2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
75 #include <opr/queue.h>
79 #include "rx_atomic.h"
80 #include "rx_globals.h"
82 #include "rx_internal.h"
89 #include "rx_packet.h"
90 #include "rx_server.h"
92 #include <afs/rxgen_consts.h>
95 #ifdef AFS_PTHREAD_ENV
97 int (*registerProgram) (pid_t, char *) = 0;
98 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
101 int (*registerProgram) (PROCESS, char *) = 0;
102 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
106 /* Local static routines */
107 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
108 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
109 struct rx_call *, struct rx_peer *,
111 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
113 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
114 void *dummy, int dummy2);
115 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
116 void *dummy, int dummy2);
117 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
118 void *unused, int unused2);
119 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
120 void *unused2, int unused3);
121 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
122 struct rx_packet *packet,
123 int istack, int force);
124 static void rxi_AckAll(struct rx_call *call);
125 static struct rx_connection
126 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
127 u_short serviceId, afs_uint32 cid,
128 afs_uint32 epoch, int type, u_int securityIndex);
129 static struct rx_packet
130 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
131 int istack, osi_socket socket,
132 afs_uint32 host, u_short port, int *tnop,
133 struct rx_call **newcallp);
134 static struct rx_packet
135 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
137 static struct rx_packet
138 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
139 struct rx_packet *np, int istack);
140 static struct rx_packet
141 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
142 struct rx_packet *np, int istack);
143 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
144 int *tnop, struct rx_call **newcallp);
145 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
146 static void rxi_ClearReceiveQueue(struct rx_call *call);
147 static void rxi_ResetCall(struct rx_call *call, int newcall);
148 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
149 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
150 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
151 static void rxi_KeepAliveOn(struct rx_call *call);
152 static void rxi_GrowMTUOn(struct rx_call *call);
153 static void rxi_ChallengeOn(struct rx_connection *conn);
154 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
155 static void rxi_AckAllInTransmitQueue(struct rx_call *call);
157 #ifdef RX_ENABLE_LOCKS
159 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
160 rx_atomic_t rxi_start_in_error;
162 #endif /* RX_ENABLE_LOCKS */
164 /* Constant delay time before sending an acknowledge of the last packet
165 * received. This is to avoid sending an extra acknowledge when the
166 * client is about to make another call, anyway, or the server is
169 * The lastAckDelay may not exceeed 400ms without causing peers to
170 * unecessarily timeout.
172 struct clock rx_lastAckDelay = {0, 400000};
174 /* Constant delay time before sending a soft ack when none was requested.
175 * This is to make sure we send soft acks before the sender times out,
176 * Normally we wait and send a hard ack when the receiver consumes the packet
178 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
179 * will require changes to the peer's RTT calculations.
181 struct clock rx_softAckDelay = {0, 100000};
184 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
185 * currently allocated within rx. This number is used to allocate the
186 * memory required to return the statistics when queried.
187 * Protected by the rx_rpc_stats mutex.
190 static unsigned int rxi_rpc_peer_stat_cnt;
193 * rxi_rpc_process_stat_cnt counts the total number of local process stat
194 * structures currently allocated within rx. The number is used to allocate
195 * the memory required to return the statistics when queried.
196 * Protected by the rx_rpc_stats mutex.
199 static unsigned int rxi_rpc_process_stat_cnt;
202 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
203 * errors should be reported to the application when a call channel appears busy
204 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
205 * and there are other call channels in the connection that are not busy.
206 * If 0, we do not return errors upon receiving busy packets; we just keep
207 * trying on the same call channel until we hit a timeout.
209 static afs_int32 rxi_busyChannelError = 0;
211 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
212 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
214 /* Incoming calls wait on this queue when there are no available
215 * server processes */
216 struct opr_queue rx_incomingCallQueue;
218 /* Server processes wait on this queue when there are no appropriate
219 * calls to process */
220 struct opr_queue rx_idleServerQueue;
222 #if !defined(offsetof)
223 #include <stddef.h> /* for definition of offsetof() */
226 #ifdef RX_ENABLE_LOCKS
227 afs_kmutex_t rx_atomic_mutex;
230 /* Forward prototypes */
231 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
234 putConnection (struct rx_connection *conn) {
235 MUTEX_ENTER(&rx_refcnt_mutex);
237 MUTEX_EXIT(&rx_refcnt_mutex);
240 #ifdef AFS_PTHREAD_ENV
243 * Use procedural initialization of mutexes/condition variables
247 extern afs_kmutex_t rx_quota_mutex;
248 extern afs_kmutex_t rx_pthread_mutex;
249 extern afs_kmutex_t rx_packets_mutex;
250 extern afs_kmutex_t rx_refcnt_mutex;
251 extern afs_kmutex_t des_init_mutex;
252 extern afs_kmutex_t des_random_mutex;
253 extern afs_kmutex_t rx_clock_mutex;
254 extern afs_kmutex_t rxi_connCacheMutex;
255 extern afs_kmutex_t event_handler_mutex;
256 extern afs_kmutex_t listener_mutex;
257 extern afs_kmutex_t rx_if_init_mutex;
258 extern afs_kmutex_t rx_if_mutex;
260 extern afs_kcondvar_t rx_event_handler_cond;
261 extern afs_kcondvar_t rx_listener_cond;
263 static afs_kmutex_t epoch_mutex;
264 static afs_kmutex_t rx_init_mutex;
265 static afs_kmutex_t rx_debug_mutex;
266 static afs_kmutex_t rx_rpc_stats;
269 rxi_InitPthread(void)
271 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
273 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
279 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
280 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
281 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
282 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
283 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
284 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
285 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
287 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
288 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
290 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
291 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
293 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
294 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
295 #ifdef RX_ENABLE_LOCKS
298 #endif /* RX_LOCKS_DB */
299 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
300 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
302 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
304 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
306 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
308 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
309 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
310 #endif /* RX_ENABLE_LOCKS */
313 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
314 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
316 * The rx_stats_mutex mutex protects the following global variables:
317 * rxi_lowConnRefCount
318 * rxi_lowPeerRefCount
327 * The rx_quota_mutex mutex protects the following global variables:
335 * The rx_freePktQ_lock protects the following global variables:
340 * The rx_packets_mutex mutex protects the following global variables:
348 * The rx_pthread_mutex mutex protects the following global variables:
349 * rxi_fcfs_thread_num
352 #define INIT_PTHREAD_LOCKS
356 /* Variables for handling the minProcs implementation. availProcs gives the
357 * number of threads available in the pool at this moment (not counting dudes
358 * executing right now). totalMin gives the total number of procs required
359 * for handling all minProcs requests. minDeficit is a dynamic variable
360 * tracking the # of procs required to satisfy all of the remaining minProcs
362 * For fine grain locking to work, the quota check and the reservation of
363 * a server thread has to come while rxi_availProcs and rxi_minDeficit
364 * are locked. To this end, the code has been modified under #ifdef
365 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
366 * same time. A new function, ReturnToServerPool() returns the allocation.
368 * A call can be on several queue's (but only one at a time). When
369 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
370 * that no one else is touching the queue. To this end, we store the address
371 * of the queue lock in the call structure (under the call lock) when we
372 * put the call on a queue, and we clear the call_queue_lock when the
373 * call is removed from a queue (once the call lock has been obtained).
374 * This allows rxi_ResetCall to safely synchronize with others wishing
375 * to manipulate the queue.
378 #if defined(RX_ENABLE_LOCKS)
379 static afs_kmutex_t rx_rpc_stats;
382 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
383 ** pretty good that the next packet coming in is from the same connection
384 ** as the last packet, since we're send multiple packets in a transmit window.
386 struct rx_connection *rxLastConn = 0;
388 #ifdef RX_ENABLE_LOCKS
389 /* The locking hierarchy for rx fine grain locking is composed of these
392 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
393 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
394 * call->lock - locks call data fields.
395 * These are independent of each other:
396 * rx_freeCallQueue_lock
401 * serverQueueEntry->lock
402 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
404 * peer->lock - locks peer data fields.
405 * conn_data_lock - that more than one thread is not updating a conn data
406 * field at the same time.
417 * Do we need a lock to protect the peer field in the conn structure?
418 * conn->peer was previously a constant for all intents and so has no
419 * lock protecting this field. The multihomed client delta introduced
420 * a RX code change : change the peer field in the connection structure
421 * to that remote interface from which the last packet for this
422 * connection was sent out. This may become an issue if further changes
425 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
426 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
428 /* rxdb_fileID is used to identify the lock location, along with line#. */
429 static int rxdb_fileID = RXDB_FILE_RX;
430 #endif /* RX_LOCKS_DB */
431 #else /* RX_ENABLE_LOCKS */
432 #define SET_CALL_QUEUE_LOCK(C, L)
433 #define CLEAR_CALL_QUEUE_LOCK(C)
434 #endif /* RX_ENABLE_LOCKS */
435 struct rx_serverQueueEntry *rx_waitForPacket = 0;
436 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
438 /* ------------Exported Interfaces------------- */
440 /* This function allows rxkad to set the epoch to a suitably random number
441 * which rx_NewConnection will use in the future. The principle purpose is to
442 * get rxnull connections to use the same epoch as the rxkad connections do, at
443 * least once the first rxkad connection is established. This is important now
444 * that the host/port addresses aren't used in FindConnection: the uniqueness
445 * of epoch/cid matters and the start time won't do. */
447 #ifdef AFS_PTHREAD_ENV
449 * This mutex protects the following global variables:
453 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
454 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
458 #endif /* AFS_PTHREAD_ENV */
461 rx_SetEpoch(afs_uint32 epoch)
468 /* Initialize rx. A port number may be mentioned, in which case this
469 * becomes the default port number for any service installed later.
470 * If 0 is provided for the port number, a random port will be chosen
471 * by the kernel. Whether this will ever overlap anything in
472 * /etc/services is anybody's guess... Returns 0 on success, -1 on
477 int rxinit_status = 1;
478 #ifdef AFS_PTHREAD_ENV
480 * This mutex protects the following global variables:
484 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
485 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
488 #define UNLOCK_RX_INIT
492 rx_InitHost(u_int host, u_int port)
499 char *htable, *ptable;
506 if (rxinit_status == 0) {
507 tmp_status = rxinit_status;
509 return tmp_status; /* Already started; return previous error code. */
515 if (afs_winsockInit() < 0)
521 * Initialize anything necessary to provide a non-premptive threading
524 rxi_InitializeThreadSupport();
527 /* Allocate and initialize a socket for client and perhaps server
530 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
531 if (rx_socket == OSI_NULLSOCKET) {
535 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
538 #endif /* RX_LOCKS_DB */
539 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
540 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
541 MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
542 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
543 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
544 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
545 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
546 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
547 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
548 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
550 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
552 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
554 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
556 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
557 #if defined(AFS_HPUX110_ENV)
559 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
560 #endif /* AFS_HPUX110_ENV */
561 #endif /* RX_ENABLE_LOCKS && KERNEL */
564 rx_connDeadTime = 12;
565 rx_tranquil = 0; /* reset flag */
566 rxi_ResetStatistics();
567 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
568 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
569 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
570 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
571 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
572 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
574 /* Malloc up a bunch of packets & buffers */
576 opr_queue_Init(&rx_freePacketQueue);
577 rxi_NeedMorePackets = FALSE;
578 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
580 /* enforce a minimum number of allocated packets */
581 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
582 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
584 /* allocate the initial free packet pool */
585 #ifdef RX_ENABLE_TSFPQ
586 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
587 #else /* RX_ENABLE_TSFPQ */
588 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
589 #endif /* RX_ENABLE_TSFPQ */
596 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
597 tv.tv_sec = clock_now.sec;
598 tv.tv_usec = clock_now.usec;
599 srand((unsigned int)tv.tv_usec);
606 #if defined(KERNEL) && !defined(UKERNEL)
607 /* Really, this should never happen in a real kernel */
610 struct sockaddr_in addr;
612 int addrlen = sizeof(addr);
614 socklen_t addrlen = sizeof(addr);
616 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
618 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
621 rx_port = addr.sin_port;
624 rx_stats.minRtt.sec = 9999999;
626 rx_SetEpoch(tv.tv_sec | 0x80000000);
628 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
629 * will provide a randomer value. */
631 MUTEX_ENTER(&rx_quota_mutex);
632 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
633 MUTEX_EXIT(&rx_quota_mutex);
634 /* *Slightly* random start time for the cid. This is just to help
635 * out with the hashing function at the peer */
636 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
637 rx_connHashTable = (struct rx_connection **)htable;
638 rx_peerHashTable = (struct rx_peer **)ptable;
640 rx_hardAckDelay.sec = 0;
641 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
643 rxevent_Init(20, rxi_ReScheduleEvents);
645 /* Initialize various global queues */
646 opr_queue_Init(&rx_idleServerQueue);
647 opr_queue_Init(&rx_incomingCallQueue);
648 opr_queue_Init(&rx_freeCallQueue);
650 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
651 /* Initialize our list of usable IP addresses. */
655 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
656 /* Start listener process (exact function is dependent on the
657 * implementation environment--kernel or user space) */
662 tmp_status = rxinit_status = 0;
670 return rx_InitHost(htonl(INADDR_ANY), port);
676 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
677 * maintaing the round trip timer.
682 * Start a new RTT timer for a given call and packet.
684 * There must be no resendEvent already listed for this call, otherwise this
685 * will leak events - intended for internal use within the RTO code only
688 * the RX call to start the timer for
689 * @param[in] lastPacket
690 * a flag indicating whether the last packet has been sent or not
692 * @pre call must be locked before calling this function
696 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
698 struct clock now, retryTime;
703 clock_Add(&retryTime, &call->rto);
705 /* If we're sending the last packet, and we're the client, then the server
706 * may wait for an additional 400ms before returning the ACK, wait for it
707 * rather than hitting a timeout */
708 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
709 clock_Addmsec(&retryTime, 400);
711 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
712 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
717 * Cancel an RTT timer for a given call.
721 * the RX call to cancel the timer for
723 * @pre call must be locked before calling this function
728 rxi_rto_cancel(struct rx_call *call)
730 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
734 * Tell the RTO timer that we have sent a packet.
736 * If the timer isn't already running, then start it. If the timer is running,
740 * the RX call that the packet has been sent on
741 * @param[in] lastPacket
742 * A flag which is true if this is the last packet for the call
744 * @pre The call must be locked before calling this function
749 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
751 if (call->resendEvent)
754 rxi_rto_startTimer(call, lastPacket, istack);
758 * Tell the RTO timer that we have received an new ACK message
760 * This function should be called whenever a call receives an ACK that
761 * acknowledges new packets. Whatever happens, we stop the current timer.
762 * If there are unacked packets in the queue which have been sent, then
763 * we restart the timer from now. Otherwise, we leave it stopped.
766 * the RX call that the ACK has been received on
770 rxi_rto_packet_acked(struct rx_call *call, int istack)
772 struct opr_queue *cursor;
774 rxi_rto_cancel(call);
776 if (opr_queue_IsEmpty(&call->tq))
779 for (opr_queue_Scan(&call->tq, cursor)) {
780 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
781 if (p->header.seq > call->tfirst + call->twind)
784 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
785 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
793 * Set an initial round trip timeout for a peer connection
795 * @param[in] secs The timeout to set in seconds
799 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
800 peer->rtt = secs * 8000;
804 * Enables or disables the busy call channel error (RX_CALL_BUSY).
806 * @param[in] onoff Non-zero to enable busy call channel errors.
808 * @pre Neither rx_Init nor rx_InitHost have been called yet
811 rx_SetBusyChannelError(afs_int32 onoff)
813 osi_Assert(rxinit_status != 0);
814 rxi_busyChannelError = onoff ? 1 : 0;
818 * Set a delayed ack event on the specified call for the given time
820 * @param[in] call - the call on which to set the event
821 * @param[in] offset - the delay from now after which the event fires
824 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
826 struct clock now, when;
830 clock_Add(&when, offset);
832 if (!call->delayedAckEvent
833 || clock_Gt(&call->delayedAckTime, &when)) {
835 rxevent_Cancel(&call->delayedAckEvent, call,
836 RX_CALL_REFCOUNT_DELAY);
837 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
839 call->delayedAckEvent = rxevent_Post(&when, &now,
842 call->delayedAckTime = when;
846 /* called with unincremented nRequestsRunning to see if it is OK to start
847 * a new thread in this service. Could be "no" for two reasons: over the
848 * max quota, or would prevent others from reaching their min quota.
850 #ifdef RX_ENABLE_LOCKS
851 /* This verion of QuotaOK reserves quota if it's ok while the
852 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
855 QuotaOK(struct rx_service *aservice)
857 /* check if over max quota */
858 if (aservice->nRequestsRunning >= aservice->maxProcs) {
862 /* under min quota, we're OK */
863 /* otherwise, can use only if there are enough to allow everyone
864 * to go to their min quota after this guy starts.
867 MUTEX_ENTER(&rx_quota_mutex);
868 if ((aservice->nRequestsRunning < aservice->minProcs)
869 || (rxi_availProcs > rxi_minDeficit)) {
870 aservice->nRequestsRunning++;
871 /* just started call in minProcs pool, need fewer to maintain
873 if (aservice->nRequestsRunning <= aservice->minProcs)
876 MUTEX_EXIT(&rx_quota_mutex);
879 MUTEX_EXIT(&rx_quota_mutex);
885 ReturnToServerPool(struct rx_service *aservice)
887 aservice->nRequestsRunning--;
888 MUTEX_ENTER(&rx_quota_mutex);
889 if (aservice->nRequestsRunning < aservice->minProcs)
892 MUTEX_EXIT(&rx_quota_mutex);
895 #else /* RX_ENABLE_LOCKS */
897 QuotaOK(struct rx_service *aservice)
900 /* under min quota, we're OK */
901 if (aservice->nRequestsRunning < aservice->minProcs)
904 /* check if over max quota */
905 if (aservice->nRequestsRunning >= aservice->maxProcs)
908 /* otherwise, can use only if there are enough to allow everyone
909 * to go to their min quota after this guy starts.
911 MUTEX_ENTER(&rx_quota_mutex);
912 if (rxi_availProcs > rxi_minDeficit)
914 MUTEX_EXIT(&rx_quota_mutex);
917 #endif /* RX_ENABLE_LOCKS */
920 /* Called by rx_StartServer to start up lwp's to service calls.
921 NExistingProcs gives the number of procs already existing, and which
922 therefore needn't be created. */
924 rxi_StartServerProcs(int nExistingProcs)
926 struct rx_service *service;
931 /* For each service, reserve N processes, where N is the "minimum"
932 * number of processes that MUST be able to execute a request in parallel,
933 * at any time, for that process. Also compute the maximum difference
934 * between any service's maximum number of processes that can run
935 * (i.e. the maximum number that ever will be run, and a guarantee
936 * that this number will run if other services aren't running), and its
937 * minimum number. The result is the extra number of processes that
938 * we need in order to provide the latter guarantee */
939 for (i = 0; i < RX_MAX_SERVICES; i++) {
941 service = rx_services[i];
942 if (service == (struct rx_service *)0)
944 nProcs += service->minProcs;
945 diff = service->maxProcs - service->minProcs;
949 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
950 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
951 for (i = 0; i < nProcs; i++) {
952 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
958 /* This routine is only required on Windows */
960 rx_StartClientThread(void)
962 #ifdef AFS_PTHREAD_ENV
964 pid = pthread_self();
965 #endif /* AFS_PTHREAD_ENV */
967 #endif /* AFS_NT40_ENV */
969 /* This routine must be called if any services are exported. If the
970 * donateMe flag is set, the calling process is donated to the server
973 rx_StartServer(int donateMe)
975 struct rx_service *service;
981 /* Start server processes, if necessary (exact function is dependent
982 * on the implementation environment--kernel or user space). DonateMe
983 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
984 * case, one less new proc will be created rx_StartServerProcs.
986 rxi_StartServerProcs(donateMe);
988 /* count up the # of threads in minProcs, and add set the min deficit to
989 * be that value, too.
991 for (i = 0; i < RX_MAX_SERVICES; i++) {
992 service = rx_services[i];
993 if (service == (struct rx_service *)0)
995 MUTEX_ENTER(&rx_quota_mutex);
996 rxi_totalMin += service->minProcs;
997 /* below works even if a thread is running, since minDeficit would
998 * still have been decremented and later re-incremented.
1000 rxi_minDeficit += service->minProcs;
1001 MUTEX_EXIT(&rx_quota_mutex);
1004 /* Turn on reaping of idle server connections */
1005 rxi_ReapConnections(NULL, NULL, NULL, 0);
1010 #ifndef AFS_NT40_ENV
1014 #ifdef AFS_PTHREAD_ENV
1016 pid = afs_pointer_to_int(pthread_self());
1017 #else /* AFS_PTHREAD_ENV */
1019 LWP_CurrentProcess(&pid);
1020 #endif /* AFS_PTHREAD_ENV */
1022 sprintf(name, "srv_%d", ++nProcs);
1023 if (registerProgram)
1024 (*registerProgram) (pid, name);
1026 #endif /* AFS_NT40_ENV */
1027 rx_ServerProc(NULL); /* Never returns */
1029 #ifdef RX_ENABLE_TSFPQ
1030 /* no use leaving packets around in this thread's local queue if
1031 * it isn't getting donated to the server thread pool.
