LOCKS
6341 if (call->flags & RX_CALL_TQ_BUSY) {
6342 /* Call is active and will be reset by rxi_Start if it's
6343 * in an error state.
6348 /* RTT + 8*MDEV, rounded up to the next second. */
6349 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6350 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6352 deadTime = conn->secondsUntilDead + fudgeFactor;
6353 /* These are computed to the second (+- 1 second). But that's
6354 * good enough for these values, which should be a significant
6355 * number of seconds. */
6356 if (now > (call->lastReceiveTime + deadTime)) {
6357 if (call->state == RX_STATE_ACTIVE) {
6358 #ifdef AFS_ADAPT_PMTU
6359 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6361 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6362 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6363 ip_stack_t *ipst = ns->netstack_ip;
6365 ire = ire_cache_lookup(conn->peer->host
6366 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6368 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6370 # if defined(GLOBAL_NETSTACKID)
6377 if (ire && ire->ire_max_frag > 0)
6378 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6380 # if defined(GLOBAL_NETSTACKID)
6384 #endif /* AFS_ADAPT_PMTU */
6385 cerror = RX_CALL_DEAD;
6388 #ifdef RX_ENABLE_LOCKS
6389 /* Cancel pending events */
6390 rxi_CancelDelayedAckEvent(call);
6391 rxi_rto_cancel(call);
6392 rxi_CancelKeepAliveEvent(call);
6393 rxi_CancelGrowMTUEvent(call);
6394 MUTEX_ENTER(&rx_refcnt_mutex);
6395 /* if rxi_FreeCall returns 1 it has freed the call */
6396 if (call->refCount == 0 &&
6397 rxi_FreeCall(call, haveCTLock))
6399 MUTEX_EXIT(&rx_refcnt_mutex);
6402 MUTEX_EXIT(&rx_refcnt_mutex);
6404 #else /* RX_ENABLE_LOCKS */
6405 rxi_FreeCall(call, 0);
6407 #endif /* RX_ENABLE_LOCKS */
6409 /* Non-active calls are destroyed if they are not responding
6410 * to pings; active calls are simply flagged in error, so the
6411 * attached process can die reasonably gracefully. */
6414 if (conn->idleDeadDetection) {
6415 if (conn->idleDeadTime) {
6416 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6420 /* see if we have a non-activity timeout */
6421 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6422 (call->flags & RX_CALL_READER_WAIT)) {
6423 if (call->state == RX_STATE_ACTIVE) {
6424 cerror = RX_CALL_TIMEOUT;
6429 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6430 if (call->state == RX_STATE_ACTIVE) {
6431 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6439 if (conn->hardDeadTime) {
6440 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6443 /* see if we have a hard timeout */
6445 && (now > (hardDeadTime + call->startTime.sec))) {
6446 if (call->state == RX_STATE_ACTIVE)
6447 rxi_CallError(call, RX_CALL_TIMEOUT);
6452 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6453 call->lastReceiveTime) {
6454 int oldMTU = conn->peer->ifMTU;
6456 /* if we thought we could send more, perhaps things got worse */
6457 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6458 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6459 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6460 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6462 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6464 /* minimum capped in SetPeerMtu */
6465 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6468 conn->lastPacketSize = 0;
6470 /* needed so ResetCall doesn't clobber us. */
6471 call->MTU = conn->peer->ifMTU;
6473 /* if we never succeeded, let the error pass out as-is */
6474 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6475 cerror = conn->msgsizeRetryErr;
6478 rxi_CallError(call, cerror);
6483 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6484 void *dummy, int dummy2)
6486 struct rx_connection *conn = arg1;
6487 struct rx_header theader;
6488 char tbuffer[1 + sizeof(struct rx_header)];
6489 struct sockaddr_in taddr;
6492 struct iovec tmpiov[2];
6495 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6498 tp = &tbuffer[sizeof(struct rx_header)];
6499 taddr.sin_family = AF_INET;
6500 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6501 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6502 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6503 taddr.sin_len = sizeof(struct sockaddr_in);
6505 memset(&theader, 0, sizeof(theader));
6506 theader.epoch = htonl(999);
6508 theader.callNumber = 0;
6511 theader.type = RX_PACKET_TYPE_VERSION;
6512 theader.flags = RX_LAST_PACKET;
6513 theader.serviceId = 0;
6515 memcpy(tbuffer, &theader, sizeof(theader));
6516 memcpy(tp, &a, sizeof(a));
6517 tmpiov[0].iov_base = tbuffer;
6518 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6520 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6522 MUTEX_ENTER(&conn->conn_data_lock);
6523 MUTEX_ENTER(&rx_refcnt_mutex);
6524 /* Only reschedule ourselves if the connection would not be destroyed */
6525 if (conn->refCount <= 1) {
6526 rxevent_Put(&conn->natKeepAliveEvent);
6527 MUTEX_EXIT(&rx_refcnt_mutex);
6528 MUTEX_EXIT(&conn->conn_data_lock);
6529 rx_DestroyConnection(conn); /* drop the reference for this */
6531 conn->refCount--; /* drop the reference for this */
6532 MUTEX_EXIT(&rx_refcnt_mutex);
6533 rxevent_Put(&conn->natKeepAliveEvent);
6534 rxi_ScheduleNatKeepAliveEvent(conn);
6535 MUTEX_EXIT(&conn->conn_data_lock);
6540 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6542 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6543 struct clock when, now;
6544 clock_GetTime(&now);
6546 when.sec += conn->secondsUntilNatPing;
6547 MUTEX_ENTER(&rx_refcnt_mutex);
6548 conn->refCount++; /* hold a reference for this */
6549 MUTEX_EXIT(&rx_refcnt_mutex);
6550 conn->natKeepAliveEvent =
6551 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6556 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6558 MUTEX_ENTER(&conn->conn_data_lock);
6559 conn->secondsUntilNatPing = seconds;
6561 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6562 rxi_ScheduleNatKeepAliveEvent(conn);
6564 conn->flags |= RX_CONN_NAT_PING;
6566 MUTEX_EXIT(&conn->conn_data_lock);
6569 /* When a call is in progress, this routine is called occasionally to
6570 * make sure that some traffic has arrived (or been sent to) the peer.
6571 * If nothing has arrived in a reasonable amount of time, the call is
6572 * declared dead; if nothing has been sent for a while, we send a
6573 * keep-alive packet (if we're actually trying to keep the call alive)
6576 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6579 struct rx_call *call = arg1;
6580 struct rx_connection *conn;
6583 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6584 MUTEX_ENTER(&call->lock);
6586 if (event == call->keepAliveEvent)
6587 rxevent_Put(&call->keepAliveEvent);
6591 if (rxi_CheckCall(call, 0)) {
6592 MUTEX_EXIT(&call->lock);
6596 /* Don't try to keep alive dallying calls */
6597 if (call->state == RX_STATE_DALLY) {
6598 MUTEX_EXIT(&call->lock);
6603 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6604 /* Don't try to send keepalives if there is unacknowledged data */
6605 /* the rexmit code should be good enough, this little hack
6606 * doesn't quite work XXX */
6607 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6609 rxi_ScheduleKeepAliveEvent(call);
6610 MUTEX_EXIT(&call->lock);
6613 /* Does what's on the nameplate. */
6615 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6617 struct rx_call *call = arg1;
6618 struct rx_connection *conn;
6620 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6621 MUTEX_ENTER(&call->lock);
6623 if (event == call->growMTUEvent)
6624 rxevent_Put(&call->growMTUEvent);
6626 if (rxi_CheckCall(call, 0)) {
6627 MUTEX_EXIT(&call->lock);
6631 /* Don't bother with dallying calls */
6632 if (call->state == RX_STATE_DALLY) {
6633 MUTEX_EXIT(&call->lock);
6640 * keep being scheduled, just don't do anything if we're at peak,
6641 * or we're not set up to be properly handled (idle timeout required)
6643 if ((conn->peer->maxPacketSize != 0) &&
6644 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6645 conn->idleDeadDetection)
6646 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6647 rxi_ScheduleGrowMTUEvent(call, 0);
6648 MUTEX_EXIT(&call->lock);
6652 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6654 if (!call->keepAliveEvent) {
6655 struct clock when, now;
6656 clock_GetTime(&now);
6658 when.sec += call->conn->secondsUntilPing;
6659 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6660 call->keepAliveEvent =
6661 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6666 rxi_CancelKeepAliveEvent(struct rx_call *call) {
6667 if (call->keepAliveEvent) {
6668 rxevent_Cancel(&call->keepAliveEvent);
6669 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6674 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6676 if (!call->growMTUEvent) {
6677 struct clock when, now;
6679 clock_GetTime(&now);
6682 if (call->conn->secondsUntilPing)
6683 secs = (6*call->conn->secondsUntilPing)-1;
6685 if (call->conn->secondsUntilDead)
6686 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6690 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6691 call->growMTUEvent =
6692 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6697 rxi_CancelGrowMTUEvent(struct rx_call *call)
6699 if (call->growMTUEvent) {
6700 rxevent_Cancel(&call->growMTUEvent);
6701 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6706 rxi_KeepAliveOn(struct rx_call *call)
6708 /* Pretend last packet received was received now--i.e. if another
6709 * packet isn't received within the keep alive time, then the call
6710 * will die; Initialize last send time to the current time--even
6711 * if a packet hasn't been sent yet. This will guarantee that a
6712 * keep-alive is sent within the ping time */
6713 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6714 rxi_ScheduleKeepAliveEvent(call);
6718 rx_KeepAliveOff(struct rx_call *call)
6720 MUTEX_ENTER(&call->lock);
6721 rxi_CancelKeepAliveEvent(call);
6722 MUTEX_EXIT(&call->lock);
6726 rx_KeepAliveOn(struct rx_call *call)
6728 MUTEX_ENTER(&call->lock);
6729 rxi_KeepAliveOn(call);
6730 MUTEX_EXIT(&call->lock);
6734 rxi_GrowMTUOn(struct rx_call *call)
6736 struct rx_connection *conn = call->conn;
6737 MUTEX_ENTER(&conn->conn_data_lock);
6738 conn->lastPingSizeSer = conn->lastPingSize = 0;
6739 MUTEX_EXIT(&conn->conn_data_lock);
6740 rxi_ScheduleGrowMTUEvent(call, 1);
6743 /* This routine is called to send connection abort messages
6744 * that have been delayed to throttle looping clients. */
6746 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6749 struct rx_connection *conn = arg1;
6752 struct rx_packet *packet;
6754 MUTEX_ENTER(&conn->conn_data_lock);
6755 rxevent_Put(&conn->delayedAbortEvent);
6756 error = htonl(conn->error);
6758 MUTEX_EXIT(&conn->conn_data_lock);
6759 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6762 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6763 RX_PACKET_TYPE_ABORT, (char *)&error,
6765 rxi_FreePacket(packet);
6769 /* This routine is called to send call abort messages
6770 * that have been delayed to throttle looping clients. */
6772 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6775 struct rx_call *call = arg1;
6778 struct rx_packet *packet;
6780 MUTEX_ENTER(&call->lock);
6781 rxevent_Put(&call->delayedAbortEvent);
6782 error = htonl(call->error);
6784 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6787 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6788 (char *)&error, sizeof(error), 0);
6789 rxi_FreePacket(packet);
6791 MUTEX_EXIT(&call->lock);
6792 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6795 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6796 * seconds) to ask the client to authenticate itself. The routine
6797 * issues a challenge to the client, which is obtained from the
6798 * security object associated with the connection */
6800 rxi_ChallengeEvent(struct rxevent *event,
6801 void *arg0, void *arg1, int tries)
6803 struct rx_connection *conn = arg0;
6806 rxevent_Put(&conn->challengeEvent);
6808 /* If there are no active calls it is not worth re-issuing the
6809 * challenge. If the client issues another call on this connection
6810 * the challenge can be requested at that time.
6812 if (!rxi_HasActiveCalls(conn))
6815 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6816 struct rx_packet *packet;
6817 struct clock when, now;
6820 /* We've failed to authenticate for too long.
6821 * Reset any calls waiting for authentication;
6822 * they are all in RX_STATE_PRECALL.
6826 MUTEX_ENTER(&conn->conn_call_lock);
6827 for (i = 0; i < RX_MAXCALLS; i++) {
6828 struct rx_call *call = conn->call[i];
6830 MUTEX_ENTER(&call->lock);
6831 if (call->state == RX_STATE_PRECALL) {
6832 rxi_CallError(call, RX_CALL_DEAD);
6833 rxi_SendCallAbort(call, NULL, 0, 0);
6835 MUTEX_EXIT(&call->lock);
6838 MUTEX_EXIT(&conn->conn_call_lock);
6842 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6844 /* If there's no packet available, do this later. */
6845 RXS_GetChallenge(conn->securityObject, conn, packet);
6846 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6847 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6848 rxi_FreePacket(packet);
6850 clock_GetTime(&now);
6852 when.sec += RX_CHALLENGE_TIMEOUT;
6853 conn->challengeEvent =
6854 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6859 /* Call this routine to start requesting the client to authenticate
6860 * itself. This will continue until authentication is established,
6861 * the call times out, or an invalid response is returned. The
6862 * security object associated with the connection is asked to create
6863 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6864 * defined earlier. */
6866 rxi_ChallengeOn(struct rx_connection *conn)
6868 if (!conn->challengeEvent) {
6869 RXS_CreateChallenge(conn->securityObject, conn);
6870 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6875 /* rxi_ComputeRoundTripTime is called with peer locked. */
6876 /* peer may be null */
6878 rxi_ComputeRoundTripTime(struct rx_packet *p,
6879 struct rx_ackPacket *ack,
6880 struct rx_call *call,
6881 struct rx_peer *peer,
6884 struct clock thisRtt, *sentp;
6888 /* If the ACK is delayed, then do nothing */
6889 if (ack->reason == RX_ACK_DELAY)
6892 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6893 * their RTT multiple times, so only include the RTT of the last packet
6895 if (p->flags & RX_JUMBO_PACKET)
6898 /* Use the serial number to determine which transmission the ACK is for,
6899 * and set the sent time to match this. If we have no serial number, then
6900 * only use the ACK for RTT calculations if the packet has not been
6904 serial = ntohl(ack->serial);
6906 if (serial == p->header.serial) {
6907 sentp = &p->timeSent;
6908 } else if (serial == p->firstSerial) {
6909 sentp = &p->firstSent;
6910 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6911 sentp = &p->firstSent;
6915 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6916 sentp = &p->firstSent;
6923 if (clock_Lt(&thisRtt, sentp))
6924 return; /* somebody set the clock back, don't count this time. */
6926 clock_Sub(&thisRtt, sentp);
6927 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6928 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6930 if (clock_IsZero(&thisRtt)) {
6932 * The actual round trip time is shorter than the
6933 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6934 * Since we can't tell which at the moment we will assume 1ms.
6936 thisRtt.usec = 1000;
6939 if (rx_stats_active) {
6940 MUTEX_ENTER(&rx_stats_mutex);
6941 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6942 rx_stats.minRtt = thisRtt;
6943 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6944 if (thisRtt.sec > 60) {
6945 MUTEX_EXIT(&rx_stats_mutex);
6946 return; /* somebody set the clock ahead */
6948 rx_stats.maxRtt = thisRtt;
6950 clock_Add(&rx_stats.totalRtt, &thisRtt);
6951 rx_atomic_inc(&rx_stats.nRttSamples);
6952 MUTEX_EXIT(&rx_stats_mutex);
6955 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6957 /* Apply VanJacobson round-trip estimations */
6962 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6963 * srtt is stored as fixed point with 3 bits after the binary
6964 * point (i.e., scaled by 8). The following magic is
6965 * equivalent to the smoothing algorithm in rfc793 with an
6966 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6967 * srtt'*8 = rtt + srtt*7
6968 * srtt'*8 = srtt*8 + rtt - srtt
6969 * srtt' = srtt + rtt/8 - srtt/8
6970 * srtt' = srtt + (rtt - srtt)/8
6973 delta = _8THMSEC(&thisRtt) - call->rtt;
6974 call->rtt += (delta >> 3);
6977 * We accumulate a smoothed rtt variance (actually, a smoothed
6978 * mean difference), then set the retransmit timer to smoothed
6979 * rtt + 4 times the smoothed variance (was 2x in van's original
6980 * paper, but 4x works better for me, and apparently for him as
6982 * rttvar is stored as
6983 * fixed point with 2 bits after the binary point (scaled by
6984 * 4). The following is equivalent to rfc793 smoothing with
6985 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6986 * rttvar'*4 = rttvar*3 + |delta|
6987 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6988 * rttvar' = rttvar + |delta|/4 - rttvar/4
6989 * rttvar' = rttvar + (|delta| - rttvar)/4
6990 * This replaces rfc793's wired-in beta.
6991 * dev*4 = dev*4 + (|actual - expected| - dev)
6997 delta -= (call->rtt_dev << 1);
6998 call->rtt_dev += (delta >> 3);
7000 /* I don't have a stored RTT so I start with this value. Since I'm
7001 * probably just starting a call, and will be pushing more data down
7002 * this, I expect congestion to increase rapidly. So I fudge a
7003 * little, and I set deviance to half the rtt. In practice,
7004 * deviance tends to approach something a little less than
7005 * half the smoothed rtt. */
7006 call->rtt = _8THMSEC(&thisRtt) + 8;
7007 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
7009 /* the smoothed RTT time is RTT + 4*MDEV
7011 * We allow a user specified minimum to be set for this, to allow clamping
7012 * at a minimum value in the same way as TCP. In addition, we have to allow
7013 * for the possibility that this packet is answered by a delayed ACK, so we
7014 * add on a fixed 200ms to account for that timer expiring.