1033 rxi_FlushLocalPacketsTSFPQ();
1034 #endif /* RX_ENABLE_TSFPQ */
1038 /* Create a new client connection to the specified service, using the
1039 * specified security object to implement the security model for this
1041 struct rx_connection *
1042 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1043 struct rx_securityClass *securityObject,
1044 int serviceSecurityIndex)
1048 struct rx_connection *conn;
1053 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1054 "serviceSecurityIndex %d)\n",
1055 ntohl(shost), ntohs(sport), sservice, securityObject,
1056 serviceSecurityIndex));
1058 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1059 * the case of kmem_alloc? */
1060 conn = rxi_AllocConnection();
1061 #ifdef RX_ENABLE_LOCKS
1062 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1063 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1064 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1067 MUTEX_ENTER(&rx_connHashTable_lock);
1068 cid = (rx_nextCid += RX_MAXCALLS);
1069 conn->type = RX_CLIENT_CONNECTION;
1071 conn->epoch = rx_epoch;
1072 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1073 conn->serviceId = sservice;
1074 conn->securityObject = securityObject;
1075 conn->securityData = (void *) 0;
1076 conn->securityIndex = serviceSecurityIndex;
1077 rx_SetConnDeadTime(conn, rx_connDeadTime);
1078 rx_SetConnSecondsUntilNatPing(conn, 0);
1079 conn->ackRate = RX_FAST_ACK_RATE;
1080 conn->nSpecific = 0;
1081 conn->specific = NULL;
1082 conn->challengeEvent = NULL;
1083 conn->delayedAbortEvent = NULL;
1084 conn->abortCount = 0;
1086 for (i = 0; i < RX_MAXCALLS; i++) {
1087 conn->twind[i] = rx_initSendWindow;
1088 conn->rwind[i] = rx_initReceiveWindow;
1089 conn->lastBusy[i] = 0;
1092 RXS_NewConnection(securityObject, conn);
1094 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1096 conn->refCount++; /* no lock required since only this thread knows... */
1097 conn->next = rx_connHashTable[hashindex];
1098 rx_connHashTable[hashindex] = conn;
1099 if (rx_stats_active)
1100 rx_atomic_inc(&rx_stats.nClientConns);
1101 MUTEX_EXIT(&rx_connHashTable_lock);
1107 * Ensure a connection's timeout values are valid.
1109 * @param[in] conn The connection to check
1111 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1112 * unless idleDeadTime and/or hardDeadTime are not set
1116 rxi_CheckConnTimeouts(struct rx_connection *conn)
1118 /* a connection's timeouts must have the relationship
1119 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1120 * total loss of network to a peer may cause an idle timeout instead of a
1121 * dead timeout, simply because the idle timeout gets hit first. Also set
1122 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1123 /* this logic is slightly complicated by the fact that
1124 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1126 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1127 if (conn->idleDeadTime) {
1128 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1130 if (conn->hardDeadTime) {
1131 if (conn->idleDeadTime) {
1132 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1134 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1140 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1142 /* The idea is to set the dead time to a value that allows several
1143 * keepalives to be dropped without timing out the connection. */
1144 conn->secondsUntilDead = seconds;
1145 rxi_CheckConnTimeouts(conn);
1146 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1150 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1152 conn->hardDeadTime = seconds;
1153 rxi_CheckConnTimeouts(conn);
1157 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1159 conn->idleDeadTime = seconds;
1160 conn->idleDeadDetection = (seconds ? 1 : 0);
1161 rxi_CheckConnTimeouts(conn);
1164 int rxi_lowPeerRefCount = 0;
1165 int rxi_lowConnRefCount = 0;
1168 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1169 * NOTE: must not be called with rx_connHashTable_lock held.
1172 rxi_CleanupConnection(struct rx_connection *conn)
1174 /* Notify the service exporter, if requested, that this connection
1175 * is being destroyed */
1176 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1177 (*conn->service->destroyConnProc) (conn);
1179 /* Notify the security module that this connection is being destroyed */
1180 RXS_DestroyConnection(conn->securityObject, conn);
1182 /* If this is the last connection using the rx_peer struct, set its
1183 * idle time to now. rxi_ReapConnections will reap it if it's still
1184 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1186 MUTEX_ENTER(&rx_peerHashTable_lock);
1187 if (conn->peer->refCount < 2) {
1188 conn->peer->idleWhen = clock_Sec();
1189 if (conn->peer->refCount < 1) {
1190 conn->peer->refCount = 1;
1191 if (rx_stats_active) {
1192 MUTEX_ENTER(&rx_stats_mutex);
1193 rxi_lowPeerRefCount++;
1194 MUTEX_EXIT(&rx_stats_mutex);
1198 conn->peer->refCount--;
1199 MUTEX_EXIT(&rx_peerHashTable_lock);
1201 if (rx_stats_active)
1203 if (conn->type == RX_SERVER_CONNECTION)
1204 rx_atomic_dec(&rx_stats.nServerConns);
1206 rx_atomic_dec(&rx_stats.nClientConns);
1209 if (conn->specific) {
1211 for (i = 0; i < conn->nSpecific; i++) {
1212 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1213 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1214 conn->specific[i] = NULL;
1216 free(conn->specific);
1218 conn->specific = NULL;
1219 conn->nSpecific = 0;
1220 #endif /* !KERNEL */
1222 MUTEX_DESTROY(&conn->conn_call_lock);
1223 MUTEX_DESTROY(&conn->conn_data_lock);
1224 CV_DESTROY(&conn->conn_call_cv);
1226 rxi_FreeConnection(conn);
1229 /* Destroy the specified connection */
1231 rxi_DestroyConnection(struct rx_connection *conn)
1233 MUTEX_ENTER(&rx_connHashTable_lock);
1234 rxi_DestroyConnectionNoLock(conn);
1235 /* conn should be at the head of the cleanup list */
1236 if (conn == rx_connCleanup_list) {
1237 rx_connCleanup_list = rx_connCleanup_list->next;
1238 MUTEX_EXIT(&rx_connHashTable_lock);
1239 rxi_CleanupConnection(conn);
1241 #ifdef RX_ENABLE_LOCKS
1243 MUTEX_EXIT(&rx_connHashTable_lock);
1245 #endif /* RX_ENABLE_LOCKS */
1249 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1251 struct rx_connection **conn_ptr;
1253 struct rx_packet *packet;
1260 MUTEX_ENTER(&conn->conn_data_lock);
1261 MUTEX_ENTER(&rx_refcnt_mutex);
1262 if (conn->refCount > 0)
1265 if (rx_stats_active) {
1266 MUTEX_ENTER(&rx_stats_mutex);
1267 rxi_lowConnRefCount++;
1268 MUTEX_EXIT(&rx_stats_mutex);
1272 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1273 /* Busy; wait till the last guy before proceeding */
1274 MUTEX_EXIT(&rx_refcnt_mutex);
1275 MUTEX_EXIT(&conn->conn_data_lock);
1280 /* If the client previously called rx_NewCall, but it is still
1281 * waiting, treat this as a running call, and wait to destroy the
1282 * connection later when the call completes. */
1283 if ((conn->type == RX_CLIENT_CONNECTION)
1284 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1285 conn->flags |= RX_CONN_DESTROY_ME;
1286 MUTEX_EXIT(&conn->conn_data_lock);
1290 MUTEX_EXIT(&rx_refcnt_mutex);
1291 MUTEX_EXIT(&conn->conn_data_lock);
1293 /* Check for extant references to this connection */
1294 MUTEX_ENTER(&conn->conn_call_lock);
1295 for (i = 0; i < RX_MAXCALLS; i++) {
1296 struct rx_call *call = conn->call[i];
1299 if (conn->type == RX_CLIENT_CONNECTION) {
1300 MUTEX_ENTER(&call->lock);
1301 if (call->delayedAckEvent) {
1302 /* Push the final acknowledgment out now--there
1303 * won't be a subsequent call to acknowledge the
1304 * last reply packets */
1305 rxevent_Cancel(&call->delayedAckEvent, call,
1306 RX_CALL_REFCOUNT_DELAY);
1307 if (call->state == RX_STATE_PRECALL
1308 || call->state == RX_STATE_ACTIVE) {
1309 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1314 MUTEX_EXIT(&call->lock);
1318 MUTEX_EXIT(&conn->conn_call_lock);
1320 #ifdef RX_ENABLE_LOCKS
1322 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1323 MUTEX_EXIT(&conn->conn_data_lock);
1325 /* Someone is accessing a packet right now. */
1329 #endif /* RX_ENABLE_LOCKS */
1332 /* Don't destroy the connection if there are any call
1333 * structures still in use */
1334 MUTEX_ENTER(&conn->conn_data_lock);
1335 conn->flags |= RX_CONN_DESTROY_ME;
1336 MUTEX_EXIT(&conn->conn_data_lock);
1341 if (conn->natKeepAliveEvent) {
1342 rxi_NatKeepAliveOff(conn);
1345 if (conn->delayedAbortEvent) {
1346 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1347 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1349 MUTEX_ENTER(&conn->conn_data_lock);
1350 rxi_SendConnectionAbort(conn, packet, 0, 1);
1351 MUTEX_EXIT(&conn->conn_data_lock);
1352 rxi_FreePacket(packet);
1356 /* Remove from connection hash table before proceeding */
1358 &rx_connHashTable[CONN_HASH
1359 (peer->host, peer->port, conn->cid, conn->epoch,
1361 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1362 if (*conn_ptr == conn) {
1363 *conn_ptr = conn->next;
1367 /* if the conn that we are destroying was the last connection, then we
1368 * clear rxLastConn as well */
1369 if (rxLastConn == conn)
1372 /* Make sure the connection is completely reset before deleting it. */
1373 /* get rid of pending events that could zap us later */
1374 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1375 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1376 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1378 /* Add the connection to the list of destroyed connections that
1379 * need to be cleaned up. This is necessary to avoid deadlocks
1380 * in the routines we call to inform others that this connection is
1381 * being destroyed. */
1382 conn->next = rx_connCleanup_list;
1383 rx_connCleanup_list = conn;
1386 /* Externally available version */
1388 rx_DestroyConnection(struct rx_connection *conn)
1393 rxi_DestroyConnection(conn);
1398 rx_GetConnection(struct rx_connection *conn)
1403 MUTEX_ENTER(&rx_refcnt_mutex);
1405 MUTEX_EXIT(&rx_refcnt_mutex);
1409 #ifdef RX_ENABLE_LOCKS
1410 /* Wait for the transmit queue to no longer be busy.
1411 * requires the call->lock to be held */
1413 rxi_WaitforTQBusy(struct rx_call *call) {
1414 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1415 call->flags |= RX_CALL_TQ_WAIT;
1417 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1418 CV_WAIT(&call->cv_tq, &call->lock);
1420 if (call->tqWaiters == 0) {
1421 call->flags &= ~RX_CALL_TQ_WAIT;
1428 rxi_WakeUpTransmitQueue(struct rx_call *call)
1430 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1431 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1432 call, call->tqWaiters, call->flags));
1433 #ifdef RX_ENABLE_LOCKS
1434 osirx_AssertMine(&call->lock, "rxi_Start start");
1435 CV_BROADCAST(&call->cv_tq);
1436 #else /* RX_ENABLE_LOCKS */
1437 osi_rxWakeup(&call->tq);
1438 #endif /* RX_ENABLE_LOCKS */
1442 /* Start a new rx remote procedure call, on the specified connection.
1443 * If wait is set to 1, wait for a free call channel; otherwise return
1444 * 0. Maxtime gives the maximum number of seconds this call may take,
1445 * after rx_NewCall returns. After this time interval, a call to any
1446 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1447 * For fine grain locking, we hold the conn_call_lock in order to
1448 * to ensure that we don't get signalle after we found a call in an active
1449 * state and before we go to sleep.
1452 rx_NewCall(struct rx_connection *conn)
1454 int i, wait, ignoreBusy = 1;
1455 struct rx_call *call;
1456 struct clock queueTime;
1457 afs_uint32 leastBusy = 0;
1461 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1464 clock_GetTime(&queueTime);
1466 * Check if there are others waiting for a new call.
1467 * If so, let them go first to avoid starving them.
1468 * This is a fairly simple scheme, and might not be
1469 * a complete solution for large numbers of waiters.
1471 * makeCallWaiters keeps track of the number of
1472 * threads waiting to make calls and the
1473 * RX_CONN_MAKECALL_WAITING flag bit is used to
1474 * indicate that there are indeed calls waiting.
1475 * The flag is set when the waiter is incremented.
1476 * It is only cleared when makeCallWaiters is 0.
1477 * This prevents us from accidently destroying the
1478 * connection while it is potentially about to be used.
1480 MUTEX_ENTER(&conn->conn_call_lock);
1481 MUTEX_ENTER(&conn->conn_data_lock);
1482 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1483 conn->flags |= RX_CONN_MAKECALL_WAITING;
1484 conn->makeCallWaiters++;
1485 MUTEX_EXIT(&conn->conn_data_lock);
1487 #ifdef RX_ENABLE_LOCKS
1488 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1492 MUTEX_ENTER(&conn->conn_data_lock);
1493 conn->makeCallWaiters--;
1494 if (conn->makeCallWaiters == 0)
1495 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1498 /* We are now the active thread in rx_NewCall */
1499 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1500 MUTEX_EXIT(&conn->conn_data_lock);
1505 for (i = 0; i < RX_MAXCALLS; i++) {
1506 call = conn->call[i];
1508 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1509 /* we're not ignoring busy call slots; only look at the
1510 * call slot that is the "least" busy */
1514 if (call->state == RX_STATE_DALLY) {
1515 MUTEX_ENTER(&call->lock);
1516 if (call->state == RX_STATE_DALLY) {
1517 if (ignoreBusy && conn->lastBusy[i]) {
1518 /* if we're ignoring busy call slots, skip any ones that
1519 * have lastBusy set */
1520 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1521 leastBusy = conn->lastBusy[i];
1523 MUTEX_EXIT(&call->lock);
1528 * We are setting the state to RX_STATE_RESET to
1529 * ensure that no one else will attempt to use this
1530 * call once we drop the conn->conn_call_lock and
1531 * call->lock. We must drop the conn->conn_call_lock
1532 * before calling rxi_ResetCall because the process
1533 * of clearing the transmit queue can block for an
1534 * extended period of time. If we block while holding
1535 * the conn->conn_call_lock, then all rx_EndCall
1536 * processing will block as well. This has a detrimental
1537 * effect on overall system performance.
1539 call->state = RX_STATE_RESET;
1540 (*call->callNumber)++;
1541 MUTEX_EXIT(&conn->conn_call_lock);
1542 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1543 rxi_ResetCall(call, 0);
1544 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1548 * If we failed to be able to safely obtain the
1549 * conn->conn_call_lock we will have to drop the
1550 * call->lock to avoid a deadlock. When the call->lock
1551 * is released the state of the call can change. If it
1552 * is no longer RX_STATE_RESET then some other thread is
1555 MUTEX_EXIT(&call->lock);
1556 MUTEX_ENTER(&conn->conn_call_lock);
1557 MUTEX_ENTER(&call->lock);
1559 if (call->state == RX_STATE_RESET)
1563 * If we get here it means that after dropping
1564 * the conn->conn_call_lock and call->lock that
1565 * the call is no longer ours. If we can't find
1566 * a free call in the remaining slots we should
1567 * not go immediately to RX_CONN_MAKECALL_WAITING
1568 * because by dropping the conn->conn_call_lock
1569 * we have given up synchronization with rx_EndCall.
1570 * Instead, cycle through one more time to see if
1571 * we can find a call that can call our own.
1573 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1576 MUTEX_EXIT(&call->lock);
1579 if (ignoreBusy && conn->lastBusy[i]) {
1580 /* if we're ignoring busy call slots, skip any ones that
1581 * have lastBusy set */
1582 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1583 leastBusy = conn->lastBusy[i];
1588 /* rxi_NewCall returns with mutex locked */
1589 call = rxi_NewCall(conn, i);
1590 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1594 if (i < RX_MAXCALLS) {
1595 conn->lastBusy[i] = 0;
1596 call->flags &= ~RX_CALL_PEER_BUSY;
1601 if (leastBusy && ignoreBusy) {
1602 /* we didn't find a useable call slot, but we did see at least one
1603 * 'busy' slot; look again and only use a slot with the 'least
1609 MUTEX_ENTER(&conn->conn_data_lock);
1610 conn->flags |= RX_CONN_MAKECALL_WAITING;
1611 conn->makeCallWaiters++;
1612 MUTEX_EXIT(&conn->conn_data_lock);
1614 #ifdef RX_ENABLE_LOCKS
1615 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1619 MUTEX_ENTER(&conn->conn_data_lock);
1620 conn->makeCallWaiters--;
1621 if (conn->makeCallWaiters == 0)
1622 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1623 MUTEX_EXIT(&conn->conn_data_lock);
1625 /* Client is initially in send mode */
1626 call->state = RX_STATE_ACTIVE;
1627 call->error = conn->error;
1629 call->app.mode = RX_MODE_ERROR;
1631 call->app.mode = RX_MODE_SENDING;
1633 #ifdef AFS_RXERRQ_ENV
1634 /* remember how many network errors the peer has when we started, so if
1635 * more errors are encountered after the call starts, we know the other endpoint won't be
1636 * responding to us */
1637 call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
1640 /* remember start time for call in case we have hard dead time limit */
1641 call->queueTime = queueTime;
1642 clock_GetTime(&call->startTime);
1643 call->app.bytesSent = 0;
1644 call->app.bytesRcvd = 0;
1646 /* Turn on busy protocol. */
1647 rxi_KeepAliveOn(call);
1649 /* Attempt MTU discovery */
1650 rxi_GrowMTUOn(call);
1653 * We are no longer the active thread in rx_NewCall
1655 MUTEX_ENTER(&conn->conn_data_lock);
1656 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1657 MUTEX_EXIT(&conn->conn_data_lock);
1660 * Wake up anyone else who might be giving us a chance to
1661 * run (see code above that avoids resource starvation).
1663 #ifdef RX_ENABLE_LOCKS
1664 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1665 osi_Panic("rx_NewCall call about to be used without an empty tq");
1668 CV_BROADCAST(&conn->conn_call_cv);
1672 MUTEX_EXIT(&conn->conn_call_lock);
1673 MUTEX_EXIT(&call->lock);
1676 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1681 rxi_HasActiveCalls(struct rx_connection *aconn)
1684 struct rx_call *tcall;
1688 for (i = 0; i < RX_MAXCALLS; i++) {
1689 if ((tcall = aconn->call[i])) {
1690 if ((tcall->state == RX_STATE_ACTIVE)
1691 || (tcall->state == RX_STATE_PRECALL)) {
1702 rxi_GetCallNumberVector(struct rx_connection *aconn,
1703 afs_int32 * aint32s)
1706 struct rx_call *tcall;
1710 MUTEX_ENTER(&aconn->conn_call_lock);
1711 for (i = 0; i < RX_MAXCALLS; i++) {
1712 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1713 aint32s[i] = aconn->callNumber[i] + 1;
1715 aint32s[i] = aconn->callNumber[i];
1717 MUTEX_EXIT(&aconn->conn_call_lock);
1723 rxi_SetCallNumberVector(struct rx_connection *aconn,
1724 afs_int32 * aint32s)
1727 struct rx_call *tcall;
1731 MUTEX_ENTER(&aconn->conn_call_lock);
1732 for (i = 0; i < RX_MAXCALLS; i++) {
1733 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1734 aconn->callNumber[i] = aint32s[i] - 1;
1736 aconn->callNumber[i] = aint32s[i];
1738 MUTEX_EXIT(&aconn->conn_call_lock);
1743 /* Advertise a new service. A service is named locally by a UDP port
1744 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1747 char *serviceName; Name for identification purposes (e.g. the
1748 service name might be used for probing for
1751 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1752 char *serviceName, struct rx_securityClass **securityObjects,
1753 int nSecurityObjects,
1754 afs_int32(*serviceProc) (struct rx_call * acall))
1756 osi_socket socket = OSI_NULLSOCKET;
1757 struct rx_service *tservice;
1763 if (serviceId == 0) {
1765 "rx_NewService: service id for service %s is not non-zero.\n",
1772 "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
1780 tservice = rxi_AllocService();
1783 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1785 for (i = 0; i < RX_MAX_SERVICES; i++) {
1786 struct rx_service *service = rx_services[i];
1788 if (port == service->servicePort && host == service->serviceHost) {
1789 if (service->serviceId == serviceId) {
1790 /* The identical service has already been
1791 * installed; if the caller was intending to
1792 * change the security classes used by this
1793 * service, he/she loses. */
1795 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1796 serviceName, serviceId, service->serviceName);
1798 rxi_FreeService(tservice);
1801 /* Different service, same port: re-use the socket
1802 * which is bound to the same port */
1803 socket = service->socket;
1806 if (socket == OSI_NULLSOCKET) {
1807 /* If we don't already have a socket (from another
1808 * service on same port) get a new one */
1809 socket = rxi_GetHostUDPSocket(host, port);
1810 if (socket == OSI_NULLSOCKET) {
1812 rxi_FreeService(tservice);
1817 service->socket = socket;
1818 service->serviceHost = host;
1819 service->servicePort = port;
1820 service->serviceId = serviceId;
1821 service->serviceName = serviceName;
1822 service->nSecurityObjects = nSecurityObjects;
1823 service->securityObjects = securityObjects;
1824 service->minProcs = 0;
1825 service->maxProcs = 1;
1826 service->idleDeadTime = 60;
1827 service->idleDeadErr = 0;
1828 service->connDeadTime = rx_connDeadTime;
1829 service->executeRequestProc = serviceProc;
1830 service->checkReach = 0;
1831 service->nSpecific = 0;
1832 service->specific = NULL;
1833 rx_services[i] = service; /* not visible until now */
1839 rxi_FreeService(tservice);
1840 (osi_Msg "rx_NewService: cannot support > %d services\n",
1845 /* Set configuration options for all of a service's security objects */
1848 rx_SetSecurityConfiguration(struct rx_service *service,
1849 rx_securityConfigVariables type,
1853 for (i = 0; i<service->nSecurityObjects; i++) {
1854 if (service->securityObjects[i]) {
1855 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1863 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1864 struct rx_securityClass **securityObjects, int nSecurityObjects,
1865 afs_int32(*serviceProc) (struct rx_call * acall))
1867 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1870 /* Generic request processing loop. This routine should be called
1871 * by the implementation dependent rx_ServerProc. If socketp is
1872 * non-null, it will be set to the file descriptor that this thread
1873 * is now listening on. If socketp is null, this routine will never
1876 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1878 struct rx_call *call;
1880 struct rx_service *tservice = NULL;
1887 call = rx_GetCall(threadID, tservice, socketp);
1888 if (socketp && *socketp != OSI_NULLSOCKET) {
1889 /* We are now a listener thread */
1895 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1896 #ifdef RX_ENABLE_LOCKS
1898 #endif /* RX_ENABLE_LOCKS */
1899 afs_termState = AFSOP_STOP_AFS;
1900 afs_osi_Wakeup(&afs_termState);
1901 #ifdef RX_ENABLE_LOCKS
1903 #endif /* RX_ENABLE_LOCKS */
1908 /* if server is restarting( typically smooth shutdown) then do not
1909 * allow any new calls.