7017 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
7018 rx_minPeerTimeout) + 200;
7019 clock_Zero(&call->rto);
7020 clock_Addmsec(&call->rto, rtt_timeout);
7022 /* Update the peer, so any new calls start with our values */
7023 peer->rtt_dev = call->rtt_dev;
7024 peer->rtt = call->rtt;
7026 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
7027 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
7031 /* Find all server connections that have not been active for a long time, and
7034 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
7037 struct clock now, when;
7038 struct rxevent *event;
7039 clock_GetTime(&now);
7041 /* Find server connection structures that haven't been used for
7042 * greater than rx_idleConnectionTime */
7044 struct rx_connection **conn_ptr, **conn_end;
7045 int i, havecalls = 0;
7046 MUTEX_ENTER(&rx_connHashTable_lock);
7047 for (conn_ptr = &rx_connHashTable[0], conn_end =
7048 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
7050 struct rx_connection *conn, *next;
7051 struct rx_call *call;
7055 for (conn = *conn_ptr; conn; conn = next) {
7056 /* XXX -- Shouldn't the connection be locked? */
7059 for (i = 0; i < RX_MAXCALLS; i++) {
7060 call = conn->call[i];
7064 code = MUTEX_TRYENTER(&call->lock);
7067 result = rxi_CheckCall(call, 1);
7068 MUTEX_EXIT(&call->lock);
7070 /* If CheckCall freed the call, it might
7071 * have destroyed the connection as well,
7072 * which screws up the linked lists.
7078 if (conn->type == RX_SERVER_CONNECTION) {
7079 /* This only actually destroys the connection if
7080 * there are no outstanding calls */
7081 MUTEX_ENTER(&conn->conn_data_lock);
7082 MUTEX_ENTER(&rx_refcnt_mutex);
7083 if (!havecalls && !conn->refCount
7084 && ((conn->lastSendTime + rx_idleConnectionTime) <
7086 conn->refCount++; /* it will be decr in rx_DestroyConn */
7087 MUTEX_EXIT(&rx_refcnt_mutex);
7088 MUTEX_EXIT(&conn->conn_data_lock);
7089 #ifdef RX_ENABLE_LOCKS
7090 rxi_DestroyConnectionNoLock(conn);
7091 #else /* RX_ENABLE_LOCKS */
7092 rxi_DestroyConnection(conn);
7093 #endif /* RX_ENABLE_LOCKS */
7095 #ifdef RX_ENABLE_LOCKS
7097 MUTEX_EXIT(&rx_refcnt_mutex);
7098 MUTEX_EXIT(&conn->conn_data_lock);
7100 #endif /* RX_ENABLE_LOCKS */
7104 #ifdef RX_ENABLE_LOCKS
7105 while (rx_connCleanup_list) {
7106 struct rx_connection *conn;
7107 conn = rx_connCleanup_list;
7108 rx_connCleanup_list = rx_connCleanup_list->next;
7109 MUTEX_EXIT(&rx_connHashTable_lock);
7110 rxi_CleanupConnection(conn);
7111 MUTEX_ENTER(&rx_connHashTable_lock);
7113 MUTEX_EXIT(&rx_connHashTable_lock);
7114 #endif /* RX_ENABLE_LOCKS */
7117 /* Find any peer structures that haven't been used (haven't had an
7118 * associated connection) for greater than rx_idlePeerTime */
7120 struct rx_peer **peer_ptr, **peer_end;
7124 * Why do we need to hold the rx_peerHashTable_lock across
7125 * the incrementing of peer_ptr since the rx_peerHashTable
7126 * array is not changing? We don't.
7128 * By dropping the lock periodically we can permit other
7129 * activities to be performed while a rxi_ReapConnections
7130 * call is in progress. The goal of reap connections
7131 * is to clean up quickly without causing large amounts
7132 * of contention. Therefore, it is important that global
7133 * mutexes not be held for extended periods of time.
7135 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7136 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7138 struct rx_peer *peer, *next, *prev;
7140 MUTEX_ENTER(&rx_peerHashTable_lock);
7141 for (prev = peer = *peer_ptr; peer; peer = next) {
7143 code = MUTEX_TRYENTER(&peer->peer_lock);
7144 if ((code) && (peer->refCount == 0)
7145 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7146 struct opr_queue *cursor, *store;
7150 * now know that this peer object is one to be
7151 * removed from the hash table. Once it is removed
7152 * it can't be referenced by other threads.
7153 * Lets remove it first and decrement the struct
7154 * nPeerStructs count.
7156 if (peer == *peer_ptr) {
7162 if (rx_stats_active)
7163 rx_atomic_dec(&rx_stats.nPeerStructs);
7166 * Now if we hold references on 'prev' and 'next'
7167 * we can safely drop the rx_peerHashTable_lock
7168 * while we destroy this 'peer' object.
7174 MUTEX_EXIT(&rx_peerHashTable_lock);
7176 MUTEX_EXIT(&peer->peer_lock);
7177 MUTEX_DESTROY(&peer->peer_lock);
7179 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7180 unsigned int num_funcs;
7181 struct rx_interface_stat *rpc_stat
7182 = opr_queue_Entry(cursor, struct rx_interface_stat,
7187 opr_queue_Remove(&rpc_stat->entry);
7188 opr_queue_Remove(&rpc_stat->entryPeers);
7190 num_funcs = rpc_stat->stats[0].func_total;
7192 sizeof(rx_interface_stat_t) +
7193 rpc_stat->stats[0].func_total *
7194 sizeof(rx_function_entry_v1_t);
7196 rxi_Free(rpc_stat, space);
7198 MUTEX_ENTER(&rx_rpc_stats);
7199 rxi_rpc_peer_stat_cnt -= num_funcs;
7200 MUTEX_EXIT(&rx_rpc_stats);
7205 * Regain the rx_peerHashTable_lock and
7206 * decrement the reference count on 'prev'
7209 MUTEX_ENTER(&rx_peerHashTable_lock);
7216 MUTEX_EXIT(&peer->peer_lock);
7221 MUTEX_EXIT(&rx_peerHashTable_lock);
7225 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7226 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7227 * GC, just below. Really, we shouldn't have to keep moving packets from
7228 * one place to another, but instead ought to always know if we can
7229 * afford to hold onto a packet in its particular use. */
7230 MUTEX_ENTER(&rx_freePktQ_lock);
7231 if (rx_waitingForPackets) {
7232 rx_waitingForPackets = 0;
7233 #ifdef RX_ENABLE_LOCKS
7234 CV_BROADCAST(&rx_waitingForPackets_cv);
7236 osi_rxWakeup(&rx_waitingForPackets);
7239 MUTEX_EXIT(&rx_freePktQ_lock);
7242 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7243 event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
7244 rxevent_Put(&event);
7248 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7249 * rx.h is sort of strange this is better. This is called with a security
7250 * object before it is discarded. Each connection using a security object has
7251 * its own refcount to the object so it won't actually be freed until the last
7252 * connection is destroyed.
7254 * This is the only rxs module call. A hold could also be written but no one
7258 rxs_Release(struct rx_securityClass *aobj)
7260 return RXS_Close(aobj);
7268 #define TRACE_OPTION_RX_DEBUG 16
7276 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7277 0, KEY_QUERY_VALUE, &parmKey);
7278 if (code != ERROR_SUCCESS)
7281 dummyLen = sizeof(TraceOption);
7282 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7283 (BYTE *) &TraceOption, &dummyLen);
7284 if (code == ERROR_SUCCESS) {
7285 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7287 RegCloseKey (parmKey);
7288 #endif /* AFS_NT40_ENV */
7293 rx_DebugOnOff(int on)
7297 rxdebug_active = on;
7303 rx_StatsOnOff(int on)
7305 rx_stats_active = on;
7309 /* Don't call this debugging routine directly; use dpf */
7311 rxi_DebugPrint(char *format, ...)