1912 if (rx_tranquil && (call != NULL)) {
1916 MUTEX_ENTER(&call->lock);
1918 rxi_CallError(call, RX_RESTARTING);
1919 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1921 MUTEX_EXIT(&call->lock);
1926 tservice = call->conn->service;
1928 if (tservice->beforeProc)
1929 (*tservice->beforeProc) (call);
1931 code = tservice->executeRequestProc(call);
1933 if (tservice->afterProc)
1934 (*tservice->afterProc) (call, code);
1936 rx_EndCall(call, code);
1938 if (tservice->postProc)
1939 (*tservice->postProc) (code);
1941 if (rx_stats_active) {
1942 MUTEX_ENTER(&rx_stats_mutex);
1944 MUTEX_EXIT(&rx_stats_mutex);
1951 rx_WakeupServerProcs(void)
1953 struct rx_serverQueueEntry *np, *tqp;
1954 struct opr_queue *cursor;
1958 MUTEX_ENTER(&rx_serverPool_lock);
1960 #ifdef RX_ENABLE_LOCKS
1961 if (rx_waitForPacket)
1962 CV_BROADCAST(&rx_waitForPacket->cv);
1963 #else /* RX_ENABLE_LOCKS */
1964 if (rx_waitForPacket)
1965 osi_rxWakeup(rx_waitForPacket);
1966 #endif /* RX_ENABLE_LOCKS */
1967 MUTEX_ENTER(&freeSQEList_lock);
1968 for (np = rx_FreeSQEList; np; np = tqp) {
1969 tqp = *(struct rx_serverQueueEntry **)np;
1970 #ifdef RX_ENABLE_LOCKS
1971 CV_BROADCAST(&np->cv);
1972 #else /* RX_ENABLE_LOCKS */
1974 #endif /* RX_ENABLE_LOCKS */
1976 MUTEX_EXIT(&freeSQEList_lock);
1977 for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
1978 np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
1979 #ifdef RX_ENABLE_LOCKS
1980 CV_BROADCAST(&np->cv);
1981 #else /* RX_ENABLE_LOCKS */
1983 #endif /* RX_ENABLE_LOCKS */
1985 MUTEX_EXIT(&rx_serverPool_lock);
1990 * One thing that seems to happen is that all the server threads get
1991 * tied up on some empty or slow call, and then a whole bunch of calls
1992 * arrive at once, using up the packet pool, so now there are more
1993 * empty calls. The most critical resources here are server threads
1994 * and the free packet pool. The "doreclaim" code seems to help in
1995 * general. I think that eventually we arrive in this state: there
1996 * are lots of pending calls which do have all their packets present,
1997 * so they won't be reclaimed, are multi-packet calls, so they won't
1998 * be scheduled until later, and thus are tying up most of the free
1999 * packet pool for a very long time.
2001 * 1. schedule multi-packet calls if all the packets are present.
2002 * Probably CPU-bound operation, useful to return packets to pool.
2003 * Do what if there is a full window, but the last packet isn't here?
2004 * 3. preserve one thread which *only* runs "best" calls, otherwise
2005 * it sleeps and waits for that type of call.
2006 * 4. Don't necessarily reserve a whole window for each thread. In fact,
2007 * the current dataquota business is badly broken. The quota isn't adjusted
2008 * to reflect how many packets are presently queued for a running call.
2009 * So, when we schedule a queued call with a full window of packets queued
2010 * up for it, that *should* free up a window full of packets for other 2d-class
2011 * calls to be able to use from the packet pool. But it doesn't.
2013 * NB. Most of the time, this code doesn't run -- since idle server threads
2014 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2015 * as a new call arrives.
2017 /* Sleep until a call arrives. Returns a pointer to the call, ready
2018 * for an rx_Read. */
2019 #ifdef RX_ENABLE_LOCKS
2021 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2023 struct rx_serverQueueEntry *sq;
2024 struct rx_call *call = (struct rx_call *)0;
2025 struct rx_service *service = NULL;
2027 MUTEX_ENTER(&freeSQEList_lock);
2029 if ((sq = rx_FreeSQEList)) {
2030 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2031 MUTEX_EXIT(&freeSQEList_lock);
2032 } else { /* otherwise allocate a new one and return that */
2033 MUTEX_EXIT(&freeSQEList_lock);
2034 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2035 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2036 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2039 MUTEX_ENTER(&rx_serverPool_lock);
2040 if (cur_service != NULL) {
2041 ReturnToServerPool(cur_service);
2044 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2045 struct rx_call *tcall, *choice2 = NULL;
2046 struct opr_queue *cursor;
2048 /* Scan for eligible incoming calls. A call is not eligible
2049 * if the maximum number of calls for its service type are
2050 * already executing */
2051 /* One thread will process calls FCFS (to prevent starvation),
2052 * while the other threads may run ahead looking for calls which
2053 * have all their input data available immediately. This helps
2054 * keep threads from blocking, waiting for data from the client. */
2055 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2056 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2058 service = tcall->conn->service;
2059 if (!QuotaOK(service)) {
2062 MUTEX_ENTER(&rx_pthread_mutex);
2063 if (tno == rxi_fcfs_thread_num
2064 || opr_queue_IsEnd(&rx_incomingCallQueue, cursor)) {
2065 MUTEX_EXIT(&rx_pthread_mutex);
2066 /* If we're the fcfs thread , then we'll just use
2067 * this call. If we haven't been able to find an optimal
2068 * choice, and we're at the end of the list, then use a
2069 * 2d choice if one has been identified. Otherwise... */
2070 call = (choice2 ? choice2 : tcall);
2071 service = call->conn->service;
2073 MUTEX_EXIT(&rx_pthread_mutex);
2074 if (!opr_queue_IsEmpty(&tcall->rq)) {
2075 struct rx_packet *rp;
2076 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2078 if (rp->header.seq == 1) {
2080 || (rp->header.flags & RX_LAST_PACKET)) {
2082 } else if (rxi_2dchoice && !choice2
2083 && !(tcall->flags & RX_CALL_CLEARED)
2084 && (tcall->rprev > rxi_HardAckRate)) {
2094 ReturnToServerPool(service);
2100 opr_queue_Remove(&call->entry);
2101 MUTEX_EXIT(&rx_serverPool_lock);
2102 MUTEX_ENTER(&call->lock);
2104 if (call->flags & RX_CALL_WAIT_PROC) {
2105 call->flags &= ~RX_CALL_WAIT_PROC;
2106 rx_atomic_dec(&rx_nWaiting);
2109 if (call->state != RX_STATE_PRECALL || call->error) {
2110 MUTEX_EXIT(&call->lock);
2111 MUTEX_ENTER(&rx_serverPool_lock);
2112 ReturnToServerPool(service);
2117 if (opr_queue_IsEmpty(&call->rq)
2118 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
2119 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2121 CLEAR_CALL_QUEUE_LOCK(call);
2124 /* If there are no eligible incoming calls, add this process
2125 * to the idle server queue, to wait for one */
2129 *socketp = OSI_NULLSOCKET;
2131 sq->socketp = socketp;
2132 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2133 #ifndef AFS_AIX41_ENV
2134 rx_waitForPacket = sq;
2136 rx_waitingForPacket = sq;
2137 #endif /* AFS_AIX41_ENV */
2139 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2141 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2142 MUTEX_EXIT(&rx_serverPool_lock);
2143 return (struct rx_call *)0;
2146 } while (!(call = sq->newcall)
2147 && !(socketp && *socketp != OSI_NULLSOCKET));
2148 MUTEX_EXIT(&rx_serverPool_lock);
2150 MUTEX_ENTER(&call->lock);
2156 MUTEX_ENTER(&freeSQEList_lock);
2157 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2158 rx_FreeSQEList = sq;
2159 MUTEX_EXIT(&freeSQEList_lock);
2162 clock_GetTime(&call->startTime);
2163 call->state = RX_STATE_ACTIVE;
2164 call->app.mode = RX_MODE_RECEIVING;
2165 #ifdef RX_KERNEL_TRACE
2166 if (ICL_SETACTIVE(afs_iclSetp)) {
2167 int glockOwner = ISAFS_GLOCK();
2170 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2171 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2178 rxi_calltrace(RX_CALL_START, call);
2179 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2180 call->conn->service->servicePort, call->conn->service->serviceId,
2183 MUTEX_EXIT(&call->lock);
2184 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2186 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2191 #else /* RX_ENABLE_LOCKS */
2193 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2195 struct rx_serverQueueEntry *sq;
2196 struct rx_call *call = (struct rx_call *)0, *choice2;
2197 struct rx_service *service = NULL;
2201 MUTEX_ENTER(&freeSQEList_lock);
2203 if ((sq = rx_FreeSQEList)) {
2204 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2205 MUTEX_EXIT(&freeSQEList_lock);
2206 } else { /* otherwise allocate a new one and return that */
2207 MUTEX_EXIT(&freeSQEList_lock);
2208 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2209 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2210 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2212 MUTEX_ENTER(&sq->lock);
2214 if (cur_service != NULL) {
2215 cur_service->nRequestsRunning--;
2216 MUTEX_ENTER(&rx_quota_mutex);
2217 if (cur_service->nRequestsRunning < cur_service->minProcs)
2220 MUTEX_EXIT(&rx_quota_mutex);
2222 if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
2223 struct rx_call *tcall;
2224 struct opr_queue *cursor;
2225 /* Scan for eligible incoming calls. A call is not eligible
2226 * if the maximum number of calls for its service type are
2227 * already executing */
2228 /* One thread will process calls FCFS (to prevent starvation),
2229 * while the other threads may run ahead looking for calls which
2230 * have all their input data available immediately. This helps
2231 * keep threads from blocking, waiting for data from the client. */
2232 choice2 = (struct rx_call *)0;
2233 for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
2234 tcall = opr_queue_Entry(cursor, struct rx_call, entry);
2235 service = tcall->conn->service;
2236 if (QuotaOK(service)) {
2237 MUTEX_ENTER(&rx_pthread_mutex);
2238 /* XXX - If tcall->entry.next is NULL, then we're no longer
2239 * on a queue at all. This shouldn't happen. */
2240 if (tno == rxi_fcfs_thread_num || !tcall->entry.next) {
2241 MUTEX_EXIT(&rx_pthread_mutex);
2242 /* If we're the fcfs thread, then we'll just use
2243 * this call. If we haven't been able to find an optimal
2244 * choice, and we're at the end of the list, then use a
2245 * 2d choice if one has been identified. Otherwise... */
2246 call = (choice2 ? choice2 : tcall);
2247 service = call->conn->service;
2249 MUTEX_EXIT(&rx_pthread_mutex);
2250 if (!opr_queue_IsEmpty(&tcall->rq)) {
2251 struct rx_packet *rp;
2252 rp = opr_queue_First(&tcall->rq, struct rx_packet,
2254 if (rp->header.seq == 1
2256 || (rp->header.flags & RX_LAST_PACKET))) {
2258 } else if (rxi_2dchoice && !choice2
2259 && !(tcall->flags & RX_CALL_CLEARED)
2260 && (tcall->rprev > rxi_HardAckRate)) {
2273 opr_queue_Remove(&call->entry);
2274 /* we can't schedule a call if there's no data!!! */
2275 /* send an ack if there's no data, if we're missing the
2276 * first packet, or we're missing something between first
2277 * and last -- there's a "hole" in the incoming data. */
2278 if (opr_queue_IsEmpty(&call->rq)
2279 || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
2280 || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
2281 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2283 call->flags &= (~RX_CALL_WAIT_PROC);
2284 service->nRequestsRunning++;
2285 /* just started call in minProcs pool, need fewer to maintain
2287 MUTEX_ENTER(&rx_quota_mutex);
2288 if (service->nRequestsRunning <= service->minProcs)
2291 MUTEX_EXIT(&rx_quota_mutex);
2292 rx_atomic_dec(&rx_nWaiting);
2293 /* MUTEX_EXIT(&call->lock); */
2295 /* If there are no eligible incoming calls, add this process
2296 * to the idle server queue, to wait for one */
2299 *socketp = OSI_NULLSOCKET;
2301 sq->socketp = socketp;
2302 opr_queue_Append(&rx_idleServerQueue, &sq->entry);
2306 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2308 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2309 return (struct rx_call *)0;
2312 } while (!(call = sq->newcall)
2313 && !(socketp && *socketp != OSI_NULLSOCKET));
2315 MUTEX_EXIT(&sq->lock);
2317 MUTEX_ENTER(&freeSQEList_lock);
2318 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2319 rx_FreeSQEList = sq;
2320 MUTEX_EXIT(&freeSQEList_lock);
2323 clock_GetTime(&call->startTime);
2324 call->state = RX_STATE_ACTIVE;
2325 call->app.mode = RX_MODE_RECEIVING;
2326 #ifdef RX_KERNEL_TRACE
2327 if (ICL_SETACTIVE(afs_iclSetp)) {
2328 int glockOwner = ISAFS_GLOCK();
2331 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2332 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2339 rxi_calltrace(RX_CALL_START, call);
2340 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2341 call->conn->service->servicePort, call->conn->service->serviceId,
2344 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2351 #endif /* RX_ENABLE_LOCKS */
2355 /* Establish a procedure to be called when a packet arrives for a
2356 * call. This routine will be called at most once after each call,
2357 * and will also be called if there is an error condition on the or
2358 * the call is complete. Used by multi rx to build a selection
2359 * function which determines which of several calls is likely to be a
2360 * good one to read from.
2361 * NOTE: the way this is currently implemented it is probably only a
2362 * good idea to (1) use it immediately after a newcall (clients only)
2363 * and (2) only use it once. Other uses currently void your warranty
2366 rx_SetArrivalProc(struct rx_call *call,
2367 void (*proc) (struct rx_call * call,
2370 void * handle, int arg)
2372 call->arrivalProc = proc;
2373 call->arrivalProcHandle = handle;
2374 call->arrivalProcArg = arg;
2377 /* Call is finished (possibly prematurely). Return rc to the peer, if
2378 * appropriate, and return the final error code from the conversation
2382 rx_EndCall(struct rx_call *call, afs_int32 rc)
2384 struct rx_connection *conn = call->conn;
2388 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2389 call, rc, call->error, call->abortCode));
2392 MUTEX_ENTER(&call->lock);
2394 if (rc == 0 && call->error == 0) {
2395 call->abortCode = 0;
2396 call->abortCount = 0;
2399 call->arrivalProc = (void (*)())0;
2400 if (rc && call->error == 0) {
2401 rxi_CallError(call, rc);
2402 call->app.mode = RX_MODE_ERROR;
2403 /* Send an abort message to the peer if this error code has
2404 * only just been set. If it was set previously, assume the
2405 * peer has already been sent the error code or will request it
2407 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2409 if (conn->type == RX_SERVER_CONNECTION) {
2410 /* Make sure reply or at least dummy reply is sent */
2411 if (call->app.mode == RX_MODE_RECEIVING) {
2412 MUTEX_EXIT(&call->lock);
2413 rxi_WriteProc(call, 0, 0);
2414 MUTEX_ENTER(&call->lock);
2416 if (call->app.mode == RX_MODE_SENDING) {
2417 MUTEX_EXIT(&call->lock);
2418 rxi_FlushWrite(call);
2419 MUTEX_ENTER(&call->lock);
2421 rxi_calltrace(RX_CALL_END, call);
2422 /* Call goes to hold state until reply packets are acknowledged */
2423 if (call->tfirst + call->nSoftAcked < call->tnext) {
2424 call->state = RX_STATE_HOLD;
2426 call->state = RX_STATE_DALLY;
2427 rxi_ClearTransmitQueue(call, 0);
2428 rxi_rto_cancel(call);
2429 rxevent_Cancel(&call->keepAliveEvent, call,
2430 RX_CALL_REFCOUNT_ALIVE);
2432 } else { /* Client connection */
2434 /* Make sure server receives input packets, in the case where
2435 * no reply arguments are expected */
2437 if ((call->app.mode == RX_MODE_SENDING)
2438 || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2439 MUTEX_EXIT(&call->lock);
2440 (void)rxi_ReadProc(call, &dummy, 1);
2441 MUTEX_ENTER(&call->lock);
2444 /* If we had an outstanding delayed ack, be nice to the server
2445 * and force-send it now.
2447 if (call->delayedAckEvent) {
2448 rxevent_Cancel(&call->delayedAckEvent, call,
2449 RX_CALL_REFCOUNT_DELAY);
2450 rxi_SendDelayedAck(NULL, call, NULL, 0);
2453 /* We need to release the call lock since it's lower than the
2454 * conn_call_lock and we don't want to hold the conn_call_lock
2455 * over the rx_ReadProc call. The conn_call_lock needs to be held
2456 * here for the case where rx_NewCall is perusing the calls on
2457 * the connection structure. We don't want to signal until
2458 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2459 * have checked this call, found it active and by the time it
2460 * goes to sleep, will have missed the signal.
2462 MUTEX_EXIT(&call->lock);
2463 MUTEX_ENTER(&conn->conn_call_lock);
2464 MUTEX_ENTER(&call->lock);
2466 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2467 conn->lastBusy[call->channel] = 0;
2470 MUTEX_ENTER(&conn->conn_data_lock);
2471 conn->flags |= RX_CONN_BUSY;
2472 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2473 MUTEX_EXIT(&conn->conn_data_lock);
2474 #ifdef RX_ENABLE_LOCKS
2475 CV_BROADCAST(&conn->conn_call_cv);
2480 #ifdef RX_ENABLE_LOCKS
2482 MUTEX_EXIT(&conn->conn_data_lock);
2484 #endif /* RX_ENABLE_LOCKS */
2485 call->state = RX_STATE_DALLY;
2487 error = call->error;
2489 /* currentPacket, nLeft, and NFree must be zeroed here, because
2490 * ResetCall cannot: ResetCall may be called at splnet(), in the
2491 * kernel version, and may interrupt the macros rx_Read or
2492 * rx_Write, which run at normal priority for efficiency. */
2493 if (call->app.currentPacket) {
2494 #ifdef RX_TRACK_PACKETS
2495 call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
2497 rxi_FreePacket(call->app.currentPacket);
2498 call->app.currentPacket = (struct rx_packet *)0;
2501 call->app.nLeft = call->app.nFree = call->app.curlen = 0;
2503 /* Free any packets from the last call to ReadvProc/WritevProc */
2504 #ifdef RXDEBUG_PACKET
2506 #endif /* RXDEBUG_PACKET */
2507 rxi_FreePackets(0, &call->app.iovq);
2508 MUTEX_EXIT(&call->lock);
2510 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2511 if (conn->type == RX_CLIENT_CONNECTION) {
2512 MUTEX_ENTER(&conn->conn_data_lock);
2513 conn->flags &= ~RX_CONN_BUSY;
2514 MUTEX_EXIT(&conn->conn_data_lock);
2515 MUTEX_EXIT(&conn->conn_call_lock);
2519 * Map errors to the local host's errno.h format.
2521 error = ntoh_syserr_conv(error);
2525 #if !defined(KERNEL)
2527 /* Call this routine when shutting down a server or client (especially
2528 * clients). This will allow Rx to gracefully garbage collect server
2529 * connections, and reduce the number of retries that a server might
2530 * make to a dead client.
2531 * This is not quite right, since some calls may still be ongoing and
2532 * we can't lock them to destroy them. */
2536 struct rx_connection **conn_ptr, **conn_end;
2540 if (rxinit_status == 1) {
2542 return; /* Already shutdown. */
2544 rxi_DeleteCachedConnections();
2545 if (rx_connHashTable) {
2546 MUTEX_ENTER(&rx_connHashTable_lock);
2547 for (conn_ptr = &rx_connHashTable[0], conn_end =
2548 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2550 struct rx_connection *conn, *next;
2551 for (conn = *conn_ptr; conn; conn = next) {
2553 if (conn->type == RX_CLIENT_CONNECTION) {
2554 MUTEX_ENTER(&rx_refcnt_mutex);
2556 MUTEX_EXIT(&rx_refcnt_mutex);
2557 #ifdef RX_ENABLE_LOCKS
2558 rxi_DestroyConnectionNoLock(conn);
2559 #else /* RX_ENABLE_LOCKS */
2560 rxi_DestroyConnection(conn);
2561 #endif /* RX_ENABLE_LOCKS */
2565 #ifdef RX_ENABLE_LOCKS
2566 while (rx_connCleanup_list) {
2567 struct rx_connection *conn;
2568 conn = rx_connCleanup_list;
2569 rx_connCleanup_list = rx_connCleanup_list->next;
2570 MUTEX_EXIT(&rx_connHashTable_lock);
2571 rxi_CleanupConnection(conn);
2572 MUTEX_ENTER(&rx_connHashTable_lock);
2574 MUTEX_EXIT(&rx_connHashTable_lock);
2575 #endif /* RX_ENABLE_LOCKS */
2580 afs_winsockCleanup();
2588 /* if we wakeup packet waiter too often, can get in loop with two
2589 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2591 rxi_PacketsUnWait(void)
2593 if (!rx_waitingForPackets) {
2597 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2598 return; /* still over quota */
2601 rx_waitingForPackets = 0;
2602 #ifdef RX_ENABLE_LOCKS
2603 CV_BROADCAST(&rx_waitingForPackets_cv);
2605 osi_rxWakeup(&rx_waitingForPackets);
2611 /* ------------------Internal interfaces------------------------- */
2613 /* Return this process's service structure for the
2614 * specified socket and service */
2615 static struct rx_service *
2616 rxi_FindService(osi_socket socket, u_short serviceId)
2618 struct rx_service **sp;
2619 for (sp = &rx_services[0]; *sp; sp++) {
2620 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2626 #ifdef RXDEBUG_PACKET
2627 #ifdef KDUMP_RX_LOCK
2628 static struct rx_call_rx_lock *rx_allCallsp = 0;
2630 static struct rx_call *rx_allCallsp = 0;
2632 #endif /* RXDEBUG_PACKET */
2634 /* Allocate a call structure, for the indicated channel of the
2635 * supplied connection. The mode and state of the call must be set by
2636 * the caller. Returns the call with mutex locked. */
2637 static struct rx_call *
2638 rxi_NewCall(struct rx_connection *conn, int channel)
2640 struct rx_call *call;
2641 #ifdef RX_ENABLE_LOCKS
2642 struct rx_call *cp; /* Call pointer temp */
2643 struct opr_queue *cursor;
2646 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2648 /* Grab an existing call structure, or allocate a new one.
2649 * Existing call structures are assumed to have been left reset by
2651 MUTEX_ENTER(&rx_freeCallQueue_lock);
2653 #ifdef RX_ENABLE_LOCKS
2655 * EXCEPT that the TQ might not yet be cleared out.
2656 * Skip over those with in-use TQs.
2659 for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
2660 cp = opr_queue_Entry(cursor, struct rx_call, entry);
2661 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2667 #else /* RX_ENABLE_LOCKS */
2668 if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
2669 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
2670 #endif /* RX_ENABLE_LOCKS */
2671 opr_queue_Remove(&call->entry);
2672 if (rx_stats_active)
2673 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2674 MUTEX_EXIT(&rx_freeCallQueue_lock);
2675 MUTEX_ENTER(&call->lock);
2676 CLEAR_CALL_QUEUE_LOCK(call);
2677 #ifdef RX_ENABLE_LOCKS
2678 /* Now, if TQ wasn't cleared earlier, do it now. */
2679 rxi_WaitforTQBusy(call);
2680 if (call->flags & RX_CALL_TQ_CLEARME) {
2681 rxi_ClearTransmitQueue(call, 1);
2682 /*queue_Init(&call->tq);*/
2684 #endif /* RX_ENABLE_LOCKS */
2685 /* Bind the call to its connection structure */
2687 rxi_ResetCall(call, 1);
2690 call = rxi_Alloc(sizeof(struct rx_call));
2691 #ifdef RXDEBUG_PACKET
2692 call->allNextp = rx_allCallsp;
2693 rx_allCallsp = call;
2695 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2696 #else /* RXDEBUG_PACKET */
2697 rx_atomic_inc(&rx_stats.nCallStructs);
2698 #endif /* RXDEBUG_PACKET */
2700 MUTEX_EXIT(&rx_freeCallQueue_lock);
2701 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2702 MUTEX_ENTER(&call->lock);
2703 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2704 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2705 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2707 /* Initialize once-only items */
2708 opr_queue_Init(&call->tq);
2709 opr_queue_Init(&call->rq);
2710 opr_queue_Init(&call->app.iovq);
2711 #ifdef RXDEBUG_PACKET
2712 call->rqc = call->tqc = call->iovqc = 0;
2713 #endif /* RXDEBUG_PACKET */
2714 /* Bind the call to its connection structure (prereq for reset) */
2716 rxi_ResetCall(call, 1);
2718 call->channel = channel;
2719 call->callNumber = &conn->callNumber[channel];
2720 call->rwind = conn->rwind[channel];
2721 call->twind = conn->twind[channel];
2722 /* Note that the next expected call number is retained (in
2723 * conn->callNumber[i]), even if we reallocate the call structure
2725 conn->call[channel] = call;
2726 /* if the channel's never been used (== 0), we should start at 1, otherwise
2727 * the call number is valid from the last time this channel was used */
2728 if (*call->callNumber == 0)
2729 *call->callNumber = 1;
2734 /* A call has been inactive long enough that so we can throw away
2735 * state, including the call structure, which is placed on the call
2738 * call->lock amd rx_refcnt_mutex are held upon entry.
2739 * haveCTLock is set when called from rxi_ReapConnections.
2741 * return 1 if the call is freed, 0 if not.
2744 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2746 int channel = call->channel;
2747 struct rx_connection *conn = call->conn;
2748 u_char state = call->state;
2751 * We are setting the state to RX_STATE_RESET to
2752 * ensure that no one else will attempt to use this
2753 * call once we drop the refcnt lock. We must drop
2754 * the refcnt lock before calling rxi_ResetCall
2755 * because it cannot be held across acquiring the
2756 * freepktQ lock. NewCall does the same.
2758 call->state = RX_STATE_RESET;
2759 MUTEX_EXIT(&rx_refcnt_mutex);
2760 rxi_ResetCall(call, 0);
2762 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2764 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2765 (*call->callNumber)++;
2767 if (call->conn->call[channel] == call)
2768 call->conn->call[channel] = 0;
2769 MUTEX_EXIT(&conn->conn_call_lock);
2772 * We couldn't obtain the conn_call_lock so we can't
2773 * disconnect the call from the connection. Set the
2774 * call state to dally so that the call can be reused.
2776 MUTEX_ENTER(&rx_refcnt_mutex);
2777 call->state = RX_STATE_DALLY;
2781 MUTEX_ENTER(&rx_freeCallQueue_lock);
2782 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2783 #ifdef RX_ENABLE_LOCKS
2784 /* A call may be free even though its transmit queue is still in use.
2785 * Since we search the call list from head to tail, put busy calls at
2786 * the head of the list, and idle calls at the tail.
2788 if (call->flags & RX_CALL_TQ_BUSY)
2789 opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
2791 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2792 #else /* RX_ENABLE_LOCKS */
2793 opr_queue_Append(&rx_freeCallQueue, &call->entry);
2794 #endif /* RX_ENABLE_LOCKS */
2795 if (rx_stats_active)
2796 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2797 MUTEX_EXIT(&rx_freeCallQueue_lock);
2799 /* Destroy the connection if it was previously slated for
2800 * destruction, i.e. the Rx client code previously called
2801 * rx_DestroyConnection (client connections), or
2802 * rxi_ReapConnections called the same routine (server
2803 * connections). Only do this, however, if there are no
2804 * outstanding calls. Note that for fine grain locking, there appears
2805 * to be a deadlock in that rxi_FreeCall has a call locked and
2806 * DestroyConnectionNoLock locks each call in the conn. But note a
2807 * few lines up where we have removed this call from the conn.
2808 * If someone else destroys a connection, they either have no
2809 * call lock held or are going through this section of code.
2811 MUTEX_ENTER(&conn->conn_data_lock);
2812 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2813 MUTEX_ENTER(&rx_refcnt_mutex);
2815 MUTEX_EXIT(&rx_refcnt_mutex);
2816 MUTEX_EXIT(&conn->conn_data_lock);
2817 #ifdef RX_ENABLE_LOCKS
2819 rxi_DestroyConnectionNoLock(conn);
2821 rxi_DestroyConnection(conn);
2822 #else /* RX_ENABLE_LOCKS */
2823 rxi_DestroyConnection(conn);
2824 #endif /* RX_ENABLE_LOCKS */
2826 MUTEX_EXIT(&conn->conn_data_lock);
2828 MUTEX_ENTER(&rx_refcnt_mutex);
2832 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2833 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2836 rxi_Alloc(size_t size)
2840 if (rx_stats_active) {
2841 rx_atomic_add(&rxi_Allocsize, (int) size);
2842 rx_atomic_inc(&rxi_Alloccnt);
2846 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2847 afs_osi_Alloc_NoSleep(size);
2852 osi_Panic("rxi_Alloc error");
2858 rxi_Free(void *addr, size_t size)
2860 if (rx_stats_active) {
2861 rx_atomic_sub(&rxi_Allocsize, (int) size);
2862 rx_atomic_dec(&rxi_Alloccnt);
2864 osi_Free(addr, size);
2868 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2870 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2871 struct rx_peer *next = NULL;
2875 MUTEX_ENTER(&rx_peerHashTable_lock);
2877 peer_ptr = &rx_peerHashTable[0];
2878 peer_end = &rx_peerHashTable[rx_hashTableSize];
2881 for ( ; peer_ptr < peer_end; peer_ptr++) {
2884 for ( ; peer; peer = next) {
2886 if (host == peer->host)
2891 hashIndex = PEER_HASH(host, port);
2892 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2893 if ((peer->host == host) && (peer->port == port))
2898 MUTEX_ENTER(&rx_peerHashTable_lock);
2903 MUTEX_EXIT(&rx_peerHashTable_lock);
2905 MUTEX_ENTER(&peer->peer_lock);
2906 /* We don't handle dropping below min, so don't */
2907 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2908 peer->ifMTU=MIN(mtu, peer->ifMTU);
2909 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2910 /* if we tweaked this down, need to tune our peer MTU too */
2911 peer->MTU = MIN(peer->MTU, peer->natMTU);
2912 /* if we discovered a sub-1500 mtu, degrade */
2913 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2914 peer->maxDgramPackets = 1;
2915 /* We no longer have valid peer packet information */
2916 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2917 peer->maxPacketSize = 0;
2918 MUTEX_EXIT(&peer->peer_lock);
2920 MUTEX_ENTER(&rx_peerHashTable_lock);
2922 if (host && !port) {
2924 /* pick up where we left off */
2928 MUTEX_EXIT(&rx_peerHashTable_lock);
2931 #ifdef AFS_RXERRQ_ENV
2933 rxi_SetPeerDead(afs_uint32 host, afs_uint16 port)
2935 int hashIndex = PEER_HASH(host, port);
2936 struct rx_peer *peer;
2938 MUTEX_ENTER(&rx_peerHashTable_lock);
2940 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2941 if (peer->host == host && peer->port == port) {
2947 rx_atomic_inc(&peer->neterrs);
2950 MUTEX_EXIT(&rx_peerHashTable_lock);
2954 rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
2956 # ifdef AFS_ADAPT_PMTU
2957 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
2958 rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
2962 if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
2963 switch (err->ee_code) {
2964 case ICMP_NET_UNREACH:
2965 case ICMP_HOST_UNREACH:
2966 case ICMP_PORT_UNREACH:
2969 rxi_SetPeerDead(addr, port);
2974 #endif /* AFS_RXERRQ_ENV */
2976 /* Find the peer process represented by the supplied (host,port)
2977 * combination. If there is no appropriate active peer structure, a
2978 * new one will be allocated and initialized
2979 * The origPeer, if set, is a pointer to a peer structure on which the
2980 * refcount will be be decremented. This is used to replace the peer
2981 * structure hanging off a connection structure */
2983 rxi_FindPeer(afs_uint32 host, u_short port,
2984 struct rx_peer *origPeer, int create)
2988 hashIndex = PEER_HASH(host, port);
2989 MUTEX_ENTER(&rx_peerHashTable_lock);
2990 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2991 if ((pp->host == host) && (pp->port == port))
2996 pp = rxi_AllocPeer(); /* This bzero's *pp */
2997 pp->host = host; /* set here or in InitPeerParams is zero */
2999 #ifdef AFS_RXERRQ_ENV
3000 rx_atomic_set(&pp->neterrs, 0);
3002 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
3003 opr_queue_Init(&pp->rpcStats);
3004 pp->next = rx_peerHashTable[hashIndex];
3005 rx_peerHashTable[hashIndex] = pp;
3006 rxi_InitPeerParams(pp);
3007 if (rx_stats_active)
3008 rx_atomic_inc(&rx_stats.nPeerStructs);
3015 origPeer->refCount--;
3016 MUTEX_EXIT(&rx_peerHashTable_lock);
3021 /* Find the connection at (host, port) started at epoch, and with the
3022 * given connection id. Creates the server connection if necessary.
3023 * The type specifies whether a client connection or a server
3024 * connection is desired. In both cases, (host, port) specify the
3025 * peer's (host, pair) pair. Client connections are not made
3026 * automatically by this routine. The parameter socket gives the
3027 * socket descriptor on which the packet was received. This is used,
3028 * in the case of server connections, to check that *new* connections
3029 * come via a valid (port, serviceId). Finally, the securityIndex
3030 * parameter must match the existing index for the connection. If a
3031 * server connection is created, it will be created using the supplied
3032 * index, if the index is valid for this service */
3033 static struct rx_connection *
3034 rxi_FindConnection(osi_socket socket, afs_uint32 host,
3035 u_short port, u_short serviceId, afs_uint32 cid,
3036 afs_uint32 epoch, int type, u_int securityIndex)
3038 int hashindex, flag, i;
3039 struct rx_connection *conn;
3040 hashindex = CONN_HASH(host, port, cid, epoch, type);
3041 MUTEX_ENTER(&rx_connHashTable_lock);
3042 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
3043 rx_connHashTable[hashindex],
3046 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
3047 && (epoch == conn->epoch)) {
3048 struct rx_peer *pp = conn->peer;
3049 if (securityIndex != conn->securityIndex) {
3050 /* this isn't supposed to happen, but someone could forge a packet
3051 * like this, and there seems to be some CM bug that makes this
3052 * happen from time to time -- in which case, the fileserver
3054 MUTEX_EXIT(&rx_connHashTable_lock);
3055 return (struct rx_connection *)0;
3057 if (pp->host == host && pp->port == port)
3059 if (type == RX_CLIENT_CONNECTION && pp->port == port)
3061 /* So what happens when it's a callback connection? */
3062 if ( /*type == RX_CLIENT_CONNECTION && */
3063 (conn->epoch & 0x80000000))
3067 /* the connection rxLastConn that was used the last time is not the
3068 ** one we are looking for now. Hence, start searching in the hash */
3070 conn = rx_connHashTable[hashindex];
3075 struct rx_service *service;
3076 if (type == RX_CLIENT_CONNECTION) {
3077 MUTEX_EXIT(&rx_connHashTable_lock);
3078 return (struct rx_connection *)0;
3080 service = rxi_FindService(socket, serviceId);
3081 if (!service || (securityIndex >= service->nSecurityObjects)
3082 || (service->securityObjects[securityIndex] == 0)) {
3083 MUTEX_EXIT(&rx_connHashTable_lock);
3084 return (struct rx_connection *)0;
3086 conn = rxi_AllocConnection(); /* This bzero's the connection */
3087 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3088 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3089 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3090 conn->next = rx_connHashTable[hashindex];
3091 rx_connHashTable[hashindex] = conn;
3092 conn->peer = rxi_FindPeer(host, port, 0, 1);
3093 conn->type = RX_SERVER_CONNECTION;
3094 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3095 conn->epoch = epoch;
3096 conn->cid = cid & RX_CIDMASK;
3097 conn->ackRate = RX_FAST_ACK_RATE;
3098 conn->service = service;
3099 conn->serviceId = serviceId;
3100 conn->securityIndex = securityIndex;
3101 conn->securityObject = service->securityObjects[securityIndex];
3102 conn->nSpecific = 0;
3103 conn->specific = NULL;
3104 rx_SetConnDeadTime(conn, service->connDeadTime);
3105 conn->idleDeadTime = service->idleDeadTime;
3106 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3107 for (i = 0; i < RX_MAXCALLS; i++) {
3108 conn->twind[i] = rx_initSendWindow;
3109 conn->rwind[i] = rx_initReceiveWindow;
3111 /* Notify security object of the new connection */
3112 RXS_NewConnection(conn->securityObject, conn);
3113 /* XXXX Connection timeout? */
3114 if (service->newConnProc)
3115 (*service->newConnProc) (conn);
3116 if (rx_stats_active)
3117 rx_atomic_inc(&rx_stats.nServerConns);
3120 MUTEX_ENTER(&rx_refcnt_mutex);
3122 MUTEX_EXIT(&rx_refcnt_mutex);
3124 rxLastConn = conn; /* store this connection as the last conn used */
3125 MUTEX_EXIT(&rx_connHashTable_lock);
3130 * Timeout a call on a busy call channel if appropriate.
3132 * @param[in] call The busy call.
3134 * @pre 'call' is marked as busy (namely,
3135 * call->conn->lastBusy[call->channel] != 0)
3137 * @pre call->lock is held
3138 * @pre rxi_busyChannelError is nonzero
3140 * @note call->lock is dropped and reacquired
3143 rxi_CheckBusy(struct rx_call *call)
3145 struct rx_connection *conn = call->conn;
3146 int channel = call->channel;
3147 int freechannel = 0;
3149 afs_uint32 callNumber;
3151 MUTEX_EXIT(&call->lock);
3153 MUTEX_ENTER(&conn->conn_call_lock);
3154 callNumber = *call->callNumber;
3156 /* Are there any other call slots on this conn that we should try? Look for
3157 * slots that are empty and are either non-busy, or were marked as busy
3158 * longer than conn->secondsUntilDead seconds before this call started. */
3160 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3162 /* only look at channels that aren't us */
3166 if (conn->lastBusy[i]) {
3167 /* if this channel looked busy too recently, don't look at it */
3168 if (conn->lastBusy[i] >= call->startTime.sec) {
3171 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3176 if (conn->call[i]) {
3177 struct rx_call *tcall = conn->call[i];
3178 MUTEX_ENTER(&tcall->lock);
3179 if (tcall->state == RX_STATE_DALLY) {
3182 MUTEX_EXIT(&tcall->lock);
3188 MUTEX_ENTER(&call->lock);
3190 /* Since the call->lock and conn->conn_call_lock have been released it is
3191 * possible that (1) the call may no longer be busy and/or (2) the call may
3192 * have been reused by another waiting thread. Therefore, we must confirm
3193 * that the call state has not changed when deciding whether or not to
3194 * force this application thread to retry by forcing a Timeout error. */
3196 if (freechannel && *call->callNumber == callNumber &&
3197 (call->flags & RX_CALL_PEER_BUSY)) {
3198 /* Since 'freechannel' is set, there exists another channel in this
3199 * rx_conn that the application thread might be able to use. We know
3200 * that we have the correct call since callNumber is unchanged, and we
3201 * know that the call is still busy. So, set the call error state to
3202 * rxi_busyChannelError so the application can retry the request,
3203 * presumably on a less-busy call channel. */
3205 rxi_CallError(call, RX_CALL_BUSY);
3207 MUTEX_EXIT(&conn->conn_call_lock);
3211 * Abort the call if the server is over the busy threshold. This
3212 * can be used without requiring a call structure be initialised,
3213 * or connected to a particular channel
3216 rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
3217 struct rx_packet *np)
3219 if ((rx_BusyThreshold > 0) &&
3220 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3221 rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
3222 rx_BusyError, np, 0);
3223 if (rx_stats_active)
3224 rx_atomic_inc(&rx_stats.nBusies);
3231 static_inline struct rx_call *
3232 rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
3235 struct rx_call *call;
3237 channel = np->header.cid & RX_CHANNELMASK;
3238 MUTEX_ENTER(&conn->conn_call_lock);
3239 call = conn->call[channel];
3240 if (!call || conn->callNumber[channel] != np->header.callNumber) {
3241 MUTEX_EXIT(&conn->conn_call_lock);
3242 if (rx_stats_active)
3243 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3247 MUTEX_ENTER(&call->lock);
3248 MUTEX_EXIT(&conn->conn_call_lock);
3250 if ((call->state == RX_STATE_DALLY)
3251 && np->header.type == RX_PACKET_TYPE_ACK) {
3252 if (rx_stats_active)
3253 rx_atomic_inc(&rx_stats.ignorePacketDally);
3254 MUTEX_EXIT(&call->lock);
3261 static_inline struct rx_call *
3262 rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
3263 struct rx_connection *conn)
3266 struct rx_call *call;
3268 channel = np->header.cid & RX_CHANNELMASK;
3269 MUTEX_ENTER(&conn->conn_call_lock);
3270 call = conn->call[channel];
3273 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3274 MUTEX_EXIT(&conn->conn_call_lock);
3278 call = rxi_NewCall(conn, channel); /* returns locked call */
3279 *call->callNumber = np->header.callNumber;
3280 MUTEX_EXIT(&conn->conn_call_lock);
3282 call->state = RX_STATE_PRECALL;
3283 clock_GetTime(&call->queueTime);
3284 call->app.bytesSent = 0;
3285 call->app.bytesRcvd = 0;
3286 rxi_KeepAliveOn(call);
3291 if (np->header.callNumber == conn->callNumber[channel]) {
3292 MUTEX_ENTER(&call->lock);
3293 MUTEX_EXIT(&conn->conn_call_lock);
3297 if (np->header.callNumber < conn->callNumber[channel]) {
3298 MUTEX_EXIT(&conn->conn_call_lock);
3299 if (rx_stats_active)
3300 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3304 MUTEX_ENTER(&call->lock);
3305 MUTEX_EXIT(&conn->conn_call_lock);
3307 /* Wait until the transmit queue is idle before deciding
3308 * whether to reset the current call. Chances are that the
3309 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3312 #ifdef RX_ENABLE_LOCKS
3313 if (call->state == RX_STATE_ACTIVE) {
3314 rxi_WaitforTQBusy(call);
3315 /* If we entered error state while waiting,
3316 * must call rxi_CallError to permit rxi_ResetCall
3317 * to processed when the tqWaiter count hits zero.
3320 rxi_CallError(call, call->error);
3321 MUTEX_EXIT(&call->lock);
3325 #endif /* RX_ENABLE_LOCKS */
3326 /* If the new call cannot be taken right now send a busy and set
3327 * the error condition in this call, so that it terminates as
3328 * quickly as possible */
3329 if (call->state == RX_STATE_ACTIVE) {
3330 rxi_CallError(call, RX_CALL_DEAD);
3331 rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
3333 MUTEX_EXIT(&call->lock);
3337 if (rxi_AbortIfServerBusy(socket, conn, np)) {
3338 MUTEX_EXIT(&call->lock);
3342 rxi_ResetCall(call, 0);
3343 /* The conn_call_lock is not held but no one else should be
3344 * using this call channel while we are processing this incoming
3345 * packet. This assignment should be safe.