7320 va_start(ap, format);
7322 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7325 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7327 OutputDebugString(msg);
7333 va_start(ap, format);
7335 clock_GetTime(&now);
7336 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7337 (unsigned int)now.usec);
7338 vfprintf(rx_Log, format, ap);
7346 * This function is used to process the rx_stats structure that is local
7347 * to a process as well as an rx_stats structure received from a remote
7348 * process (via rxdebug). Therefore, it needs to do minimal version
7352 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7353 afs_int32 freePackets, char version)
7357 if (size != sizeof(struct rx_statistics)) {
7359 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7360 size, sizeof(struct rx_statistics));
7363 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7366 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7367 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7368 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7369 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7370 s->specialPktAllocFailures);
7372 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7373 s->receivePktAllocFailures, s->sendPktAllocFailures,
7374 s->specialPktAllocFailures);
7378 " greedy %u, " "bogusReads %u (last from host %x), "
7379 "noPackets %u, " "noBuffers %u, " "selects %u, "
7380 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7381 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7382 s->selects, s->sendSelects);
7384 fprintf(file, " packets read: ");
7385 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7386 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7388 fprintf(file, "\n");
7391 " other read counters: data %u, " "ack %u, " "dup %u "
7392 "spurious %u " "dally %u\n", s->dataPacketsRead,
7393 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7394 s->ignorePacketDally);
7396 fprintf(file, " packets sent: ");
7397 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7398 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7400 fprintf(file, "\n");
7403 " other send counters: ack %u, " "data %u (not resends), "
7404 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7405 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7406 s->dataPacketsPushed, s->ignoreAckedPacket);
7409 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7410 s->netSendFailures, (int)s->fatalErrors);
7412 if (s->nRttSamples) {
7413 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7414 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7416 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7417 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7421 " %d server connections, " "%d client connections, "
7422 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7423 s->nServerConns, s->nClientConns, s->nPeerStructs,
7424 s->nCallStructs, s->nFreeCallStructs);
7426 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7427 fprintf(file, " %d clock updates\n", clock_nUpdates);
7431 /* for backward compatibility */
7433 rx_PrintStats(FILE * file)
7435 MUTEX_ENTER(&rx_stats_mutex);
7436 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7437 sizeof(rx_stats), rx_nFreePackets,
7439 MUTEX_EXIT(&rx_stats_mutex);
7443 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7445 fprintf(file, "Peer %x.%d.\n",
7446 ntohl(peer->host), (int)ntohs(peer->port));
7449 " Rtt %d, " "total sent %d, " "resent %d\n",
7450 peer->rtt, peer->nSent, peer->reSends);
7452 fprintf(file, " Packet size %d\n", peer->ifMTU);
7456 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7458 * This mutex protects the following static variables:
7462 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7463 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7465 #define LOCK_RX_DEBUG
7466 #define UNLOCK_RX_DEBUG
7467 #endif /* AFS_PTHREAD_ENV */
7469 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7471 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7472 u_char type, void *inputData, size_t inputLength,
7473 void *outputData, size_t outputLength)
7475 static afs_int32 counter = 100;
7476 time_t waitTime, waitCount;
7477 struct rx_header theader;
7480 struct timeval tv_now, tv_wake, tv_delta;
7481 struct sockaddr_in taddr, faddr;
7495 tp = &tbuffer[sizeof(struct rx_header)];
7496 taddr.sin_family = AF_INET;
7497 taddr.sin_port = remotePort;
7498 taddr.sin_addr.s_addr = remoteAddr;
7499 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7500 taddr.sin_len = sizeof(struct sockaddr_in);
7503 memset(&theader, 0, sizeof(theader));
7504 theader.epoch = htonl(999);
7506 theader.callNumber = htonl(counter);
7509 theader.type = type;
7510 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7511 theader.serviceId = 0;
7513 memcpy(tbuffer, &theader, sizeof(theader));
7514 memcpy(tp, inputData, inputLength);
7516 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7517 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7519 /* see if there's a packet available */
7520 gettimeofday(&tv_wake, NULL);
7521 tv_wake.tv_sec += waitTime;
7524 FD_SET(socket, &imask);
7525 tv_delta.tv_sec = tv_wake.tv_sec;
7526 tv_delta.tv_usec = tv_wake.tv_usec;
7527 gettimeofday(&tv_now, NULL);
7529 if (tv_delta.tv_usec < tv_now.tv_usec) {
7531 tv_delta.tv_usec += 1000000;
7534 tv_delta.tv_usec -= tv_now.tv_usec;
7536 if (tv_delta.tv_sec < tv_now.tv_sec) {
7540 tv_delta.tv_sec -= tv_now.tv_sec;
7543 code = select(0, &imask, 0, 0, &tv_delta);
7544 #else /* AFS_NT40_ENV */
7545 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7546 #endif /* AFS_NT40_ENV */
7547 if (code == 1 && FD_ISSET(socket, &imask)) {
7548 /* now receive a packet */
7549 faddrLen = sizeof(struct sockaddr_in);
7551 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7552 (struct sockaddr *)&faddr, &faddrLen);
7555 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7556 if (counter == ntohl(theader.callNumber))
7564 /* see if we've timed out */
7572 code -= sizeof(struct rx_header);
7573 if (code > outputLength)
7574 code = outputLength;
7575 memcpy(outputData, tp, code);
7578 #endif /* RXDEBUG */
7581 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7582 afs_uint16 remotePort, struct rx_debugStats * stat,
7583 afs_uint32 * supportedValues)
7585 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7587 struct rx_debugIn in;
7589 *supportedValues = 0;
7590 in.type = htonl(RX_DEBUGI_GETSTATS);
7593 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7594 &in, sizeof(in), stat, sizeof(*stat));
7597 * If the call was successful, fixup the version and indicate
7598 * what contents of the stat structure are valid.
7599 * Also do net to host conversion of fields here.
7603 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7604 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7606 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7607 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7609 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7610 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7612 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7613 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7615 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7616 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7618 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7619 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7621 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7622 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7624 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7625 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7627 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7628 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7630 stat->nFreePackets = ntohl(stat->nFreePackets);
7631 stat->packetReclaims = ntohl(stat->packetReclaims);
7632 stat->callsExecuted = ntohl(stat->callsExecuted);
7633 stat->nWaiting = ntohl(stat->nWaiting);
7634 stat->idleThreads = ntohl(stat->idleThreads);
7635 stat->nWaited = ntohl(stat->nWaited);
7636 stat->nPackets = ntohl(stat->nPackets);
7645 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7646 afs_uint16 remotePort, struct rx_statistics * stat,
7647 afs_uint32 * supportedValues)
7649 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7651 struct rx_debugIn in;
7652 afs_int32 *lp = (afs_int32 *) stat;
7656 * supportedValues is currently unused, but added to allow future
7657 * versioning of this function.
7660 *supportedValues = 0;
7661 in.type = htonl(RX_DEBUGI_RXSTATS);
7663 memset(stat, 0, sizeof(*stat));
7665 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7666 &in, sizeof(in), stat, sizeof(*stat));
7671 * Do net to host conversion here
7674 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7685 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7686 afs_uint16 remotePort, size_t version_length,
7689 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7691 return MakeDebugCall(socket, remoteAddr, remotePort,
7692 RX_PACKET_TYPE_VERSION, a, 1, version,
7700 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7701 afs_uint16 remotePort, afs_int32 * nextConnection,
7702 int allConnections, afs_uint32 debugSupportedValues,
7703 struct rx_debugConn * conn,
7704 afs_uint32 * supportedValues)
7706 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7708 struct rx_debugIn in;
7712 * supportedValues is currently unused, but added to allow future
7713 * versioning of this function.
7716 *supportedValues = 0;
7717 if (allConnections) {
7718 in.type = htonl(RX_DEBUGI_GETALLCONN);
7720 in.type = htonl(RX_DEBUGI_GETCONN);
7722 in.index = htonl(*nextConnection);
7723 memset(conn, 0, sizeof(*conn));
7725 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7726 &in, sizeof(in), conn, sizeof(*conn));
7729 *nextConnection += 1;
7732 * Convert old connection format to new structure.
7735 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7736 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7737 #define MOVEvL(a) (conn->a = vL->a)
7739 /* any old or unrecognized version... */
7740 for (i = 0; i < RX_MAXCALLS; i++) {
7741 MOVEvL(callState[i]);
7742 MOVEvL(callMode[i]);
7743 MOVEvL(callFlags[i]);
7744 MOVEvL(callOther[i]);
7746 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7747 MOVEvL(secStats.type);
7748 MOVEvL(secStats.level);
7749 MOVEvL(secStats.flags);
7750 MOVEvL(secStats.expires);
7751 MOVEvL(secStats.packetsReceived);
7752 MOVEvL(secStats.packetsSent);
7753 MOVEvL(secStats.bytesReceived);
7754 MOVEvL(secStats.bytesSent);
7759 * Do net to host conversion here
7761 * I don't convert host or port since we are most likely
7762 * going to want these in NBO.
7764 conn->cid = ntohl(conn->cid);
7765 conn->serial = ntohl(conn->serial);
7766 for (i = 0; i < RX_MAXCALLS; i++) {
7767 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7769 conn->error = ntohl(conn->error);
7770 conn->secStats.flags = ntohl(conn->secStats.flags);
7771 conn->secStats.expires = ntohl(conn->secStats.expires);
7772 conn->secStats.packetsReceived =
7773 ntohl(conn->secStats.packetsReceived);
7774 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7775 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7776 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7777 conn->epoch = ntohl(conn->epoch);
7778 conn->natMTU = ntohl(conn->natMTU);
7787 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7788 afs_uint16 remotePort, afs_int32 * nextPeer,
7789 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7790 afs_uint32 * supportedValues)
7792 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7794 struct rx_debugIn in;
7797 * supportedValues is currently unused, but added to allow future
7798 * versioning of this function.
7801 *supportedValues = 0;
7802 in.type = htonl(RX_DEBUGI_GETPEER);
7803 in.index = htonl(*nextPeer);
7804 memset(peer, 0, sizeof(*peer));
7806 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7807 &in, sizeof(in), peer, sizeof(*peer));
7813 * Do net to host conversion here
7815 * I don't convert host or port since we are most likely
7816 * going to want these in NBO.
7818 peer->ifMTU = ntohs(peer->ifMTU);
7819 peer->idleWhen = ntohl(peer->idleWhen);
7820 peer->refCount = ntohs(peer->refCount);
7821 peer->rtt = ntohl(peer->rtt);
7822 peer->rtt_dev = ntohl(peer->rtt_dev);
7823 peer->timeout.sec = 0;
7824 peer->timeout.usec = 0;
7825 peer->nSent = ntohl(peer->nSent);
7826 peer->reSends = ntohl(peer->reSends);
7827 peer->natMTU = ntohs(peer->natMTU);
7828 peer->maxMTU = ntohs(peer->maxMTU);
7829 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7830 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7831 peer->MTU = ntohs(peer->MTU);
7832 peer->cwind = ntohs(peer->cwind);
7833 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7834 peer->congestSeq = ntohs(peer->congestSeq);
7835 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7836 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7837 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7838 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7847 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7848 struct rx_debugPeer * peerStats)
7851 afs_int32 error = 1; /* default to "did not succeed" */
7852 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7854 MUTEX_ENTER(&rx_peerHashTable_lock);
7855 for(tp = rx_peerHashTable[hashValue];
7856 tp != NULL; tp = tp->next) {
7857 if (tp->host == peerHost)
7863 MUTEX_EXIT(&rx_peerHashTable_lock);
7867 MUTEX_ENTER(&tp->peer_lock);
7868 peerStats->host = tp->host;
7869 peerStats->port = tp->port;
7870 peerStats->ifMTU = tp->ifMTU;
7871 peerStats->idleWhen = tp->idleWhen;
7872 peerStats->refCount = tp->refCount;
7873 peerStats->burstSize = 0;
7874 peerStats->burst = 0;
7875 peerStats->burstWait.sec = 0;
7876 peerStats->burstWait.usec = 0;
7877 peerStats->rtt = tp->rtt;
7878 peerStats->rtt_dev = tp->rtt_dev;
7879 peerStats->timeout.sec = 0;
7880 peerStats->timeout.usec = 0;
7881 peerStats->nSent = tp->nSent;
7882 peerStats->reSends = tp->reSends;
7883 peerStats->natMTU = tp->natMTU;
7884 peerStats->maxMTU = tp->maxMTU;
7885 peerStats->maxDgramPackets = tp->maxDgramPackets;
7886 peerStats->ifDgramPackets = tp->ifDgramPackets;
7887 peerStats->MTU = tp->MTU;
7888 peerStats->cwind = tp->cwind;
7889 peerStats->nDgramPackets = tp->nDgramPackets;
7890 peerStats->congestSeq = tp->congestSeq;
7891 peerStats->bytesSent.high = tp->bytesSent >> 32;
7892 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7893 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7894 peerStats->bytesReceived.low
7895 = tp->bytesReceived & MAX_AFS_UINT32;
7896 MUTEX_EXIT(&tp->peer_lock);
7898 MUTEX_ENTER(&rx_peerHashTable_lock);
7901 MUTEX_EXIT(&rx_peerHashTable_lock);
7909 struct rx_serverQueueEntry *np;
7912 struct rx_call *call;
7913 struct rx_serverQueueEntry *sq;
7916 if (rx_atomic_test_and_set_bit(&rxinit_status, 0))
7917 return; /* Already shutdown. */
7921 #ifndef AFS_PTHREAD_ENV
7922 FD_ZERO(&rx_selectMask);
7923 #endif /* AFS_PTHREAD_ENV */
7924 rxi_dataQuota = RX_MAX_QUOTA;
7925 #ifndef AFS_PTHREAD_ENV
7927 #endif /* AFS_PTHREAD_ENV */
7930 #ifndef AFS_PTHREAD_ENV
7931 #ifndef AFS_USE_GETTIMEOFDAY
7933 #endif /* AFS_USE_GETTIMEOFDAY */
7934 #endif /* AFS_PTHREAD_ENV */
7936 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7937 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7938 opr_queue_Remove(&call->entry);
7939 rxi_Free(call, sizeof(struct rx_call));
7942 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7943 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7945 opr_queue_Remove(&sq->entry);
7950 struct rx_peer **peer_ptr, **peer_end;
7951 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7952 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7954 struct rx_peer *peer, *next;
7956 MUTEX_ENTER(&rx_peerHashTable_lock);
7957 for (peer = *peer_ptr; peer; peer = next) {
7958 struct opr_queue *cursor, *store;
7961 MUTEX_ENTER(&rx_rpc_stats);
7962 MUTEX_ENTER(&peer->peer_lock);
7963 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7964 unsigned int num_funcs;
7965 struct rx_interface_stat *rpc_stat
7966 = opr_queue_Entry(cursor, struct rx_interface_stat,
7970 opr_queue_Remove(&rpc_stat->entry);
7971 opr_queue_Remove(&rpc_stat->entryPeers);
7972 num_funcs = rpc_stat->stats[0].func_total;
7974 sizeof(rx_interface_stat_t) +
7975 rpc_stat->stats[0].func_total *
7976 sizeof(rx_function_entry_v1_t);
7978 rxi_Free(rpc_stat, space);
7980 /* rx_rpc_stats must be held */
7981 rxi_rpc_peer_stat_cnt -= num_funcs;
7983 MUTEX_EXIT(&peer->peer_lock);
7984 MUTEX_EXIT(&rx_rpc_stats);
7988 if (rx_stats_active)
7989 rx_atomic_dec(&rx_stats.nPeerStructs);
7991 MUTEX_EXIT(&rx_peerHashTable_lock);
7994 for (i = 0; i < RX_MAX_SERVICES; i++) {
7996 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7998 for (i = 0; i < rx_hashTableSize; i++) {
7999 struct rx_connection *tc, *ntc;
8000 MUTEX_ENTER(&rx_connHashTable_lock);
8001 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
8003 for (j = 0; j < RX_MAXCALLS; j++) {
8005 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
8008 rxi_Free(tc, sizeof(*tc));
8010 MUTEX_EXIT(&rx_connHashTable_lock);
8013 MUTEX_ENTER(&freeSQEList_lock);
8015 while ((np = rx_FreeSQEList)) {
8016 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
8017 MUTEX_DESTROY(&np->lock);
8018 rxi_Free(np, sizeof(*np));
8021 MUTEX_EXIT(&freeSQEList_lock);
8022 MUTEX_DESTROY(&freeSQEList_lock);
8023 MUTEX_DESTROY(&rx_freeCallQueue_lock);
8024 MUTEX_DESTROY(&rx_connHashTable_lock);
8025 MUTEX_DESTROY(&rx_peerHashTable_lock);
8026 MUTEX_DESTROY(&rx_serverPool_lock);
8028 osi_Free(rx_connHashTable,
8029 rx_hashTableSize * sizeof(struct rx_connection *));
8030 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8032 UNPIN(rx_connHashTable,
8033 rx_hashTableSize * sizeof(struct rx_connection *));
8034 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
8036 rxi_FreeAllPackets();
8038 MUTEX_ENTER(&rx_quota_mutex);
8039 rxi_dataQuota = RX_MAX_QUOTA;
8040 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
8041 MUTEX_EXIT(&rx_quota_mutex);
8047 * Routines to implement connection specific data.