3347 *call->callNumber = np->header.callNumber;
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);
3358 /* There are two packet tracing routines available for testing and monitoring
3359 * Rx. One is called just after every packet is received and the other is
3360 * called just before every packet is sent. Received packets, have had their
3361 * headers decoded, and packets to be sent have not yet had their headers
3362 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3363 * containing the network address. Both can be modified. The return value, if
3364 * non-zero, indicates that the packet should be dropped. */
3366 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3367 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3369 /* A packet has been received off the interface. Np is the packet, socket is
3370 * the socket number it was received from (useful in determining which service
3371 * this packet corresponds to), and (host, port) reflect the host,port of the
3372 * sender. This call returns the packet to the caller if it is finished with
3373 * it, rather than de-allocating it, just as a small performance hack */
3376 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3377 afs_uint32 host, u_short port, int *tnop,
3378 struct rx_call **newcallp)
3380 struct rx_call *call;
3381 struct rx_connection *conn;
3386 struct rx_packet *tnp;
3389 /* We don't print out the packet until now because (1) the time may not be
3390 * accurate enough until now in the lwp implementation (rx_Listener only gets
3391 * the time after the packet is read) and (2) from a protocol point of view,
3392 * this is the first time the packet has been seen */
3393 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3394 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3395 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3396 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3397 np->header.epoch, np->header.cid, np->header.callNumber,
3398 np->header.seq, np->header.flags, np));
3401 /* Account for connectionless packets */
3402 if (rx_stats_active &&
3403 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3404 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3405 struct rx_peer *peer;
3407 /* Try to look up the peer structure, but don't create one */
3408 peer = rxi_FindPeer(host, port, 0, 0);
3410 /* Since this may not be associated with a connection, it may have
3411 * no refCount, meaning we could race with ReapConnections
3414 if (peer && (peer->refCount > 0)) {
3415 #ifdef AFS_RXERRQ_ENV
3416 if (rx_atomic_read(&peer->neterrs)) {
3417 rx_atomic_set(&peer->neterrs, 0);
3420 MUTEX_ENTER(&peer->peer_lock);
3421 peer->bytesReceived += np->length;
3422 MUTEX_EXIT(&peer->peer_lock);
3426 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3427 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3430 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3431 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3434 /* If an input tracer function is defined, call it with the packet and
3435 * network address. Note this function may modify its arguments. */
3436 if (rx_justReceived) {
3437 struct sockaddr_in addr;
3439 addr.sin_family = AF_INET;
3440 addr.sin_port = port;
3441 addr.sin_addr.s_addr = host;
3442 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3443 addr.sin_len = sizeof(addr);
3444 #endif /* AFS_OSF_ENV */
3445 drop = (*rx_justReceived) (np, &addr);
3446 /* drop packet if return value is non-zero */
3449 port = addr.sin_port; /* in case fcn changed addr */
3450 host = addr.sin_addr.s_addr;
3454 /* If packet was not sent by the client, then *we* must be the client */
3455 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3456 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3458 /* Find the connection (or fabricate one, if we're the server & if
3459 * necessary) associated with this packet */
3461 rxi_FindConnection(socket, host, port, np->header.serviceId,
3462 np->header.cid, np->header.epoch, type,
3463 np->header.securityIndex);
3465 /* To avoid having 2 connections just abort at each other,
3466 don't abort an abort. */
3468 if (np->header.type != RX_PACKET_TYPE_ABORT)
3469 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3474 #ifdef AFS_RXERRQ_ENV
3475 if (rx_atomic_read(&conn->peer->neterrs)) {
3476 rx_atomic_set(&conn->peer->neterrs, 0);
3480 /* If we're doing statistics, then account for the incoming packet */
3481 if (rx_stats_active) {
3482 MUTEX_ENTER(&conn->peer->peer_lock);
3483 conn->peer->bytesReceived += np->length;
3484 MUTEX_EXIT(&conn->peer->peer_lock);
3487 /* If the connection is in an error state, send an abort packet and ignore
3488 * the incoming packet */
3490 /* Don't respond to an abort packet--we don't want loops! */
3491 MUTEX_ENTER(&conn->conn_data_lock);
3492 if (np->header.type != RX_PACKET_TYPE_ABORT)
3493 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3494 putConnection(conn);
3495 MUTEX_EXIT(&conn->conn_data_lock);
3499 /* Check for connection-only requests (i.e. not call specific). */
3500 if (np->header.callNumber == 0) {
3501 switch (np->header.type) {
3502 case RX_PACKET_TYPE_ABORT: {
3503 /* What if the supplied error is zero? */
3504 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3505 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3506 rxi_ConnectionError(conn, errcode);
3507 putConnection(conn);
3510 case RX_PACKET_TYPE_CHALLENGE:
3511 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3512 putConnection(conn);
3514 case RX_PACKET_TYPE_RESPONSE:
3515 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3516 putConnection(conn);
3518 case RX_PACKET_TYPE_PARAMS:
3519 case RX_PACKET_TYPE_PARAMS + 1:
3520 case RX_PACKET_TYPE_PARAMS + 2:
3521 /* ignore these packet types for now */
3522 putConnection(conn);
3526 /* Should not reach here, unless the peer is broken: send an
3528 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3529 MUTEX_ENTER(&conn->conn_data_lock);
3530 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3531 putConnection(conn);
3532 MUTEX_EXIT(&conn->conn_data_lock);
3537 if (type == RX_SERVER_CONNECTION)
3538 call = rxi_ReceiveServerCall(socket, np, conn);
3540 call = rxi_ReceiveClientCall(np, conn);
3543 putConnection(conn);
3547 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3548 /* Set remote user defined status from packet */
3549 call->remoteStatus = np->header.userStatus;
3551 /* Now do packet type-specific processing */
3552 switch (np->header.type) {
3553 case RX_PACKET_TYPE_DATA:
3554 /* If we're a client, and receiving a response, then all the packets
3555 * we transmitted packets are implicitly acknowledged. */
3556 if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
3557 rxi_AckAllInTransmitQueue(call);
3559 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3562 case RX_PACKET_TYPE_ACK:
3563 /* Respond immediately to ack packets requesting acknowledgement
3565 if (np->header.flags & RX_REQUEST_ACK) {
3567 (void)rxi_SendCallAbort(call, 0, 1, 0);
3569 (void)rxi_SendAck(call, 0, np->header.serial,
3570 RX_ACK_PING_RESPONSE, 1);
3572 np = rxi_ReceiveAckPacket(call, np, 1);
3574 case RX_PACKET_TYPE_ABORT: {
3575 /* An abort packet: reset the call, passing the error up to the user. */
3576 /* What if error is zero? */
3577 /* What if the error is -1? the application will treat it as a timeout. */
3578 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3579 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3580 rxi_CallError(call, errdata);
3581 MUTEX_EXIT(&call->lock);
3582 putConnection(conn);
3583 return np; /* xmitting; drop packet */
3585 case RX_PACKET_TYPE_BUSY: {
3586 struct clock busyTime;
3588 clock_GetTime(&busyTime);
3590 MUTEX_EXIT(&call->lock);
3592 MUTEX_ENTER(&conn->conn_call_lock);
3593 MUTEX_ENTER(&call->lock);
3594 conn->lastBusy[call->channel] = busyTime.sec;
3595 call->flags |= RX_CALL_PEER_BUSY;
3596 MUTEX_EXIT(&call->lock);
3597 MUTEX_EXIT(&conn->conn_call_lock);
3599 putConnection(conn);
3603 case RX_PACKET_TYPE_ACKALL:
3604 /* All packets acknowledged, so we can drop all packets previously
3605 * readied for sending */
3606 rxi_AckAllInTransmitQueue(call);
3609 /* Should not reach here, unless the peer is broken: send an abort
3611 rxi_CallError(call, RX_PROTOCOL_ERROR);
3612 np = rxi_SendCallAbort(call, np, 1, 0);
3615 /* Note when this last legitimate packet was received, for keep-alive
3616 * processing. Note, we delay getting the time until now in the hope that
3617 * the packet will be delivered to the user before any get time is required
3618 * (if not, then the time won't actually be re-evaluated here). */
3619 call->lastReceiveTime = clock_Sec();
3620 /* we've received a legit packet, so the channel is not busy */
3621 call->flags &= ~RX_CALL_PEER_BUSY;
3622 MUTEX_EXIT(&call->lock);
3623 putConnection(conn);
3627 /* return true if this is an "interesting" connection from the point of view
3628 of someone trying to debug the system */
3630 rxi_IsConnInteresting(struct rx_connection *aconn)
3633 struct rx_call *tcall;
3635 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3638 for (i = 0; i < RX_MAXCALLS; i++) {
3639 tcall = aconn->call[i];
3641 if ((tcall->state == RX_STATE_PRECALL)
3642 || (tcall->state == RX_STATE_ACTIVE))
3644 if ((tcall->app.mode == RX_MODE_SENDING)
3645 || (tcall->app.mode == RX_MODE_RECEIVING))
3653 /* if this is one of the last few packets AND it wouldn't be used by the
3654 receiving call to immediately satisfy a read request, then drop it on
3655 the floor, since accepting it might prevent a lock-holding thread from
3656 making progress in its reading. If a call has been cleared while in
3657 the precall state then ignore all subsequent packets until the call
3658 is assigned to a thread. */
3661 TooLow(struct rx_packet *ap, struct rx_call *acall)
3665 MUTEX_ENTER(&rx_quota_mutex);
3666 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3667 && (acall->state == RX_STATE_PRECALL))
3668 || ((rx_nFreePackets < rxi_dataQuota + 2)
3669 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3670 && (acall->flags & RX_CALL_READER_WAIT)))) {
3673 MUTEX_EXIT(&rx_quota_mutex);
3679 * Clear the attach wait flag on a connection and proceed.
3681 * Any processing waiting for a connection to be attached should be
3682 * unblocked. We clear the flag and do any other needed tasks.
3685 * the conn to unmark waiting for attach
3687 * @pre conn's conn_data_lock must be locked before calling this function
3691 rxi_ConnClearAttachWait(struct rx_connection *conn)
3693 /* Indicate that rxi_CheckReachEvent is no longer running by
3694 * clearing the flag. Must be atomic under conn_data_lock to
3695 * avoid a new call slipping by: rxi_CheckConnReach holds
3696 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3698 conn->flags &= ~RX_CONN_ATTACHWAIT;
3699 if (conn->flags & RX_CONN_NAT_PING) {
3700 conn->flags &= ~RX_CONN_NAT_PING;
3701 rxi_ScheduleNatKeepAliveEvent(conn);
3706 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3708 struct rx_connection *conn = arg1;
3709 struct rx_call *acall = arg2;
3710 struct rx_call *call = acall;
3711 struct clock when, now;
3714 MUTEX_ENTER(&conn->conn_data_lock);
3717 rxevent_Put(conn->checkReachEvent);
3718 conn->checkReachEvent = NULL;
3721 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3723 putConnection(conn);
3725 MUTEX_EXIT(&conn->conn_data_lock);
3729 MUTEX_ENTER(&conn->conn_call_lock);
3730 MUTEX_ENTER(&conn->conn_data_lock);
3731 for (i = 0; i < RX_MAXCALLS; i++) {
3732 struct rx_call *tc = conn->call[i];
3733 if (tc && tc->state == RX_STATE_PRECALL) {
3739 rxi_ConnClearAttachWait(conn);
3740 MUTEX_EXIT(&conn->conn_data_lock);
3741 MUTEX_EXIT(&conn->conn_call_lock);
3746 MUTEX_ENTER(&call->lock);
3747 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3749 MUTEX_EXIT(&call->lock);
3751 clock_GetTime(&now);
3753 when.sec += RX_CHECKREACH_TIMEOUT;
3754 MUTEX_ENTER(&conn->conn_data_lock);
3755 if (!conn->checkReachEvent) {
3756 MUTEX_ENTER(&rx_refcnt_mutex);
3758 MUTEX_EXIT(&rx_refcnt_mutex);
3759 conn->checkReachEvent = rxevent_Post(&when, &now,
3760 rxi_CheckReachEvent, conn,
3763 MUTEX_EXIT(&conn->conn_data_lock);
3769 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3771 struct rx_service *service = conn->service;
3772 struct rx_peer *peer = conn->peer;
3773 afs_uint32 now, lastReach;
3775 if (service->checkReach == 0)
3779 MUTEX_ENTER(&peer->peer_lock);
3780 lastReach = peer->lastReachTime;
3781 MUTEX_EXIT(&peer->peer_lock);
3782 if (now - lastReach < RX_CHECKREACH_TTL)
3785 MUTEX_ENTER(&conn->conn_data_lock);
3786 if (conn->flags & RX_CONN_ATTACHWAIT) {
3787 MUTEX_EXIT(&conn->conn_data_lock);
3790 conn->flags |= RX_CONN_ATTACHWAIT;
3791 MUTEX_EXIT(&conn->conn_data_lock);
3792 if (!conn->checkReachEvent)
3793 rxi_CheckReachEvent(NULL, conn, call, 0);
3798 /* try to attach call, if authentication is complete */
3800 TryAttach(struct rx_call *acall, osi_socket socket,
3801 int *tnop, struct rx_call **newcallp,
3804 struct rx_connection *conn = acall->conn;
3806 if (conn->type == RX_SERVER_CONNECTION
3807 && acall->state == RX_STATE_PRECALL) {
3808 /* Don't attach until we have any req'd. authentication. */
3809 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3810 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3811 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3812 /* Note: this does not necessarily succeed; there
3813 * may not any proc available
3816 rxi_ChallengeOn(acall->conn);
3821 /* A data packet has been received off the interface. This packet is
3822 * appropriate to the call (the call is in the right state, etc.). This
3823 * routine can return a packet to the caller, for re-use */
3825 static struct rx_packet *
3826 rxi_ReceiveDataPacket(struct rx_call *call,
3827 struct rx_packet *np, int istack,
3828 osi_socket socket, afs_uint32 host, u_short port,
3829 int *tnop, struct rx_call **newcallp)
3831 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3836 afs_uint32 serial=0, flags=0;
3838 struct rx_packet *tnp;
3839 if (rx_stats_active)
3840 rx_atomic_inc(&rx_stats.dataPacketsRead);
3843 /* If there are no packet buffers, drop this new packet, unless we can find
3844 * packet buffers from inactive calls */
3846 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3847 MUTEX_ENTER(&rx_freePktQ_lock);
3848 rxi_NeedMorePackets = TRUE;
3849 MUTEX_EXIT(&rx_freePktQ_lock);
3850 if (rx_stats_active)
3851 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3852 rxi_calltrace(RX_TRACE_DROP, call);
3853 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3854 /* We used to clear the receive queue here, in an attempt to free
3855 * packets. However this is unsafe if the queue has received a
3856 * soft ACK for the final packet */
3857 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3863 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3864 * packet is one of several packets transmitted as a single
3865 * datagram. Do not send any soft or hard acks until all packets
3866 * in a jumbogram have been processed. Send negative acks right away.
3868 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3869 /* tnp is non-null when there are more packets in the
3870 * current jumbo gram */
3877 seq = np->header.seq;
3878 serial = np->header.serial;
3879 flags = np->header.flags;
3881 /* If the call is in an error state, send an abort message */
3883 return rxi_SendCallAbort(call, np, istack, 0);
3885 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3886 * AFS 3.5 jumbogram. */
3887 if (flags & RX_JUMBO_PACKET) {
3888 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3893 if (np->header.spare != 0) {
3894 MUTEX_ENTER(&call->conn->conn_data_lock);
3895 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3896 MUTEX_EXIT(&call->conn->conn_data_lock);
3899 /* The usual case is that this is the expected next packet */
3900 if (seq == call->rnext) {
3902 /* Check to make sure it is not a duplicate of one already queued */
3903 if (!opr_queue_IsEmpty(&call->rq)
3904 && opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq == seq) {
3905 if (rx_stats_active)
3906 rx_atomic_inc(&rx_stats.dupPacketsRead);
3907 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3908 rxevent_Cancel(&call->delayedAckEvent, call,
3909 RX_CALL_REFCOUNT_DELAY);
3910 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3916 /* It's the next packet. Stick it on the receive queue
3917 * for this call. Set newPackets to make sure we wake
3918 * the reader once all packets have been processed */
3919 #ifdef RX_TRACK_PACKETS
3920 np->flags |= RX_PKTFLAG_RQ;
3922 opr_queue_Prepend(&call->rq, &np->entry);
3923 #ifdef RXDEBUG_PACKET
3925 #endif /* RXDEBUG_PACKET */
3927 np = NULL; /* We can't use this anymore */
3930 /* If an ack is requested then set a flag to make sure we
3931 * send an acknowledgement for this packet */
3932 if (flags & RX_REQUEST_ACK) {
3933 ackNeeded = RX_ACK_REQUESTED;
3936 /* Keep track of whether we have received the last packet */
3937 if (flags & RX_LAST_PACKET) {
3938 call->flags |= RX_CALL_HAVE_LAST;
3942 /* Check whether we have all of the packets for this call */
3943 if (call->flags & RX_CALL_HAVE_LAST) {
3944 afs_uint32 tseq; /* temporary sequence number */
3945 struct opr_queue *cursor;
3947 for (tseq = seq, opr_queue_Scan(&call->rq, cursor)) {
3948 struct rx_packet *tp;
3950 tp = opr_queue_Entry(cursor, struct rx_packet, entry);
3951 if (tseq != tp->header.seq)
3953 if (tp->header.flags & RX_LAST_PACKET) {
3954 call->flags |= RX_CALL_RECEIVE_DONE;
3961 /* Provide asynchronous notification for those who want it
3962 * (e.g. multi rx) */
3963 if (call->arrivalProc) {
3964 (*call->arrivalProc) (call, call->arrivalProcHandle,
3965 call->arrivalProcArg);
3966 call->arrivalProc = (void (*)())0;
3969 /* Update last packet received */
3972 /* If there is no server process serving this call, grab
3973 * one, if available. We only need to do this once. If a
3974 * server thread is available, this thread becomes a server
3975 * thread and the server thread becomes a listener thread. */
3977 TryAttach(call, socket, tnop, newcallp, 0);
3980 /* This is not the expected next packet. */
3982 /* Determine whether this is a new or old packet, and if it's
3983 * a new one, whether it fits into the current receive window.
3984 * Also figure out whether the packet was delivered in sequence.
3985 * We use the prev variable to determine whether the new packet
3986 * is the successor of its immediate predecessor in the
3987 * receive queue, and the missing flag to determine whether
3988 * any of this packets predecessors are missing. */
3990 afs_uint32 prev; /* "Previous packet" sequence number */
3991 struct opr_queue *cursor;
3992 int missing; /* Are any predecessors missing? */
3994 /* If the new packet's sequence number has been sent to the
3995 * application already, then this is a duplicate */
3996 if (seq < call->rnext) {
3997 if (rx_stats_active)
3998 rx_atomic_inc(&rx_stats.dupPacketsRead);
3999 rxevent_Cancel(&call->delayedAckEvent, call,
4000 RX_CALL_REFCOUNT_DELAY);
4001 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4007 /* If the sequence number is greater than what can be
4008 * accomodated by the current window, then send a negative
4009 * acknowledge and drop the packet */
4010 if ((call->rnext + call->rwind) <= seq) {
4011 rxevent_Cancel(&call->delayedAckEvent, call,
4012 RX_CALL_REFCOUNT_DELAY);
4013 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4020 /* Look for the packet in the queue of old received packets */
4021 prev = call->rnext - 1;
4023 for (opr_queue_Scan(&call->rq, cursor)) {
4024 struct rx_packet *tp
4025 = opr_queue_Entry(cursor, struct rx_packet, entry);
4027 /*Check for duplicate packet */
4028 if (seq == tp->header.seq) {
4029 if (rx_stats_active)
4030 rx_atomic_inc(&rx_stats.dupPacketsRead);
4031 rxevent_Cancel(&call->delayedAckEvent, call,
4032 RX_CALL_REFCOUNT_DELAY);
4033 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4039 /* If we find a higher sequence packet, break out and
4040 * insert the new packet here. */
4041 if (seq < tp->header.seq)
4043 /* Check for missing packet */
4044 if (tp->header.seq != prev + 1) {
4048 prev = tp->header.seq;
4051 /* Keep track of whether we have received the last packet. */
4052 if (flags & RX_LAST_PACKET) {
4053 call->flags |= RX_CALL_HAVE_LAST;
4056 /* It's within the window: add it to the the receive queue.