8051 rx_KeyCreate(rx_destructor_t rtn)
8054 MUTEX_ENTER(&rxi_keyCreate_lock);
8055 key = rxi_keyCreate_counter++;
8056 rxi_keyCreate_destructor = (rx_destructor_t *)
8057 realloc((void *)rxi_keyCreate_destructor,
8058 (key + 1) * sizeof(rx_destructor_t));
8059 rxi_keyCreate_destructor[key] = rtn;
8060 MUTEX_EXIT(&rxi_keyCreate_lock);
8065 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8068 MUTEX_ENTER(&conn->conn_data_lock);
8069 if (!conn->specific) {
8070 conn->specific = malloc((key + 1) * sizeof(void *));
8071 for (i = 0; i < key; i++)
8072 conn->specific[i] = NULL;
8073 conn->nSpecific = key + 1;
8074 conn->specific[key] = ptr;
8075 } else if (key >= conn->nSpecific) {
8076 conn->specific = (void **)
8077 realloc(conn->specific, (key + 1) * sizeof(void *));
8078 for (i = conn->nSpecific; i < key; i++)
8079 conn->specific[i] = NULL;
8080 conn->nSpecific = key + 1;
8081 conn->specific[key] = ptr;
8083 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8084 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8085 conn->specific[key] = ptr;
8087 MUTEX_EXIT(&conn->conn_data_lock);
8091 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8094 MUTEX_ENTER(&svc->svc_data_lock);
8095 if (!svc->specific) {
8096 svc->specific = malloc((key + 1) * sizeof(void *));
8097 for (i = 0; i < key; i++)
8098 svc->specific[i] = NULL;
8099 svc->nSpecific = key + 1;
8100 svc->specific[key] = ptr;
8101 } else if (key >= svc->nSpecific) {
8102 svc->specific = (void **)
8103 realloc(svc->specific, (key + 1) * sizeof(void *));
8104 for (i = svc->nSpecific; i < key; i++)
8105 svc->specific[i] = NULL;
8106 svc->nSpecific = key + 1;
8107 svc->specific[key] = ptr;
8109 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8110 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8111 svc->specific[key] = ptr;
8113 MUTEX_EXIT(&svc->svc_data_lock);
8117 rx_GetSpecific(struct rx_connection *conn, int key)
8120 MUTEX_ENTER(&conn->conn_data_lock);
8121 if (key >= conn->nSpecific)
8124 ptr = conn->specific[key];
8125 MUTEX_EXIT(&conn->conn_data_lock);
8130 rx_GetServiceSpecific(struct rx_service *svc, int key)
8133 MUTEX_ENTER(&svc->svc_data_lock);
8134 if (key >= svc->nSpecific)
8137 ptr = svc->specific[key];
8138 MUTEX_EXIT(&svc->svc_data_lock);
8143 #endif /* !KERNEL */
8146 * processStats is a queue used to store the statistics for the local
8147 * process. Its contents are similar to the contents of the rpcStats
8148 * queue on a rx_peer structure, but the actual data stored within
8149 * this queue contains totals across the lifetime of the process (assuming
8150 * the stats have not been reset) - unlike the per peer structures
8151 * which can come and go based upon the peer lifetime.
8154 static struct opr_queue processStats = { &processStats, &processStats };
8157 * peerStats is a queue used to store the statistics for all peer structs.
8158 * Its contents are the union of all the peer rpcStats queues.
8161 static struct opr_queue peerStats = { &peerStats, &peerStats };
8164 * rxi_monitor_processStats is used to turn process wide stat collection
8168 static int rxi_monitor_processStats = 0;
8171 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8174 static int rxi_monitor_peerStats = 0;
8178 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8180 rpc_stat->invocations = 0;
8181 rpc_stat->bytes_sent = 0;
8182 rpc_stat->bytes_rcvd = 0;
8183 rpc_stat->queue_time_sum.sec = 0;
8184 rpc_stat->queue_time_sum.usec = 0;
8185 rpc_stat->queue_time_sum_sqr.sec = 0;
8186 rpc_stat->queue_time_sum_sqr.usec = 0;
8187 rpc_stat->queue_time_min.sec = 9999999;
8188 rpc_stat->queue_time_min.usec = 9999999;
8189 rpc_stat->queue_time_max.sec = 0;
8190 rpc_stat->queue_time_max.usec = 0;
8191 rpc_stat->execution_time_sum.sec = 0;
8192 rpc_stat->execution_time_sum.usec = 0;
8193 rpc_stat->execution_time_sum_sqr.sec = 0;
8194 rpc_stat->execution_time_sum_sqr.usec = 0;
8195 rpc_stat->execution_time_min.sec = 9999999;
8196 rpc_stat->execution_time_min.usec = 9999999;
8197 rpc_stat->execution_time_max.sec = 0;
8198 rpc_stat->execution_time_max.usec = 0;
8202 * Given all of the information for a particular rpc
8203 * call, find or create (if requested) the stat structure for the rpc.
8206 * the queue of stats that will be updated with the new value
8208 * @param rxInterface
8209 * a unique number that identifies the rpc interface
8212 * the total number of functions in this interface. this is only
8213 * required if create is true
8216 * if true, this invocation was made to a server
8219 * the ip address of the remote host. this is only required if create
8220 * and addToPeerList are true
8223 * the port of the remote host. this is only required if create
8224 * and addToPeerList are true
8226 * @param addToPeerList
8227 * if != 0, add newly created stat to the global peer list
8230 * if a new stats structure is allocated, the counter will
8231 * be updated with the new number of allocated stat structures.
8232 * only required if create is true
8235 * if no stats structure exists, allocate one
8239 static rx_interface_stat_p
8240 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8241 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8242 afs_uint32 remotePort, int addToPeerList,
8243 unsigned int *counter, int create)
8245 rx_interface_stat_p rpc_stat = NULL;
8246 struct opr_queue *cursor;
8249 * See if there's already a structure for this interface
8252 for (opr_queue_Scan(stats, cursor)) {
8253 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8255 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8256 && (rpc_stat->stats[0].remote_is_server == isServer))
8260 /* if they didn't ask us to create, we're done */
8262 if (opr_queue_IsEnd(stats, cursor))
8268 /* can't proceed without these */
8269 if (!totalFunc || !counter)
8273 * Didn't find a match so allocate a new structure and add it to the
8277 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8278 || (rpc_stat->stats[0].interfaceId != rxInterface)
8279 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8284 sizeof(rx_interface_stat_t) +
8285 totalFunc * sizeof(rx_function_entry_v1_t);
8287 rpc_stat = rxi_Alloc(space);
8288 if (rpc_stat == NULL)
8291 *counter += totalFunc;
8292 for (i = 0; i < totalFunc; i++) {
8293 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8294 rpc_stat->stats[i].remote_peer = remoteHost;
8295 rpc_stat->stats[i].remote_port = remotePort;
8296 rpc_stat->stats[i].remote_is_server = isServer;
8297 rpc_stat->stats[i].interfaceId = rxInterface;
8298 rpc_stat->stats[i].func_total = totalFunc;
8299 rpc_stat->stats[i].func_index = i;
8301 opr_queue_Prepend(stats, &rpc_stat->entry);
8302 if (addToPeerList) {
8303 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8310 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8312 rx_interface_stat_p rpc_stat;
8315 if (rxInterface == -1)
8318 MUTEX_ENTER(&rx_rpc_stats);
8319 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8322 totalFunc = rpc_stat->stats[0].func_total;
8323 for (i = 0; i < totalFunc; i++)
8324 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8326 MUTEX_EXIT(&rx_rpc_stats);
8331 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8333 rx_interface_stat_p rpc_stat;
8335 struct rx_peer * peer;
8337 if (rxInterface == -1)
8340 peer = rxi_FindPeer(peerHost, peerPort, 0);
8344 MUTEX_ENTER(&rx_rpc_stats);
8345 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8348 totalFunc = rpc_stat->stats[0].func_total;
8349 for (i = 0; i < totalFunc; i++)
8350 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8352 MUTEX_EXIT(&rx_rpc_stats);
8357 rx_CopyProcessRPCStats(afs_uint64 op)
8359 rx_interface_stat_p rpc_stat;
8360 rx_function_entry_v1_p rpcop_stat =
8361 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8362 int currentFunc = (op & MAX_AFS_UINT32);
8363 afs_int32 rxInterface = (op >> 32);
8365 if (!rxi_monitor_processStats)
8368 if (rxInterface == -1)
8371 if (rpcop_stat == NULL)
8374 MUTEX_ENTER(&rx_rpc_stats);
8375 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8378 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8379 sizeof(rx_function_entry_v1_t));
8380 MUTEX_EXIT(&rx_rpc_stats);
8382 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8389 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8391 rx_interface_stat_p rpc_stat;
8392 rx_function_entry_v1_p rpcop_stat =
8393 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8394 int currentFunc = (op & MAX_AFS_UINT32);
8395 afs_int32 rxInterface = (op >> 32);
8396 struct rx_peer *peer;
8398 if (!rxi_monitor_peerStats)
8401 if (rxInterface == -1)
8404 if (rpcop_stat == NULL)
8407 peer = rxi_FindPeer(peerHost, peerPort, 0);
8411 MUTEX_ENTER(&rx_rpc_stats);
8412 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8415 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8416 sizeof(rx_function_entry_v1_t));
8417 MUTEX_EXIT(&rx_rpc_stats);
8419 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8426 rx_ReleaseRPCStats(void *stats)
8429 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8433 * Given all of the information for a particular rpc
8434 * call, create (if needed) and update the stat totals for the rpc.
8437 * the queue of stats that will be updated with the new value
8439 * @param rxInterface
8440 * a unique number that identifies the rpc interface
8442 * @param currentFunc
8443 * the index of the function being invoked
8446 * the total number of functions in this interface
8449 * the amount of time this function waited for a thread
8452 * the amount of time this function invocation took to execute
8455 * the number bytes sent by this invocation
8458 * the number bytes received by this invocation
8461 * if true, this invocation was made to a server
8464 * the ip address of the remote host
8467 * the port of the remote host
8469 * @param addToPeerList
8470 * if != 0, add newly created stat to the global peer list
8473 * if a new stats structure is allocated, the counter will
8474 * be updated with the new number of allocated stat structures
8479 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8480 afs_uint32 currentFunc, afs_uint32 totalFunc,
8481 struct clock *queueTime, struct clock *execTime,
8482 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8483 afs_uint32 remoteHost, afs_uint32 remotePort,
8484 int addToPeerList, unsigned int *counter)
8487 rx_interface_stat_p rpc_stat;
8489 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8490 remoteHost, remotePort, addToPeerList, counter,
8498 * Increment the stats for this function
8501 rpc_stat->stats[currentFunc].invocations++;
8502 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8503 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8504 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8505 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8506 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8507 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8509 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8510 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8512 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8513 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8515 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8516 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8518 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8519 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8527 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8528 afs_uint32 currentFunc, afs_uint32 totalFunc,
8529 struct clock *queueTime, struct clock *execTime,
8530 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8534 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8537 MUTEX_ENTER(&rx_rpc_stats);
8539 if (rxi_monitor_peerStats) {
8540 MUTEX_ENTER(&peer->peer_lock);
8541 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8542 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8543 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8544 MUTEX_EXIT(&peer->peer_lock);
8547 if (rxi_monitor_processStats) {
8548 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8549 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8550 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8553 MUTEX_EXIT(&rx_rpc_stats);
8557 * Increment the times and count for a particular rpc function.
8559 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8560 * call rx_RecordCallStatistics instead, so the public version of this
8561 * function is left purely for legacy callers.
8564 * The peer who invoked the rpc
8566 * @param rxInterface
8567 * A unique number that identifies the rpc interface
8569 * @param currentFunc
8570 * The index of the function being invoked
8573 * The total number of functions in this interface
8576 * The amount of time this function waited for a thread
8579 * The amount of time this function invocation took to execute
8582 * The number bytes sent by this invocation
8585 * The number bytes received by this invocation
8588 * If true, this invocation was made to a server
8592 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8593 afs_uint32 currentFunc, afs_uint32 totalFunc,
8594 struct clock *queueTime, struct clock *execTime,
8595 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8601 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8602 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8604 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8605 queueTime, execTime, sent64, rcvd64,
8612 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8616 * IN callerVersion - the rpc stat version of the caller.
8618 * IN count - the number of entries to marshall.
8620 * IN stats - pointer to stats to be marshalled.
8622 * OUT ptr - Where to store the marshalled data.
8629 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8630 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8636 * We only support the first version
8638 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8639 *(ptr++) = stats->remote_peer;
8640 *(ptr++) = stats->remote_port;
8641 *(ptr++) = stats->remote_is_server;
8642 *(ptr++) = stats->interfaceId;
8643 *(ptr++) = stats->func_total;
8644 *(ptr++) = stats->func_index;
8645 *(ptr++) = stats->invocations >> 32;
8646 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8647 *(ptr++) = stats->bytes_sent >> 32;
8648 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8649 *(ptr++) = stats->bytes_rcvd >> 32;
8650 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8651 *(ptr++) = stats->queue_time_sum.sec;
8652 *(ptr++) = stats->queue_time_sum.usec;
8653 *(ptr++) = stats->queue_time_sum_sqr.sec;
8654 *(ptr++) = stats->queue_time_sum_sqr.usec;
8655 *(ptr++) = stats->queue_time_min.sec;
8656 *(ptr++) = stats->queue_time_min.usec;
8657 *(ptr++) = stats->queue_time_max.sec;
8658 *(ptr++) = stats->queue_time_max.usec;
8659 *(ptr++) = stats->execution_time_sum.sec;
8660 *(ptr++) = stats->execution_time_sum.usec;
8661 *(ptr++) = stats->execution_time_sum_sqr.sec;
8662 *(ptr++) = stats->execution_time_sum_sqr.usec;
8663 *(ptr++) = stats->execution_time_min.sec;
8664 *(ptr++) = stats->execution_time_min.usec;
8665 *(ptr++) = stats->execution_time_max.sec;
8666 *(ptr++) = stats->execution_time_max.usec;
8672 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8677 * IN callerVersion - the rpc stat version of the caller
8679 * OUT myVersion - the rpc stat version of this function
8681 * OUT clock_sec - local time seconds
8683 * OUT clock_usec - local time microseconds
8685 * OUT allocSize - the number of bytes allocated to contain stats
8687 * OUT statCount - the number stats retrieved from this process.
8689 * OUT stats - the actual stats retrieved from this process.
8693 * Returns void. If successful, stats will != NULL.
8697 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8698 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8699 size_t * allocSize, afs_uint32 * statCount,
8700 afs_uint32 ** stats)
8710 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8713 * Check to see if stats are enabled
8716 MUTEX_ENTER(&rx_rpc_stats);
8717 if (!rxi_monitor_processStats) {
8718 MUTEX_EXIT(&rx_rpc_stats);
8722 clock_GetTime(&now);
8723 *clock_sec = now.sec;
8724 *clock_usec = now.usec;
8727 * Allocate the space based upon the caller version
8729 * If the client is at an older version than we are,
8730 * we return the statistic data in the older data format, but
8731 * we still return our version number so the client knows we
8732 * are maintaining more data than it can retrieve.
8735 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8736 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8737 *statCount = rxi_rpc_process_stat_cnt;
8740 * This can't happen yet, but in the future version changes
8741 * can be handled by adding additional code here
8745 if (space > (size_t) 0) {
8747 ptr = *stats = rxi_Alloc(space);
8750 struct opr_queue *cursor;
8752 for (opr_queue_Scan(&processStats, cursor)) {
8753 struct rx_interface_stat *rpc_stat =
8754 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8756 * Copy the data based upon the caller version
8758 rx_MarshallProcessRPCStats(callerVersion,
8759 rpc_stat->stats[0].func_total,
8760 rpc_stat->stats, &ptr);
8766 MUTEX_EXIT(&rx_rpc_stats);
8771 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8775 * IN callerVersion - the rpc stat version of the caller
8777 * OUT myVersion - the rpc stat version of this function
8779 * OUT clock_sec - local time seconds
8781 * OUT clock_usec - local time microseconds
8783 * OUT allocSize - the number of bytes allocated to contain stats
8785 * OUT statCount - the number of stats retrieved from the individual
8788 * OUT stats - the actual stats retrieved from the individual peer structures.
8792 * Returns void. If successful, stats will != NULL.
8796 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8797 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8798 size_t * allocSize, afs_uint32 * statCount,
8799 afs_uint32 ** stats)
8809 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8812 * Check to see if stats are enabled
8815 MUTEX_ENTER(&rx_rpc_stats);
8816 if (!rxi_monitor_peerStats) {
8817 MUTEX_EXIT(&rx_rpc_stats);
8821 clock_GetTime(&now);
8822 *clock_sec = now.sec;
8823 *clock_usec = now.usec;
8826 * Allocate the space based upon the caller version
8828 * If the client is at an older version than we are,
8829 * we return the statistic data in the older data format, but
8830 * we still return our version number so the client knows we
8831 * are maintaining more data than it can retrieve.
8834 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8835 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8836 *statCount = rxi_rpc_peer_stat_cnt;
8839 * This can't happen yet, but in the future version changes
8840 * can be handled by adding additional code here
8844 if (space > (size_t) 0) {
8846 ptr = *stats = rxi_Alloc(space);
8849 struct opr_queue *cursor;
8851 for (opr_queue_Scan(&peerStats, cursor)) {
8852 struct rx_interface_stat *rpc_stat
8853 = opr_queue_Entry(cursor, struct rx_interface_stat,
8857 * Copy the data based upon the caller version
8859 rx_MarshallProcessRPCStats(callerVersion,
8860 rpc_stat->stats[0].func_total,
8861 rpc_stat->stats, &ptr);
8867 MUTEX_EXIT(&rx_rpc_stats);
8872 * rx_FreeRPCStats - free memory allocated by
8873 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8877 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8878 * rx_RetrievePeerRPCStats
8880 * IN allocSize - the number of bytes in stats.
8888 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8890 rxi_Free(stats, allocSize);
8894 * rx_queryProcessRPCStats - see if process rpc stat collection is
8895 * currently enabled.