4057 * tp is left by the previous loop either pointing at the
4058 * packet before which to insert the new packet, or at the
4059 * queue head if the queue is empty or the packet should be
4061 #ifdef RX_TRACK_PACKETS
4062 np->flags |= RX_PKTFLAG_RQ;
4064 #ifdef RXDEBUG_PACKET
4066 #endif /* RXDEBUG_PACKET */
4067 opr_queue_InsertBefore(cursor, &np->entry);
4071 /* Check whether we have all of the packets for this call */
4072 if ((call->flags & RX_CALL_HAVE_LAST)
4073 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4074 afs_uint32 tseq; /* temporary sequence number */
4077 for (opr_queue_Scan(&call->rq, cursor)) {
4078 struct rx_packet *tp
4079 = opr_queue_Entry(cursor, struct rx_packet, entry);
4080 if (tseq != tp->header.seq)
4082 if (tp->header.flags & RX_LAST_PACKET) {
4083 call->flags |= RX_CALL_RECEIVE_DONE;
4090 /* We need to send an ack of the packet is out of sequence,
4091 * or if an ack was requested by the peer. */
4092 if (seq != prev + 1 || missing) {
4093 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4094 } else if (flags & RX_REQUEST_ACK) {
4095 ackNeeded = RX_ACK_REQUESTED;
4098 /* Acknowledge the last packet for each call */
4099 if (flags & RX_LAST_PACKET) {
4110 * If the receiver is waiting for an iovec, fill the iovec
4111 * using the data from the receive queue */
4112 if (call->flags & RX_CALL_IOVEC_WAIT) {
4113 didHardAck = rxi_FillReadVec(call, serial);
4114 /* the call may have been aborted */
4123 /* Wakeup the reader if any */
4124 if ((call->flags & RX_CALL_READER_WAIT)
4125 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4126 || (call->iovNext >= call->iovMax)
4127 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4128 call->flags &= ~RX_CALL_READER_WAIT;
4129 #ifdef RX_ENABLE_LOCKS
4130 CV_BROADCAST(&call->cv_rq);
4132 osi_rxWakeup(&call->rq);
4138 * Send an ack when requested by the peer, or once every
4139 * rxi_SoftAckRate packets until the last packet has been
4140 * received. Always send a soft ack for the last packet in
4141 * the server's reply. */
4143 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4144 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4145 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4146 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4147 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4148 } else if (call->nSoftAcks) {
4149 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4150 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4152 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4153 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4154 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4161 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4163 struct rx_peer *peer = conn->peer;
4165 MUTEX_ENTER(&peer->peer_lock);
4166 peer->lastReachTime = clock_Sec();
4167 MUTEX_EXIT(&peer->peer_lock);
4169 MUTEX_ENTER(&conn->conn_data_lock);
4170 if (conn->flags & RX_CONN_ATTACHWAIT) {
4173 rxi_ConnClearAttachWait(conn);
4174 MUTEX_EXIT(&conn->conn_data_lock);
4176 for (i = 0; i < RX_MAXCALLS; i++) {
4177 struct rx_call *call = conn->call[i];
4180 MUTEX_ENTER(&call->lock);
4181 /* tnop can be null if newcallp is null */
4182 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4184 MUTEX_EXIT(&call->lock);
4188 MUTEX_EXIT(&conn->conn_data_lock);
4191 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4193 rx_ack_reason(int reason)
4196 case RX_ACK_REQUESTED:
4198 case RX_ACK_DUPLICATE:
4200 case RX_ACK_OUT_OF_SEQUENCE:
4202 case RX_ACK_EXCEEDS_WINDOW:
4204 case RX_ACK_NOSPACE:
4208 case RX_ACK_PING_RESPONSE:
4221 /* The real smarts of the whole thing. */
4222 static struct rx_packet *
4223 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4226 struct rx_ackPacket *ap;
4228 struct rx_packet *tp;
4229 struct rx_connection *conn = call->conn;
4230 struct rx_peer *peer = conn->peer;
4231 struct opr_queue *cursor;
4232 struct clock now; /* Current time, for RTT calculations */
4240 int newAckCount = 0;
4241 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4242 int pktsize = 0; /* Set if we need to update the peer mtu */
4243 int conn_data_locked = 0;
4245 if (rx_stats_active)
4246 rx_atomic_inc(&rx_stats.ackPacketsRead);
4247 ap = (struct rx_ackPacket *)rx_DataOf(np);
4248 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4250 return np; /* truncated ack packet */
4252 /* depends on ack packet struct */
4253 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4254 first = ntohl(ap->firstPacket);
4255 prev = ntohl(ap->previousPacket);
4256 serial = ntohl(ap->serial);
4259 * Ignore ack packets received out of order while protecting
4260 * against peers that set the previousPacket field to a packet
4261 * serial number instead of a sequence number.
4263 if (first < call->tfirst ||
4264 (first == call->tfirst && prev < call->tprev && prev < call->tfirst
4271 if (np->header.flags & RX_SLOW_START_OK) {
4272 call->flags |= RX_CALL_SLOW_START_OK;
4275 if (ap->reason == RX_ACK_PING_RESPONSE)
4276 rxi_UpdatePeerReach(conn, call);
4278 if (conn->lastPacketSizeSeq) {
4279 MUTEX_ENTER(&conn->conn_data_lock);
4280 conn_data_locked = 1;
4281 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4282 pktsize = conn->lastPacketSize;
4283 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4286 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4287 if (!conn_data_locked) {
4288 MUTEX_ENTER(&conn->conn_data_lock);
4289 conn_data_locked = 1;
4291 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4292 /* process mtu ping ack */
4293 pktsize = conn->lastPingSize;
4294 conn->lastPingSizeSer = conn->lastPingSize = 0;
4298 if (conn_data_locked) {
4299 MUTEX_EXIT(&conn->conn_data_lock);
4300 conn_data_locked = 0;
4304 if (rxdebug_active) {
4308 len = _snprintf(msg, sizeof(msg),
4309 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4310 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4311 ntohl(ap->serial), ntohl(ap->previousPacket),
4312 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4313 ap->nAcks, ntohs(ap->bufferSpace) );
4317 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4318 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4322 OutputDebugString(msg);
4324 #else /* AFS_NT40_ENV */
4327 "RACK: reason %x previous %u seq %u serial %u first %u",
4328 ap->reason, ntohl(ap->previousPacket),
4329 (unsigned int)np->header.seq, (unsigned int)serial,
4330 ntohl(ap->firstPacket));
4333 for (offset = 0; offset < nAcks; offset++)
4334 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4339 #endif /* AFS_NT40_ENV */
4342 MUTEX_ENTER(&peer->peer_lock);
4345 * Start somewhere. Can't assume we can send what we can receive,
4346 * but we are clearly receiving.
4348 if (!peer->maxPacketSize)
4349 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4351 if (pktsize > peer->maxPacketSize) {
4352 peer->maxPacketSize = pktsize;
4353 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4354 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4355 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4356 rxi_ScheduleGrowMTUEvent(call, 1);
4361 clock_GetTime(&now);
4363 /* The transmit queue splits into 4 sections.
4365 * The first section is packets which have now been acknowledged
4366 * by a window size change in the ack. These have reached the
4367 * application layer, and may be discarded. These are packets
4368 * with sequence numbers < ap->firstPacket.
4370 * The second section is packets which have sequence numbers in
4371 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4372 * contents of the packet's ack array determines whether these
4373 * packets are acknowledged or not.
4375 * The third section is packets which fall above the range
4376 * addressed in the ack packet. These have not yet been received
4379 * The four section is packets which have not yet been transmitted.
4380 * These packets will have a header.serial of 0.
4383 /* First section - implicitly acknowledged packets that can be
4387 tp = opr_queue_First(&call->tq, struct rx_packet, entry);
4388 while(!opr_queue_IsEnd(&call->tq, &tp->entry) && tp->header.seq < first) {
4389 struct rx_packet *next;
4391 next = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4392 call->tfirst = tp->header.seq + 1;
4394 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4396 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4399 #ifdef RX_ENABLE_LOCKS
4400 /* XXX Hack. Because we have to release the global call lock when sending
4401 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4402 * in rxi_Start sending packets out because packets may move to the
4403 * freePacketQueue as result of being here! So we drop these packets until
4404 * we're safely out of the traversing. Really ugly!
4405 * To make it even uglier, if we're using fine grain locking, we can
4406 * set the ack bits in the packets and have rxi_Start remove the packets
4407 * when it's done transmitting.
4409 if (call->flags & RX_CALL_TQ_BUSY) {
4410 tp->flags |= RX_PKTFLAG_ACKED;
4411 call->flags |= RX_CALL_TQ_SOME_ACKED;
4413 #endif /* RX_ENABLE_LOCKS */
4415 opr_queue_Remove(&tp->entry);
4416 #ifdef RX_TRACK_PACKETS
4417 tp->flags &= ~RX_PKTFLAG_TQ;
4419 #ifdef RXDEBUG_PACKET
4421 #endif /* RXDEBUG_PACKET */
4422 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4427 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4429 /* Second section of the queue - packets for which we are receiving
4432 * Go through the explicit acks/nacks and record the results in
4433 * the waiting packets. These are packets that can't be released
4434 * yet, even with a positive acknowledge. This positive
4435 * acknowledge only means the packet has been received by the
4436 * peer, not that it will be retained long enough to be sent to
4437 * the peer's upper level. In addition, reset the transmit timers
4438 * of any missing packets (those packets that must be missing
4439 * because this packet was out of sequence) */
4441 call->nSoftAcked = 0;
4443 while (!opr_queue_IsEnd(&call->tq, &tp->entry)
4444 && tp->header.seq < first + nAcks) {
4445 /* Set the acknowledge flag per packet based on the
4446 * information in the ack packet. An acknowlegded packet can
4447 * be downgraded when the server has discarded a packet it
4448 * soacked previously, or when an ack packet is received
4449 * out of sequence. */
4450 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4451 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4453 tp->flags |= RX_PKTFLAG_ACKED;
4454 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4461 } else /* RX_ACK_TYPE_NACK */ {
4462 tp->flags &= ~RX_PKTFLAG_ACKED;
4466 tp = opr_queue_Next(&tp->entry, struct rx_packet, entry);
4469 /* We don't need to take any action with the 3rd or 4th section in the
4470 * queue - they're not addressed by the contents of this ACK packet.
4473 /* If the window has been extended by this acknowledge packet,
4474 * then wakeup a sender waiting in alloc for window space, or try
4475 * sending packets now, if he's been sitting on packets due to
4476 * lack of window space */
4477 if (call->tnext < (call->tfirst + call->twind)) {
4478 #ifdef RX_ENABLE_LOCKS
4479 CV_SIGNAL(&call->cv_twind);
4481 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4482 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4483 osi_rxWakeup(&call->twind);
4486 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4487 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4491 /* if the ack packet has a receivelen field hanging off it,
4492 * update our state */
4493 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4496 /* If the ack packet has a "recommended" size that is less than
4497 * what I am using now, reduce my size to match */
4498 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4499 (int)sizeof(afs_int32), &tSize);
4500 tSize = (afs_uint32) ntohl(tSize);
4501 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4503 /* Get the maximum packet size to send to this peer */
4504 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4506 tSize = (afs_uint32) ntohl(tSize);
4507 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4508 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4510 /* sanity check - peer might have restarted with different params.
4511 * If peer says "send less", dammit, send less... Peer should never
4512 * be unable to accept packets of the size that prior AFS versions would
4513 * send without asking. */
4514 if (peer->maxMTU != tSize) {
4515 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4517 peer->maxMTU = tSize;
4518 peer->MTU = MIN(tSize, peer->MTU);
4519 call->MTU = MIN(call->MTU, tSize);
4522 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4525 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4526 (int)sizeof(afs_int32), &tSize);
4527 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4528 if (tSize < call->twind) { /* smaller than our send */
4529 call->twind = tSize; /* window, we must send less... */
4530 call->ssthresh = MIN(call->twind, call->ssthresh);
4531 call->conn->twind[call->channel] = call->twind;
4534 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4535 * network MTU confused with the loopback MTU. Calculate the
4536 * maximum MTU here for use in the slow start code below.
4538 /* Did peer restart with older RX version? */
4539 if (peer->maxDgramPackets > 1) {
4540 peer->maxDgramPackets = 1;
4542 } else if (np->length >=
4543 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4546 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4547 sizeof(afs_int32), &tSize);
4548 tSize = (afs_uint32) ntohl(tSize);
4550 * As of AFS 3.5 we set the send window to match the receive window.
4552 if (tSize < call->twind) {
4553 call->twind = tSize;
4554 call->conn->twind[call->channel] = call->twind;
4555 call->ssthresh = MIN(call->twind, call->ssthresh);
4556 } else if (tSize > call->twind) {
4557 call->twind = tSize;
4558 call->conn->twind[call->channel] = call->twind;
4562 * As of AFS 3.5, a jumbogram is more than one fixed size
4563 * packet transmitted in a single UDP datagram. If the remote
4564 * MTU is smaller than our local MTU then never send a datagram
4565 * larger than the natural MTU.
4568 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4569 (int)sizeof(afs_int32), &tSize);
4570 maxDgramPackets = (afs_uint32) ntohl(tSize);
4571 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4573 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4574 if (maxDgramPackets > 1) {
4575 peer->maxDgramPackets = maxDgramPackets;
4576 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4578 peer->maxDgramPackets = 1;
4579 call->MTU = peer->natMTU;
4581 } else if (peer->maxDgramPackets > 1) {
4582 /* Restarted with lower version of RX */
4583 peer->maxDgramPackets = 1;
4585 } else if (peer->maxDgramPackets > 1
4586 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4587 /* Restarted with lower version of RX */
4588 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4589 peer->natMTU = OLD_MAX_PACKET_SIZE;
4590 peer->MTU = OLD_MAX_PACKET_SIZE;
4591 peer->maxDgramPackets = 1;
4592 peer->nDgramPackets = 1;
4594 call->MTU = OLD_MAX_PACKET_SIZE;
4599 * Calculate how many datagrams were successfully received after
4600 * the first missing packet and adjust the negative ack counter
4605 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4606 if (call->nNacks < nNacked) {
4607 call->nNacks = nNacked;
4610 call->nAcks += newAckCount;
4614 /* If the packet contained new acknowledgements, rather than just
4615 * being a duplicate of one we have previously seen, then we can restart
4618 if (newAckCount > 0)
4619 rxi_rto_packet_acked(call, istack);
4621 if (call->flags & RX_CALL_FAST_RECOVER) {
4622 if (newAckCount == 0) {
4623 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4625 call->flags &= ~RX_CALL_FAST_RECOVER;
4626 call->cwind = call->nextCwind;
4627 call->nextCwind = 0;
4630 call->nCwindAcks = 0;
4631 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4632 /* Three negative acks in a row trigger congestion recovery */
4633 call->flags |= RX_CALL_FAST_RECOVER;
4634 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4636 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4637 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4638 call->nextCwind = call->ssthresh;
4641 peer->MTU = call->MTU;
4642 peer->cwind = call->nextCwind;
4643 peer->nDgramPackets = call->nDgramPackets;
4645 call->congestSeq = peer->congestSeq;
4647 /* Reset the resend times on the packets that were nacked
4648 * so we will retransmit as soon as the window permits
4652 for (opr_queue_ScanBackwards(&call->tq, cursor)) {
4653 struct rx_packet *tp =
4654 opr_queue_Entry(cursor, struct rx_packet, entry);
4656 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4657 tp->flags &= ~RX_PKTFLAG_SENT;
4659 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4664 /* If cwind is smaller than ssthresh, then increase
4665 * the window one packet for each ack we receive (exponential
4667 * If cwind is greater than or equal to ssthresh then increase
4668 * the congestion window by one packet for each cwind acks we
4669 * receive (linear growth). */
4670 if (call->cwind < call->ssthresh) {
4672 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4673 call->nCwindAcks = 0;
4675 call->nCwindAcks += newAckCount;
4676 if (call->nCwindAcks >= call->cwind) {
4677 call->nCwindAcks = 0;
4678 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4682 * If we have received several acknowledgements in a row then
4683 * it is time to increase the size of our datagrams
4685 if ((int)call->nAcks > rx_nDgramThreshold) {
4686 if (peer->maxDgramPackets > 1) {
4687 if (call->nDgramPackets < peer->maxDgramPackets) {
4688 call->nDgramPackets++;
4690 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4691 } else if (call->MTU < peer->maxMTU) {
4692 /* don't upgrade if we can't handle it */
4693 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4694 call->MTU = peer->ifMTU;
4696 call->MTU += peer->natMTU;
4697 call->MTU = MIN(call->MTU, peer->maxMTU);
4704 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4706 /* Servers need to hold the call until all response packets have
4707 * been acknowledged. Soft acks are good enough since clients
4708 * are not allowed to clear their receive queues. */
4709 if (call->state == RX_STATE_HOLD
4710 && call->tfirst + call->nSoftAcked >= call->tnext) {
4711 call->state = RX_STATE_DALLY;
4712 rxi_ClearTransmitQueue(call, 0);
4713 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4714 } else if (!opr_queue_IsEmpty(&call->tq)) {
4715 rxi_Start(call, istack);
4720 /* Received a response to a challenge packet */
4721 static struct rx_packet *
4722 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4723 struct rx_packet *np, int istack)
4727 /* Ignore the packet if we're the client */
4728 if (conn->type == RX_CLIENT_CONNECTION)
4731 /* If already authenticated, ignore the packet (it's probably a retry) */
4732 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4735 /* Otherwise, have the security object evaluate the response packet */
4736 error = RXS_CheckResponse(conn->securityObject, conn, np);
4738 /* If the response is invalid, reset the connection, sending
4739 * an abort to the peer */
4743 rxi_ConnectionError(conn, error);
4744 MUTEX_ENTER(&conn->conn_data_lock);
4745 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4746 MUTEX_EXIT(&conn->conn_data_lock);
4749 /* If the response is valid, any calls waiting to attach
4750 * servers can now do so */
4753 for (i = 0; i < RX_MAXCALLS; i++) {
4754 struct rx_call *call = conn->call[i];
4756 MUTEX_ENTER(&call->lock);
4757 if (call->state == RX_STATE_PRECALL)
4758 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4759 /* tnop can be null if newcallp is null */
4760 MUTEX_EXIT(&call->lock);
4764 /* Update the peer reachability information, just in case
4765 * some calls went into attach-wait while we were waiting
4766 * for authentication..
4768 rxi_UpdatePeerReach(conn, NULL);
4773 /* A client has received an authentication challenge: the security
4774 * object is asked to cough up a respectable response packet to send
4775 * back to the server. The server is responsible for retrying the
4776 * challenge if it fails to get a response. */
4778 static struct rx_packet *
4779 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4780 struct rx_packet *np, int istack)
4784 /* Ignore the challenge if we're the server */
4785 if (conn->type == RX_SERVER_CONNECTION)
4788 /* Ignore the challenge if the connection is otherwise idle; someone's
4789 * trying to use us as an oracle. */
4790 if (!rxi_HasActiveCalls(conn))
4793 /* Send the security object the challenge packet. It is expected to fill
4794 * in the response. */
4795 error = RXS_GetResponse(conn->securityObject, conn, np);
4797 /* If the security object is unable to return a valid response, reset the
4798 * connection and send an abort to the peer. Otherwise send the response
4799 * packet to the peer connection. */
4801 rxi_ConnectionError(conn, error);
4802 MUTEX_ENTER(&conn->conn_data_lock);
4803 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4804 MUTEX_EXIT(&conn->conn_data_lock);
4806 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4807 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4813 /* Find an available server process to service the current request in
4814 * the given call structure. If one isn't available, queue up this
4815 * call so it eventually gets one */
4817 rxi_AttachServerProc(struct rx_call *call,
4818 osi_socket socket, int *tnop,
4819 struct rx_call **newcallp)
4821 struct rx_serverQueueEntry *sq;
4822 struct rx_service *service = call->conn->service;
4825 /* May already be attached */
4826 if (call->state == RX_STATE_ACTIVE)
4829 MUTEX_ENTER(&rx_serverPool_lock);
4831 haveQuota = QuotaOK(service);
4832 if ((!haveQuota) || opr_queue_IsEmpty(&rx_idleServerQueue)) {
4833 /* If there are no processes available to service this call,
4834 * put the call on the incoming call queue (unless it's
4835 * already on the queue).
4837 #ifdef RX_ENABLE_LOCKS
4839 ReturnToServerPool(service);
4840 #endif /* RX_ENABLE_LOCKS */
4842 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4843 call->flags |= RX_CALL_WAIT_PROC;
4844 rx_atomic_inc(&rx_nWaiting);
4845 rx_atomic_inc(&rx_nWaited);
4846 rxi_calltrace(RX_CALL_ARRIVAL, call);
4847 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4848 opr_queue_Append(&rx_incomingCallQueue, &call->entry);
4851 sq = opr_queue_Last(&rx_idleServerQueue,
4852 struct rx_serverQueueEntry, entry);
4854 /* If hot threads are enabled, and both newcallp and sq->socketp
4855 * are non-null, then this thread will process the call, and the
4856 * idle server thread will start listening on this threads socket.
4858 opr_queue_Remove(&sq->entry);
4860 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4863 *sq->socketp = socket;
4864 clock_GetTime(&call->startTime);
4865 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4869 if (call->flags & RX_CALL_WAIT_PROC) {
4870 /* Conservative: I don't think this should happen */
4871 call->flags &= ~RX_CALL_WAIT_PROC;
4872 rx_atomic_dec(&rx_nWaiting);
4873 if (opr_queue_IsOnQueue(&call->entry)) {
4874 opr_queue_Remove(&call->entry);
4877 call->state = RX_STATE_ACTIVE;
4878 call->app.mode = RX_MODE_RECEIVING;
4879 #ifdef RX_KERNEL_TRACE
4881 int glockOwner = ISAFS_GLOCK();
4884 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4885 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4891 if (call->flags & RX_CALL_CLEARED) {
4892 /* send an ack now to start the packet flow up again */
4893 call->flags &= ~RX_CALL_CLEARED;
4894 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4896 #ifdef RX_ENABLE_LOCKS
4899 service->nRequestsRunning++;
4900 MUTEX_ENTER(&rx_quota_mutex);
4901 if (service->nRequestsRunning <= service->minProcs)
4904 MUTEX_EXIT(&rx_quota_mutex);
4908 MUTEX_EXIT(&rx_serverPool_lock);
4911 /* Delay the sending of an acknowledge event for a short while, while
4912 * a new call is being prepared (in the case of a client) or a reply
4913 * is being prepared (in the case of a server). Rather than sending
4914 * an ack packet, an ACKALL packet is sent. */
4916 rxi_AckAll(struct rx_call *call)
4918 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4920 call->flags |= RX_CALL_ACKALL_SENT;
4924 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4927 struct rx_call *call = arg1;
4928 #ifdef RX_ENABLE_LOCKS
4930 MUTEX_ENTER(&call->lock);
4931 if (event == call->delayedAckEvent) {
4932 rxevent_Put(call->delayedAckEvent);
4933 call->delayedAckEvent = NULL;
4935 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4937 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4939 MUTEX_EXIT(&call->lock);
4940 #else /* RX_ENABLE_LOCKS */
4942 rxevent_Put(call->delayedAckEvent);
4943 call->delayedAckEvent = NULL;
4945 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4946 #endif /* RX_ENABLE_LOCKS */
4949 #ifdef RX_ENABLE_LOCKS
4950 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4951 * clearing them out.