8901 * Returns 0 if stats are not enabled != 0 otherwise
8905 rx_queryProcessRPCStats(void)
8908 MUTEX_ENTER(&rx_rpc_stats);
8909 rc = rxi_monitor_processStats;
8910 MUTEX_EXIT(&rx_rpc_stats);
8915 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8921 * Returns 0 if stats are not enabled != 0 otherwise
8925 rx_queryPeerRPCStats(void)
8928 MUTEX_ENTER(&rx_rpc_stats);
8929 rc = rxi_monitor_peerStats;
8930 MUTEX_EXIT(&rx_rpc_stats);
8935 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8945 rx_enableProcessRPCStats(void)
8947 MUTEX_ENTER(&rx_rpc_stats);
8948 rx_enable_stats = 1;
8949 rxi_monitor_processStats = 1;
8950 MUTEX_EXIT(&rx_rpc_stats);
8954 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8964 rx_enablePeerRPCStats(void)
8966 MUTEX_ENTER(&rx_rpc_stats);
8967 rx_enable_stats = 1;
8968 rxi_monitor_peerStats = 1;
8969 MUTEX_EXIT(&rx_rpc_stats);
8973 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8983 rx_disableProcessRPCStats(void)
8985 struct opr_queue *cursor, *store;
8988 MUTEX_ENTER(&rx_rpc_stats);
8991 * Turn off process statistics and if peer stats is also off, turn
8995 rxi_monitor_processStats = 0;
8996 if (rxi_monitor_peerStats == 0) {
8997 rx_enable_stats = 0;
9000 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
9001 unsigned int num_funcs = 0;
9002 struct rx_interface_stat *rpc_stat
9003 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9005 opr_queue_Remove(&rpc_stat->entry);
9007 num_funcs = rpc_stat->stats[0].func_total;
9009 sizeof(rx_interface_stat_t) +
9010 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
9012 rxi_Free(rpc_stat, space);
9013 rxi_rpc_process_stat_cnt -= num_funcs;
9015 MUTEX_EXIT(&rx_rpc_stats);
9019 * rx_disablePeerRPCStats - stop rpc stat collection for peers
9029 rx_disablePeerRPCStats(void)
9031 struct rx_peer **peer_ptr, **peer_end;
9035 * Turn off peer statistics and if process stats is also off, turn
9039 rxi_monitor_peerStats = 0;
9040 if (rxi_monitor_processStats == 0) {
9041 rx_enable_stats = 0;
9044 for (peer_ptr = &rx_peerHashTable[0], peer_end =
9045 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
9047 struct rx_peer *peer, *next, *prev;
9049 MUTEX_ENTER(&rx_peerHashTable_lock);
9050 MUTEX_ENTER(&rx_rpc_stats);
9051 for (prev = peer = *peer_ptr; peer; peer = next) {
9053 code = MUTEX_TRYENTER(&peer->peer_lock);
9056 struct opr_queue *cursor, *store;
9058 if (prev == *peer_ptr) {
9069 MUTEX_EXIT(&rx_peerHashTable_lock);
9071 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9072 unsigned int num_funcs = 0;
9073 struct rx_interface_stat *rpc_stat
9074 = opr_queue_Entry(cursor, struct rx_interface_stat,
9077 opr_queue_Remove(&rpc_stat->entry);
9078 opr_queue_Remove(&rpc_stat->entryPeers);
9079 num_funcs = rpc_stat->stats[0].func_total;
9081 sizeof(rx_interface_stat_t) +
9082 rpc_stat->stats[0].func_total *
9083 sizeof(rx_function_entry_v1_t);
9085 rxi_Free(rpc_stat, space);
9086 rxi_rpc_peer_stat_cnt -= num_funcs;
9088 MUTEX_EXIT(&peer->peer_lock);
9090 MUTEX_ENTER(&rx_peerHashTable_lock);
9100 MUTEX_EXIT(&rx_rpc_stats);
9101 MUTEX_EXIT(&rx_peerHashTable_lock);
9106 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9111 * IN clearFlag - flag indicating which stats to clear
9119 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9121 struct opr_queue *cursor;
9123 MUTEX_ENTER(&rx_rpc_stats);
9125 for (opr_queue_Scan(&processStats, cursor)) {
9126 unsigned int num_funcs = 0, i;
9127 struct rx_interface_stat *rpc_stat
9128 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
9130 num_funcs = rpc_stat->stats[0].func_total;
9131 for (i = 0; i < num_funcs; i++) {
9132 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9133 rpc_stat->stats[i].invocations = 0;
9135 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9136 rpc_stat->stats[i].bytes_sent = 0;
9138 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9139 rpc_stat->stats[i].bytes_rcvd = 0;
9141 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9142 rpc_stat->stats[i].queue_time_sum.sec = 0;
9143 rpc_stat->stats[i].queue_time_sum.usec = 0;
9145 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9146 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9147 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9149 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9150 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9151 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9153 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9154 rpc_stat->stats[i].queue_time_max.sec = 0;
9155 rpc_stat->stats[i].queue_time_max.usec = 0;
9157 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9158 rpc_stat->stats[i].execution_time_sum.sec = 0;
9159 rpc_stat->stats[i].execution_time_sum.usec = 0;
9161 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9162 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9163 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9165 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9166 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9167 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9169 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9170 rpc_stat->stats[i].execution_time_max.sec = 0;
9171 rpc_stat->stats[i].execution_time_max.usec = 0;
9176 MUTEX_EXIT(&rx_rpc_stats);
9180 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9185 * IN clearFlag - flag indicating which stats to clear
9193 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9195 struct opr_queue *cursor;
9197 MUTEX_ENTER(&rx_rpc_stats);
9199 for (opr_queue_Scan(&peerStats, cursor)) {
9200 unsigned int num_funcs, i;
9201 struct rx_interface_stat *rpc_stat
9202 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9204 num_funcs = rpc_stat->stats[0].func_total;
9205 for (i = 0; i < num_funcs; i++) {
9206 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9207 rpc_stat->stats[i].invocations = 0;
9209 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9210 rpc_stat->stats[i].bytes_sent = 0;
9212 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9213 rpc_stat->stats[i].bytes_rcvd = 0;
9215 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9216 rpc_stat->stats[i].queue_time_sum.sec = 0;
9217 rpc_stat->stats[i].queue_time_sum.usec = 0;
9219 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9220 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9221 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9223 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9224 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9225 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9227 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9228 rpc_stat->stats[i].queue_time_max.sec = 0;
9229 rpc_stat->stats[i].queue_time_max.usec = 0;
9231 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9232 rpc_stat->stats[i].execution_time_sum.sec = 0;
9233 rpc_stat->stats[i].execution_time_sum.usec = 0;
9235 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9236 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9237 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9239 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9240 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9241 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9243 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9244 rpc_stat->stats[i].execution_time_max.sec = 0;
9245 rpc_stat->stats[i].execution_time_max.usec = 0;
9250 MUTEX_EXIT(&rx_rpc_stats);
9254 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9255 * is authorized to enable/disable/clear RX statistics.
9257 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9260 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9262 rxi_rxstat_userok = proc;
9266 rx_RxStatUserOk(struct rx_call *call)
9268 if (!rxi_rxstat_userok)
9270 return rxi_rxstat_userok(call);
9275 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9276 * function in the MSVC runtime DLL (msvcrt.dll).
9278 * Note: the system serializes calls to this function.
9281 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9282 DWORD reason, /* reason function is being called */
9283 LPVOID reserved) /* reserved for future use */
9286 case DLL_PROCESS_ATTACH:
9287 /* library is being attached to a process */
9291 case DLL_PROCESS_DETACH:
9298 #endif /* AFS_NT40_ENV */
9301 int rx_DumpCalls(FILE *outputFile, char *cookie)
9303 #ifdef RXDEBUG_PACKET
9304 #ifdef KDUMP_RX_LOCK
9305 struct rx_call_rx_lock *c;
9312 #define RXDPRINTF sprintf
9313 #define RXDPRINTOUT output
9315 #define RXDPRINTF fprintf
9316 #define RXDPRINTOUT outputFile
9319 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9321 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9324 for (c = rx_allCallsp; c; c = c->allNextp) {
9325 u_short rqc, tqc, iovqc;
9327 MUTEX_ENTER(&c->lock);
9328 rqc = opr_queue_Count(&c->rq);
9329 tqc = opr_queue_Count(&c->tq);
9330 iovqc = opr_queue_Count(&c->app.iovq);
9332 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, "
9333 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9334 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9335 "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9336 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9337 #ifdef RX_ENABLE_LOCKS
9340 #ifdef RX_REFCOUNT_CHECK
9341 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9342 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9345 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,
9346 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9347 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9348 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9349 c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9350 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9351 #ifdef RX_ENABLE_LOCKS
9352 , (afs_uint32)c->refCount
9354 #ifdef RX_REFCOUNT_CHECK
9355 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9358 MUTEX_EXIT(&c->lock);
9361 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9364 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9366 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9368 #endif /* RXDEBUG_PACKET */