4954 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4956 struct opr_queue *cursor;
4959 for (opr_queue_Scan(&call->tq, cursor)) {
4961 = opr_queue_Entry(cursor, struct rx_packet, entry);
4963 p->flags |= RX_PKTFLAG_ACKED;
4968 call->flags |= RX_CALL_TQ_CLEARME;
4969 call->flags |= RX_CALL_TQ_SOME_ACKED;
4972 rxi_rto_cancel(call);
4974 call->tfirst = call->tnext;
4975 call->nSoftAcked = 0;
4977 if (call->flags & RX_CALL_FAST_RECOVER) {
4978 call->flags &= ~RX_CALL_FAST_RECOVER;
4979 call->cwind = call->nextCwind;
4980 call->nextCwind = 0;
4983 CV_SIGNAL(&call->cv_twind);
4985 #endif /* RX_ENABLE_LOCKS */
4988 * Acknowledge the whole transmit queue.
4990 * If we're running without locks, or the transmit queue isn't busy, then
4991 * we can just clear the queue now. Otherwise, we have to mark all of the
4992 * packets as acknowledged, and let rxi_Start clear it later on
4995 rxi_AckAllInTransmitQueue(struct rx_call *call)
4997 #ifdef RX_ENABLE_LOCKS
4998 if (call->flags & RX_CALL_TQ_BUSY) {
4999 rxi_SetAcksInTransmitQueue(call);
5003 rxi_ClearTransmitQueue(call, 0);
5005 /* Clear out the transmit queue for the current call (all packets have
5006 * been received by peer) */
5008 rxi_ClearTransmitQueue(struct rx_call *call, int force)
5010 #ifdef RX_ENABLE_LOCKS
5011 struct opr_queue *cursor;
5012 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
5014 for (opr_queue_Scan(&call->tq, cursor)) {
5016 = opr_queue_Entry(cursor, struct rx_packet, entry);
5018 p->flags |= RX_PKTFLAG_ACKED;
5022 call->flags |= RX_CALL_TQ_CLEARME;
5023 call->flags |= RX_CALL_TQ_SOME_ACKED;
5026 #endif /* RX_ENABLE_LOCKS */
5027 #ifdef RXDEBUG_PACKET
5029 #endif /* RXDEBUG_PACKET */
5030 rxi_FreePackets(0, &call->tq);
5031 rxi_WakeUpTransmitQueue(call);
5032 #ifdef RX_ENABLE_LOCKS
5033 call->flags &= ~RX_CALL_TQ_CLEARME;
5037 rxi_rto_cancel(call);
5038 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5039 call->nSoftAcked = 0;
5041 if (call->flags & RX_CALL_FAST_RECOVER) {
5042 call->flags &= ~RX_CALL_FAST_RECOVER;
5043 call->cwind = call->nextCwind;
5045 #ifdef RX_ENABLE_LOCKS
5046 CV_SIGNAL(&call->cv_twind);
5048 osi_rxWakeup(&call->twind);
5053 rxi_ClearReceiveQueue(struct rx_call *call)
5055 if (!opr_queue_IsEmpty(&call->rq)) {
5058 count = rxi_FreePackets(0, &call->rq);
5059 rx_packetReclaims += count;
5060 #ifdef RXDEBUG_PACKET
5062 if ( call->rqc != 0 )
5063 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5065 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5067 if (call->state == RX_STATE_PRECALL) {
5068 call->flags |= RX_CALL_CLEARED;
5072 /* Send an abort packet for the specified call */
5073 static struct rx_packet *
5074 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5075 int istack, int force)
5077 afs_int32 error, cerror;
5078 struct clock when, now;
5083 switch (call->error) {
5086 cerror = RX_CALL_TIMEOUT;
5089 cerror = call->error;
5092 /* Clients should never delay abort messages */
5093 if (rx_IsClientConn(call->conn))
5096 if (call->abortCode != cerror) {
5097 call->abortCode = cerror;
5098 call->abortCount = 0;
5101 if (force || rxi_callAbortThreshhold == 0
5102 || call->abortCount < rxi_callAbortThreshhold) {
5103 if (call->delayedAbortEvent) {
5104 rxevent_Cancel(&call->delayedAbortEvent, call,
5105 RX_CALL_REFCOUNT_ABORT);
5107 error = htonl(cerror);
5110 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5111 (char *)&error, sizeof(error), istack);
5112 } else if (!call->delayedAbortEvent) {
5113 clock_GetTime(&now);
5115 clock_Addmsec(&when, rxi_callAbortDelay);
5116 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5117 call->delayedAbortEvent =
5118 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5123 /* Send an abort packet for the specified connection. Packet is an
5124 * optional pointer to a packet that can be used to send the abort.
5125 * Once the number of abort messages reaches the threshhold, an
5126 * event is scheduled to send the abort. Setting the force flag
5127 * overrides sending delayed abort messages.
5129 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5130 * to send the abort packet.
5133 rxi_SendConnectionAbort(struct rx_connection *conn,
5134 struct rx_packet *packet, int istack, int force)
5137 struct clock when, now;
5142 /* Clients should never delay abort messages */
5143 if (rx_IsClientConn(conn))
5146 if (force || rxi_connAbortThreshhold == 0
5147 || conn->abortCount < rxi_connAbortThreshhold) {
5149 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5150 error = htonl(conn->error);
5152 MUTEX_EXIT(&conn->conn_data_lock);
5154 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5155 RX_PACKET_TYPE_ABORT, (char *)&error,
5156 sizeof(error), istack);
5157 MUTEX_ENTER(&conn->conn_data_lock);
5158 } else if (!conn->delayedAbortEvent) {
5159 clock_GetTime(&now);
5161 clock_Addmsec(&when, rxi_connAbortDelay);
5162 conn->delayedAbortEvent =
5163 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5168 /* Associate an error all of the calls owned by a connection. Called
5169 * with error non-zero. This is only for really fatal things, like
5170 * bad authentication responses. The connection itself is set in
5171 * error at this point, so that future packets received will be
5174 rxi_ConnectionError(struct rx_connection *conn,
5180 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5182 MUTEX_ENTER(&conn->conn_data_lock);
5183 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5184 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5185 if (conn->checkReachEvent) {
5186 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5187 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5188 putConnection(conn);
5190 MUTEX_EXIT(&conn->conn_data_lock);
5191 for (i = 0; i < RX_MAXCALLS; i++) {
5192 struct rx_call *call = conn->call[i];
5194 MUTEX_ENTER(&call->lock);
5195 rxi_CallError(call, error);
5196 MUTEX_EXIT(&call->lock);
5199 conn->error = error;
5200 if (rx_stats_active)
5201 rx_atomic_inc(&rx_stats.fatalErrors);
5206 * Interrupt an in-progress call with the specified error and wakeup waiters.
5208 * @param[in] call The call to interrupt
5209 * @param[in] error The error code to send to the peer
5212 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5214 MUTEX_ENTER(&call->lock);
5215 rxi_CallError(call, error);
5216 rxi_SendCallAbort(call, NULL, 0, 1);
5217 MUTEX_EXIT(&call->lock);
5221 rxi_CallError(struct rx_call *call, afs_int32 error)
5224 osirx_AssertMine(&call->lock, "rxi_CallError");
5226 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5228 error = call->error;
5230 #ifdef RX_ENABLE_LOCKS
5231 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5232 rxi_ResetCall(call, 0);
5235 rxi_ResetCall(call, 0);
5237 call->error = error;
5240 /* Reset various fields in a call structure, and wakeup waiting
5241 * processes. Some fields aren't changed: state & mode are not
5242 * touched (these must be set by the caller), and bufptr, nLeft, and
5243 * nFree are not reset, since these fields are manipulated by
5244 * unprotected macros, and may only be reset by non-interrupting code.
5248 rxi_ResetCall(struct rx_call *call, int newcall)
5251 struct rx_peer *peer;
5252 struct rx_packet *packet;
5254 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5256 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5258 /* Notify anyone who is waiting for asynchronous packet arrival */
5259 if (call->arrivalProc) {
5260 (*call->arrivalProc) (call, call->arrivalProcHandle,
5261 call->arrivalProcArg);
5262 call->arrivalProc = (void (*)())0;
5266 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5268 if (call->delayedAbortEvent) {
5269 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5270 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5272 rxi_SendCallAbort(call, packet, 0, 1);
5273 rxi_FreePacket(packet);
5278 * Update the peer with the congestion information in this call
5279 * so other calls on this connection can pick up where this call
5280 * left off. If the congestion sequence numbers don't match then
5281 * another call experienced a retransmission.
5283 peer = call->conn->peer;
5284 MUTEX_ENTER(&peer->peer_lock);
5286 if (call->congestSeq == peer->congestSeq) {
5287 peer->cwind = MAX(peer->cwind, call->cwind);
5288 peer->MTU = MAX(peer->MTU, call->MTU);
5289 peer->nDgramPackets =
5290 MAX(peer->nDgramPackets, call->nDgramPackets);
5293 call->abortCode = 0;
5294 call->abortCount = 0;
5296 if (peer->maxDgramPackets > 1) {
5297 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5299 call->MTU = peer->MTU;
5301 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5302 call->ssthresh = rx_maxSendWindow;
5303 call->nDgramPackets = peer->nDgramPackets;
5304 call->congestSeq = peer->congestSeq;
5305 call->rtt = peer->rtt;
5306 call->rtt_dev = peer->rtt_dev;
5307 clock_Zero(&call->rto);
5308 clock_Addmsec(&call->rto,
5309 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5310 MUTEX_EXIT(&peer->peer_lock);
5312 flags = call->flags;
5313 rxi_WaitforTQBusy(call);
5315 rxi_ClearTransmitQueue(call, 1);
5316 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5317 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5321 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5322 /* The call channel is still busy; resetting the call doesn't change
5323 * that. However, if 'newcall' is set, we are processing a call
5324 * structure that has either been recycled from the free list, or has
5325 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5326 * 'newcall' is set, since it describes a completely different call
5327 * channel which we do not care about. */
5328 call->flags |= RX_CALL_PEER_BUSY;
5331 rxi_ClearReceiveQueue(call);
5332 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5336 call->twind = call->conn->twind[call->channel];
5337 call->rwind = call->conn->rwind[call->channel];
5338 call->nSoftAcked = 0;
5339 call->nextCwind = 0;
5342 call->nCwindAcks = 0;
5343 call->nSoftAcks = 0;
5344 call->nHardAcks = 0;
5346 call->tfirst = call->rnext = call->tnext = 1;
5349 call->lastAcked = 0;
5350 call->localStatus = call->remoteStatus = 0;
5352 if (flags & RX_CALL_READER_WAIT) {
5353 #ifdef RX_ENABLE_LOCKS
5354 CV_BROADCAST(&call->cv_rq);
5356 osi_rxWakeup(&call->rq);
5359 if (flags & RX_CALL_WAIT_PACKETS) {
5360 MUTEX_ENTER(&rx_freePktQ_lock);
5361 rxi_PacketsUnWait(); /* XXX */
5362 MUTEX_EXIT(&rx_freePktQ_lock);
5364 #ifdef RX_ENABLE_LOCKS
5365 CV_SIGNAL(&call->cv_twind);
5367 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5368 osi_rxWakeup(&call->twind);
5371 if (flags & RX_CALL_WAIT_PROC) {
5372 rx_atomic_dec(&rx_nWaiting);
5374 #ifdef RX_ENABLE_LOCKS
5375 /* The following ensures that we don't mess with any queue while some
5376 * other thread might also be doing so. The call_queue_lock field is
5377 * is only modified under the call lock. If the call is in the process
5378 * of being removed from a queue, the call is not locked until the
5379 * the queue lock is dropped and only then is the call_queue_lock field
5380 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5381 * Note that any other routine which removes a call from a queue has to
5382 * obtain the queue lock before examing the queue and removing the call.
5384 if (call->call_queue_lock) {
5385 MUTEX_ENTER(call->call_queue_lock);
5386 if (opr_queue_IsOnQueue(&call->entry)) {
5387 opr_queue_Remove(&call->entry);
5389 MUTEX_EXIT(call->call_queue_lock);
5390 CLEAR_CALL_QUEUE_LOCK(call);
5392 #else /* RX_ENABLE_LOCKS */
5393 if (opr_queue_IsOnQueue(&call->entry)) {
5394 opr_queue_Remove(&call->entry);
5396 #endif /* RX_ENABLE_LOCKS */
5398 rxi_KeepAliveOff(call);
5399 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5402 /* Send an acknowledge for the indicated packet (seq,serial) of the
5403 * indicated call, for the indicated reason (reason). This
5404 * acknowledge will specifically acknowledge receiving the packet, and
5405 * will also specify which other packets for this call have been
5406 * received. This routine returns the packet that was used to the
5407 * caller. The caller is responsible for freeing it or re-using it.
5408 * This acknowledgement also returns the highest sequence number
5409 * actually read out by the higher level to the sender; the sender
5410 * promises to keep around packets that have not been read by the
5411 * higher level yet (unless, of course, the sender decides to abort
5412 * the call altogether). Any of p, seq, serial, pflags, or reason may
5413 * be set to zero without ill effect. That is, if they are zero, they
5414 * will not convey any information.
5415 * NOW there is a trailer field, after the ack where it will safely be
5416 * ignored by mundanes, which indicates the maximum size packet this
5417 * host can swallow. */
5419 struct rx_packet *optionalPacket; use to send ack (or null)
5420 int seq; Sequence number of the packet we are acking
5421 int serial; Serial number of the packet
5422 int pflags; Flags field from packet header
5423 int reason; Reason an acknowledge was prompted
5427 rxi_SendAck(struct rx_call *call,
5428 struct rx_packet *optionalPacket, int serial, int reason,
5431 struct rx_ackPacket *ap;
5432 struct rx_packet *p;
5433 struct opr_queue *cursor;
5436 afs_uint32 padbytes = 0;
5437 #ifdef RX_ENABLE_TSFPQ
5438 struct rx_ts_info_t * rx_ts_info;
5442 * Open the receive window once a thread starts reading packets
5444 if (call->rnext > 1) {
5445 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5448 /* Don't attempt to grow MTU if this is a critical ping */
5449 if (reason == RX_ACK_MTU) {
5450 /* keep track of per-call attempts, if we're over max, do in small
5451 * otherwise in larger? set a size to increment by, decrease
5454 if (call->conn->peer->maxPacketSize &&
5455 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5457 padbytes = call->conn->peer->maxPacketSize+16;
5459 padbytes = call->conn->peer->maxMTU + 128;
5461 /* do always try a minimum size ping */
5462 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5464 /* subtract the ack payload */
5465 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5466 reason = RX_ACK_PING;
5469 call->nHardAcks = 0;
5470 call->nSoftAcks = 0;
5471 if (call->rnext > call->lastAcked)
5472 call->lastAcked = call->rnext;
5476 rx_computelen(p, p->length); /* reset length, you never know */
5477 } /* where that's been... */
5478 #ifdef RX_ENABLE_TSFPQ
5480 RX_TS_INFO_GET(rx_ts_info);
5481 if ((p = rx_ts_info->local_special_packet)) {
5482 rx_computelen(p, p->length);
5483 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5484 rx_ts_info->local_special_packet = p;
5485 } else { /* We won't send the ack, but don't panic. */
5486 return optionalPacket;
5490 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5491 /* We won't send the ack, but don't panic. */
5492 return optionalPacket;
5497 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5500 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5501 #ifndef RX_ENABLE_TSFPQ
5502 if (!optionalPacket)
5505 return optionalPacket;
5507 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5508 if (rx_Contiguous(p) < templ) {
5509 #ifndef RX_ENABLE_TSFPQ
5510 if (!optionalPacket)
5513 return optionalPacket;
5518 /* MTUXXX failing to send an ack is very serious. We should */
5519 /* try as hard as possible to send even a partial ack; it's */
5520 /* better than nothing. */
5521 ap = (struct rx_ackPacket *)rx_DataOf(p);
5522 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5523 ap->reason = reason;
5525 /* The skew computation used to be bogus, I think it's better now. */
5526 /* We should start paying attention to skew. XXX */
5527 ap->serial = htonl(serial);
5528 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5531 * First packet not yet forwarded to reader. When ACKALL has been
5532 * sent the peer has been told that all received packets will be
5533 * delivered to the reader. The value 'rnext' is used internally
5534 * to refer to the next packet in the receive queue that must be
5535 * delivered to the reader. From the perspective of the peer it
5536 * already has so report the last sequence number plus one if there
5537 * are packets in the receive queue awaiting processing.
5539 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5540 !opr_queue_IsEmpty(&call->rq)) {
5541 ap->firstPacket = htonl(opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq + 1);
5543 ap->firstPacket = htonl(call->rnext);
5545 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5547 /* No fear of running out of ack packet here because there can only
5548 * be at most one window full of unacknowledged packets. The window
5549 * size must be constrained to be less than the maximum ack size,
5550 * of course. Also, an ack should always fit into a single packet
5551 * -- it should not ever be fragmented. */
5553 for (opr_queue_Scan(&call->rq, cursor)) {
5554 struct rx_packet *rqp
5555 = opr_queue_Entry(cursor, struct rx_packet, entry);
5557 if (!rqp || !call->rq.next
5558 || (rqp->header.seq > (call->rnext + call->rwind))) {
5559 #ifndef RX_ENABLE_TSFPQ
5560 if (!optionalPacket)
5563 rxi_CallError(call, RX_CALL_DEAD);
5564 return optionalPacket;
5567 while (rqp->header.seq > call->rnext + offset)
5568 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5569 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5571 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5572 #ifndef RX_ENABLE_TSFPQ
5573 if (!optionalPacket)
5576 rxi_CallError(call, RX_CALL_DEAD);
5577 return optionalPacket;
5583 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5585 /* these are new for AFS 3.3 */
5586 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5587 templ = htonl(templ);
5588 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5589 templ = htonl(call->conn->peer->ifMTU);
5590 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5591 sizeof(afs_int32), &templ);
5593 /* new for AFS 3.4 */
5594 templ = htonl(call->rwind);
5595 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5596 sizeof(afs_int32), &templ);
5598 /* new for AFS 3.5 */
5599 templ = htonl(call->conn->peer->ifDgramPackets);
5600 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5601 sizeof(afs_int32), &templ);
5603 p->header.serviceId = call->conn->serviceId;
5604 p->header.cid = (call->conn->cid | call->channel);
5605 p->header.callNumber = *call->callNumber;
5607 p->header.securityIndex = call->conn->securityIndex;
5608 p->header.epoch = call->conn->epoch;
5609 p->header.type = RX_PACKET_TYPE_ACK;
5610 p->header.flags = RX_SLOW_START_OK;
5611 if (reason == RX_ACK_PING) {
5612 p->header.flags |= RX_REQUEST_ACK;
5614 p->length = padbytes +
5615 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5618 /* not fast but we can potentially use this if truncated
5619 * fragments are delivered to figure out the mtu.
5621 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5622 sizeof(afs_int32), sizeof(afs_int32),
5626 if (call->conn->type == RX_CLIENT_CONNECTION)
5627 p->header.flags |= RX_CLIENT_INITIATED;
5631 if (rxdebug_active) {
5635 len = _snprintf(msg, sizeof(msg),
5636 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5637 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5638 ntohl(ap->serial), ntohl(ap->previousPacket),
5639 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5640 ap->nAcks, ntohs(ap->bufferSpace) );
5644 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5645 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5649 OutputDebugString(msg);
5651 #else /* AFS_NT40_ENV */
5653 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5654 ap->reason, ntohl(ap->previousPacket),
5655 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5657 for (offset = 0; offset < ap->nAcks; offset++)
5658 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5663 #endif /* AFS_NT40_ENV */
5666 int i, nbytes = p->length;
5668 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5669 if (nbytes <= p->wirevec[i].iov_len) {
5672 savelen = p->wirevec[i].iov_len;
5674 p->wirevec[i].iov_len = nbytes;
5676 rxi_Send(call, p, istack);
5677 p->wirevec[i].iov_len = savelen;
5681 nbytes -= p->wirevec[i].iov_len;
5684 if (rx_stats_active)
5685 rx_atomic_inc(&rx_stats.ackPacketsSent);
5686 #ifndef RX_ENABLE_TSFPQ
5687 if (!optionalPacket)
5690 return optionalPacket; /* Return packet for re-use by caller */
5694 struct rx_packet **list;
5699 /* Send all of the packets in the list in single datagram */
5701 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5702 int istack, int moreFlag)
5708 struct rx_connection *conn = call->conn;
5709 struct rx_peer *peer = conn->peer;
5711 MUTEX_ENTER(&peer->peer_lock);
5712 peer->nSent += xmit->len;
5713 if (xmit->resending)
5714 peer->reSends += xmit->len;
5715 MUTEX_EXIT(&peer->peer_lock);
5717 if (rx_stats_active) {
5718 if (xmit->resending)
5719 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5721 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5724 clock_GetTime(&now);
5726 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5730 /* Set the packet flags and schedule the resend events */
5731 /* Only request an ack for the last packet in the list */
5732 for (i = 0; i < xmit->len; i++) {
5733 struct rx_packet *packet = xmit->list[i];
5735 /* Record the time sent */
5736 packet->timeSent = now;
5737 packet->flags |= RX_PKTFLAG_SENT;
5739 /* Ask for an ack on retransmitted packets, on every other packet
5740 * if the peer doesn't support slow start. Ask for an ack on every
5741 * packet until the congestion window reaches the ack rate. */
5742 if (packet->header.serial) {
5745 packet->firstSent = now;
5746 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5747 || (!(call->flags & RX_CALL_SLOW_START_OK)
5748 && (packet->header.seq & 1)))) {
5753 /* Tag this packet as not being the last in this group,
5754 * for the receiver's benefit */
5755 if (i < xmit->len - 1 || moreFlag) {
5756 packet->header.flags |= RX_MORE_PACKETS;
5761 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5764 /* Since we're about to send a data packet to the peer, it's
5765 * safe to nuke any scheduled end-of-packets ack */
5766 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5768 MUTEX_EXIT(&call->lock);
5769 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5770 if (xmit->len > 1) {
5771 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5773 rxi_SendPacket(call, conn, xmit->list[0], istack);
5775 MUTEX_ENTER(&call->lock);
5776 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5778 /* Tell the RTO calculation engine that we have sent a packet, and
5779 * if it was the last one */
5780 rxi_rto_packet_sent(call, lastPacket, istack);
5782 /* Update last send time for this call (for keep-alive
5783 * processing), and for the connection (so that we can discover
5784 * idle connections) */
5785 conn->lastSendTime = call->lastSendTime = clock_Sec();
5786 /* Let a set of retransmits trigger an idle timeout */
5787 if (!xmit->resending)
5788 call->lastSendData = call->lastSendTime;
5791 /* When sending packets we need to follow these rules:
5792 * 1. Never send more than maxDgramPackets in a jumbogram.
5793 * 2. Never send a packet with more than two iovecs in a jumbogram.
5794 * 3. Never send a retransmitted packet in a jumbogram.
5795 * 4. Never send more than cwind/4 packets in a jumbogram
5796 * We always keep the last list we should have sent so we
5797 * can set the RX_MORE_PACKETS flags correctly.
5801 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5806 struct xmitlist working;
5807 struct xmitlist last;
5809 struct rx_peer *peer = call->conn->peer;
5810 int morePackets = 0;
5812 memset(&last, 0, sizeof(struct xmitlist));
5813 working.list = &list[0];
5815 working.resending = 0;
5817 recovery = call->flags & RX_CALL_FAST_RECOVER;
5819 for (i = 0; i < len; i++) {
5820 /* Does the current packet force us to flush the current list? */
5822 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5823 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5825 /* This sends the 'last' list and then rolls the current working
5826 * set into the 'last' one, and resets the working set */
5829 rxi_SendList(call, &last, istack, 1);
5830 /* If the call enters an error state stop sending, or if
5831 * we entered congestion recovery mode, stop sending */
5833 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5838 working.resending = 0;
5839 working.list = &list[i];
5841 /* Add the current packet to the list if it hasn't been acked.
5842 * Otherwise adjust the list pointer to skip the current packet. */
5843 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5846 if (list[i]->header.serial)
5847 working.resending = 1;
5849 /* Do we need to flush the list? */
5850 if (working.len >= (int)peer->maxDgramPackets
5851 || working.len >= (int)call->nDgramPackets
5852 || working.len >= (int)call->cwind
5853 || list[i]->header.serial
5854 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5856 rxi_SendList(call, &last, istack, 1);
5857 /* If the call enters an error state stop sending, or if
5858 * we entered congestion recovery mode, stop sending */
5860 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5865 working.resending = 0;
5866 working.list = &list[i + 1];
5869 if (working.len != 0) {
5870 osi_Panic("rxi_SendList error");
5872 working.list = &list[i + 1];
5876 /* Send the whole list when the call is in receive mode, when
5877 * the call is in eof mode, when we are in fast recovery mode,
5878 * and when we have the last packet */
5879 /* XXX - The accesses to app.mode aren't safe, as this may be called by
5880 * the listener or event threads
5882 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5883 || (call->flags & RX_CALL_FLUSH)
5884 || (call->flags & RX_CALL_FAST_RECOVER)) {
5885 /* Check for the case where the current list contains
5886 * an acked packet. Since we always send retransmissions
5887 * in a separate packet, we only need to check the first
5888 * packet in the list */
5889 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5893 rxi_SendList(call, &last, istack, morePackets);
5894 /* If the call enters an error state stop sending, or if
5895 * we entered congestion recovery mode, stop sending */
5897 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5901 rxi_SendList(call, &working, istack, 0);
5903 } else if (last.len > 0) {
5904 rxi_SendList(call, &last, istack, 0);
5905 /* Packets which are in 'working' are not sent by this call */
5910 * Check if the peer for the given call is known to be dead
5912 * If the call's peer appears dead (it has encountered fatal network errors
5913 * since the call started) the call is killed with RX_CALL_DEAD if the call
5914 * is active. Otherwise, we do nothing.
5916 * @param[in] call The call to check
5919 * @retval 0 The call is fine, and we haven't done anything to the call
5920 * @retval nonzero The call's peer appears dead, and the call has been
5921 * terminated if it was active
5923 * @pre call->lock must be locked
5926 rxi_CheckPeerDead(struct rx_call *call)
5928 #ifdef AFS_RXERRQ_ENV
5931 if (call->state == RX_STATE_DALLY) {
5935 peererrs = rx_atomic_read(&call->conn->peer->neterrs);
5936 if (call->neterr_gen < peererrs) {
5937 /* we have received network errors since this call started; kill
5939 if (call->state == RX_STATE_ACTIVE) {
5940 rxi_CallError(call, RX_CALL_DEAD);
5944 if (call->neterr_gen > peererrs) {
5945 /* someone has reset the number of peer errors; set the call error gen
5946 * so we can detect if more errors are encountered */
5947 call->neterr_gen = peererrs;
5954 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5956 struct rx_call *call = arg0;
5957 struct rx_peer *peer;
5958 struct opr_queue *cursor;
5959 struct clock maxTimeout = { 60, 0 };
5961 MUTEX_ENTER(&call->lock);
5963 peer = call->conn->peer;
5965 /* Make sure that the event pointer is removed from the call
5966 * structure, since there is no longer a per-call retransmission
5968 if (event == call->resendEvent) {
5969 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5970 rxevent_Put(call->resendEvent);
5971 call->resendEvent = NULL;
5974 rxi_CheckPeerDead(call);
5976 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5977 rxi_CheckBusy(call);
5980 if (opr_queue_IsEmpty(&call->tq)) {
5981 /* Nothing to do. This means that we've been raced, and that an
5982 * ACK has come in between when we were triggered, and when we
5983 * actually got to run. */
5987 /* We're in loss recovery */
5988 call->flags |= RX_CALL_FAST_RECOVER;
5990 /* Mark all of the pending packets in the queue as being lost */
5991 for (opr_queue_Scan(&call->tq, cursor)) {
5992 struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
5993 if (!(p->flags & RX_PKTFLAG_ACKED))
5994 p->flags &= ~RX_PKTFLAG_SENT;
5997 /* We're resending, so we double the timeout of the call. This will be
5998 * dropped back down by the first successful ACK that we receive.
6000 * We apply a maximum value here of 60 seconds
6002 clock_Add(&call->rto, &call->rto);
6003 if (clock_Gt(&call->rto, &maxTimeout))
6004 call->rto = maxTimeout;
6006 /* Packet loss is most likely due to congestion, so drop our window size
6007 * and start again from the beginning */
6008 if (peer->maxDgramPackets >1) {
6009 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
6010 call->MTU = MIN(peer->natMTU, peer->maxMTU);
6012 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
6013 call->nDgramPackets = 1;
6015 call->nextCwind = 1;
6018 MUTEX_ENTER(&peer->peer_lock);
6019 peer->MTU = call->MTU;
6020 peer->cwind = call->cwind;
6021 peer->nDgramPackets = 1;
6023 call->congestSeq = peer->congestSeq;
6024 MUTEX_EXIT(&peer->peer_lock);
6026 rxi_Start(call, istack);
6029 MUTEX_EXIT(&call->lock);
6032 /* This routine is called when new packets are readied for
6033 * transmission and when retransmission may be necessary, or when the
6034 * transmission window or burst count are favourable. This should be
6035 * better optimized for new packets, the usual case, now that we've
6036 * got rid of queues of send packets. XXXXXXXXXXX */
6038 rxi_Start(struct rx_call *call, int istack)
6040 struct opr_queue *cursor;
6041 #ifdef RX_ENABLE_LOCKS
6042 struct opr_queue *store;
6048 #ifdef RX_ENABLE_LOCKS
6049 if (rx_stats_active)
6050 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
6055 if (!opr_queue_IsEmpty(&call->tq)) { /* If we have anything to send */
6056 /* Send (or resend) any packets that need it, subject to
6057 * window restrictions and congestion burst control
6058 * restrictions. Ask for an ack on the last packet sent in
6059 * this burst. For now, we're relying upon the window being
6060 * considerably bigger than the largest number of packets that
6061 * are typically sent at once by one initial call to
6062 * rxi_Start. This is probably bogus (perhaps we should ask
6063 * for an ack when we're half way through the current
6064 * window?). Also, for non file transfer applications, this
6065 * may end up asking for an ack for every packet. Bogus. XXXX
6068 * But check whether we're here recursively, and let the other guy
6071 #ifdef RX_ENABLE_LOCKS
6072 if (!(call->flags & RX_CALL_TQ_BUSY)) {
6073 call->flags |= RX_CALL_TQ_BUSY;
6075 #endif /* RX_ENABLE_LOCKS */
6077 #ifdef RX_ENABLE_LOCKS
6078 call->flags &= ~RX_CALL_NEED_START;
6079 #endif /* RX_ENABLE_LOCKS */
6081 maxXmitPackets = MIN(call->twind, call->cwind);
6082 for (opr_queue_Scan(&call->tq, cursor)) {
6084 = opr_queue_Entry(cursor, struct rx_packet, entry);
6086 if (p->flags & RX_PKTFLAG_ACKED) {
6087 /* Since we may block, don't trust this */
6088 if (rx_stats_active)
6089 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6090 continue; /* Ignore this packet if it has been acknowledged */
6093 /* Turn off all flags except these ones, which are the same
6094 * on each transmission */
6095 p->header.flags &= RX_PRESET_FLAGS;
6097 if (p->header.seq >=
6098 call->tfirst + MIN((int)call->twind,
6099 (int)(call->nSoftAcked +
6101 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6102 /* Note: if we're waiting for more window space, we can
6103 * still send retransmits; hence we don't return here, but
6104 * break out to schedule a retransmit event */
6105 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6106 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6111 /* Transmit the packet if it needs to be sent. */
6112 if (!(p->flags & RX_PKTFLAG_SENT)) {
6113 if (nXmitPackets == maxXmitPackets) {
6114 rxi_SendXmitList(call, call->xmitList,
6115 nXmitPackets, istack);
6118 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6119 *(call->callNumber), p));
6120 call->xmitList[nXmitPackets++] = p;
6122 } /* end of the queue_Scan */
6124 /* xmitList now hold pointers to all of the packets that are
6125 * ready to send. Now we loop to send the packets */
6126 if (nXmitPackets > 0) {
6127 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6131 #ifdef RX_ENABLE_LOCKS
6133 /* We went into the error state while sending packets. Now is
6134 * the time to reset the call. This will also inform the using
6135 * process that the call is in an error state.
6137 if (rx_stats_active)
6138 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6139 call->flags &= ~RX_CALL_TQ_BUSY;
6140 rxi_WakeUpTransmitQueue(call);
6141 rxi_CallError(call, call->error);
6145 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6147 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6148 /* Some packets have received acks. If they all have, we can clear
6149 * the transmit queue.
6152 for (opr_queue_ScanSafe(&call->tq, cursor, store)) {
6154 = opr_queue_Entry(cursor, struct rx_packet, entry);
6156 if (p->header.seq < call->tfirst
6157 && (p->flags & RX_PKTFLAG_ACKED)) {
6158 opr_queue_Remove(&p->entry);
6159 #ifdef RX_TRACK_PACKETS
6160 p->flags &= ~RX_PKTFLAG_TQ;
6162 #ifdef RXDEBUG_PACKET
6170 call->flags |= RX_CALL_TQ_CLEARME;
6172 if (call->flags & RX_CALL_TQ_CLEARME)
6173 rxi_ClearTransmitQueue(call, 1);
6174 } while (call->flags & RX_CALL_NEED_START);
6176 * TQ references no longer protected by this flag; they must remain
6177 * protected by the call lock.
6179 call->flags &= ~RX_CALL_TQ_BUSY;
6180 rxi_WakeUpTransmitQueue(call);
6182 call->flags |= RX_CALL_NEED_START;
6184 #endif /* RX_ENABLE_LOCKS */
6186 rxi_rto_cancel(call);
6190 /* Also adjusts the keep alive parameters for the call, to reflect
6191 * that we have just sent a packet (so keep alives aren't sent
6194 rxi_Send(struct rx_call *call, struct rx_packet *p,
6197 struct rx_connection *conn = call->conn;
6199 /* Stamp each packet with the user supplied status */
6200 p->header.userStatus = call->localStatus;
6202 /* Allow the security object controlling this call's security to
6203 * make any last-minute changes to the packet */
6204 RXS_SendPacket(conn->securityObject, call, p);
6206 /* Since we're about to send SOME sort of packet to the peer, it's
6207 * safe to nuke any scheduled end-of-packets ack */
6208 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6210 /* Actually send the packet, filling in more connection-specific fields */
6211 MUTEX_EXIT(&call->lock);
6212 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6213 rxi_SendPacket(call, conn, p, istack);
6214 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6215 MUTEX_ENTER(&call->lock);
6217 /* Update last send time for this call (for keep-alive
6218 * processing), and for the connection (so that we can discover
6219 * idle connections) */
6220 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6221 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6222 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6224 conn->lastSendTime = call->lastSendTime = clock_Sec();
6225 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6226 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6227 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6228 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6229 RX_ACK_PING_RESPONSE)))
6230 call->lastSendData = call->lastSendTime;
6234 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6235 * that things are fine. Also called periodically to guarantee that nothing
6236 * falls through the cracks (e.g. (error + dally) connections have keepalive
6237 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6239 * haveCTLock Set if calling from rxi_ReapConnections
6242 rxi_CheckCall(struct rx_call *call, int haveCTLock)
6244 struct rx_connection *conn = call->conn;
6246 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6247 afs_uint32 fudgeFactor;
6250 int idle_timeout = 0;
6251 afs_int32 clock_diff = 0;
6253 if (rxi_CheckPeerDead(call)) {
6259 /* Large swings in the clock can have a significant impact on
6260 * the performance of RX call processing. Forward clock shifts
6261 * will result in premature event triggering or timeouts.
6262 * Backward shifts can result in calls not completing until
6263 * the clock catches up with the original start clock value.
6265 * If a backward clock shift of more than five minutes is noticed,
6266 * just fail the call.
6268 if (now < call->lastSendTime)
6269 clock_diff = call->lastSendTime - now;
6270 if (now < call->startWait)
6271 clock_diff = MAX(clock_diff, call->startWait - now);
6272 if (now < call->lastReceiveTime)
6273 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6274 if (clock_diff > 5 * 60)
6276 if (call->state == RX_STATE_ACTIVE)
6277 rxi_CallError(call, RX_CALL_TIMEOUT);
6281 #ifdef RX_ENABLE_LOCKS
6282 if (call->flags & RX_CALL_TQ_BUSY) {
6283 /* Call is active and will be reset by rxi_Start if it's
6284 * in an error state.
6289 /* RTT + 8*MDEV, rounded up to the next second. */
6290 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6291 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6293 deadTime = conn->secondsUntilDead + fudgeFactor;
6294 /* These are computed to the second (+- 1 second). But that's
6295 * good enough for these values, which should be a significant
6296 * number of seconds. */
6297 if (now > (call->lastReceiveTime + deadTime)) {
6298 if (call->state == RX_STATE_ACTIVE) {
6299 #ifdef AFS_ADAPT_PMTU
6300 # if defined(KERNEL) && defined(AFS_SUN5_ENV)
6302 # if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6303 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6304 ip_stack_t *ipst = ns->netstack_ip;
6306 ire = ire_cache_lookup(conn->peer->host
6307 # if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6309 # if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
6311 # if defined(GLOBAL_NETSTACKID)
6318 if (ire && ire->ire_max_frag > 0)
6319 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6321 # if defined(GLOBAL_NETSTACKID)
6325 #endif /* AFS_ADAPT_PMTU */
6326 cerror = RX_CALL_DEAD;
6329 #ifdef RX_ENABLE_LOCKS
6330 /* Cancel pending events */
6331 rxevent_Cancel(&call->delayedAckEvent, call,
6332 RX_CALL_REFCOUNT_DELAY);
6333 rxi_rto_cancel(call);
6334 rxevent_Cancel(&call->keepAliveEvent, call,
6335 RX_CALL_REFCOUNT_ALIVE);
6336 rxevent_Cancel(&call->growMTUEvent, call,
6337 RX_CALL_REFCOUNT_MTU);
6338 MUTEX_ENTER(&rx_refcnt_mutex);
6339 /* if rxi_FreeCall returns 1 it has freed the call */
6340 if (call->refCount == 0 &&
6341 rxi_FreeCall(call, haveCTLock))
6343 MUTEX_EXIT(&rx_refcnt_mutex);
6346 MUTEX_EXIT(&rx_refcnt_mutex);
6348 #else /* RX_ENABLE_LOCKS */
6349 rxi_FreeCall(call, 0);
6351 #endif /* RX_ENABLE_LOCKS */
6353 /* Non-active calls are destroyed if they are not responding
6354 * to pings; active calls are simply flagged in error, so the
6355 * attached process can die reasonably gracefully. */
6358 if (conn->idleDeadDetection) {
6359 if (conn->idleDeadTime) {
6360 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6364 /* see if we have a non-activity timeout */
6365 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6366 (call->flags & RX_CALL_READER_WAIT)) {
6367 if (call->state == RX_STATE_ACTIVE) {
6368 cerror = RX_CALL_TIMEOUT;
6373 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6374 if (call->state == RX_STATE_ACTIVE) {
6375 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6383 if (conn->hardDeadTime) {
6384 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6387 /* see if we have a hard timeout */
6389 && (now > (hardDeadTime + call->startTime.sec))) {
6390 if (call->state == RX_STATE_ACTIVE)
6391 rxi_CallError(call, RX_CALL_TIMEOUT);
6396 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6397 call->lastReceiveTime) {
6398 int oldMTU = conn->peer->ifMTU;
6400 /* if we thought we could send more, perhaps things got worse */
6401 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6402 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6403 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6404 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6406 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6408 /* minimum capped in SetPeerMtu */
6409 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6412 conn->lastPacketSize = 0;
6414 /* needed so ResetCall doesn't clobber us. */
6415 call->MTU = conn->peer->ifMTU;
6417 /* if we never succeeded, let the error pass out as-is */
6418 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6419 cerror = conn->msgsizeRetryErr;
6422 rxi_CallError(call, cerror);
6427 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6428 void *dummy, int dummy2)
6430 struct rx_connection *conn = arg1;
6431 struct rx_header theader;
6432 char tbuffer[1 + sizeof(struct rx_header)];
6433 struct sockaddr_in taddr;
6436 struct iovec tmpiov[2];
6439 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6442 tp = &tbuffer[sizeof(struct rx_header)];
6443 taddr.sin_family = AF_INET;
6444 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6445 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6446 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6447 taddr.sin_len = sizeof(struct sockaddr_in);
6449 memset(&theader, 0, sizeof(theader));
6450 theader.epoch = htonl(999);
6452 theader.callNumber = 0;
6455 theader.type = RX_PACKET_TYPE_VERSION;
6456 theader.flags = RX_LAST_PACKET;
6457 theader.serviceId = 0;
6459 memcpy(tbuffer, &theader, sizeof(theader));
6460 memcpy(tp, &a, sizeof(a));
6461 tmpiov[0].iov_base = tbuffer;
6462 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6464 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6466 MUTEX_ENTER(&conn->conn_data_lock);
6467 MUTEX_ENTER(&rx_refcnt_mutex);
6468 /* Only reschedule ourselves if the connection would not be destroyed */
6469 if (conn->refCount <= 1) {
6470 rxevent_Put(conn->natKeepAliveEvent);
6471 conn->natKeepAliveEvent = NULL;
6472 MUTEX_EXIT(&rx_refcnt_mutex);
6473 MUTEX_EXIT(&conn->conn_data_lock);
6474 rx_DestroyConnection(conn); /* drop the reference for this */
6476 conn->refCount--; /* drop the reference for this */
6477 MUTEX_EXIT(&rx_refcnt_mutex);
6478 rxevent_Put(conn->natKeepAliveEvent);
6479 conn->natKeepAliveEvent = NULL;
6480 rxi_ScheduleNatKeepAliveEvent(conn);
6481 MUTEX_EXIT(&conn->conn_data_lock);
6486 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6488 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6489 struct clock when, now;
6490 clock_GetTime(&now);
6492 when.sec += conn->secondsUntilNatPing;
6493 MUTEX_ENTER(&rx_refcnt_mutex);
6494 conn->refCount++; /* hold a reference for this */
6495 MUTEX_EXIT(&rx_refcnt_mutex);
6496 conn->natKeepAliveEvent =
6497 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6502 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6504 MUTEX_ENTER(&conn->conn_data_lock);
6505 conn->secondsUntilNatPing = seconds;
6507 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6508 rxi_ScheduleNatKeepAliveEvent(conn);
6510 conn->flags |= RX_CONN_NAT_PING;
6512 MUTEX_EXIT(&conn->conn_data_lock);
6515 /* When a call is in progress, this routine is called occasionally to
6516 * make sure that some traffic has arrived (or been sent to) the peer.
6517 * If nothing has arrived in a reasonable amount of time, the call is
6518 * declared dead; if nothing has been sent for a while, we send a
6519 * keep-alive packet (if we're actually trying to keep the call alive)
6522 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6525 struct rx_call *call = arg1;
6526 struct rx_connection *conn;
6529 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6530 MUTEX_ENTER(&call->lock);
6532 if (event == call->keepAliveEvent) {
6533 rxevent_Put(call->keepAliveEvent);
6534 call->keepAliveEvent = NULL;
6539 if (rxi_CheckCall(call, 0)) {
6540 MUTEX_EXIT(&call->lock);
6544 /* Don't try to keep alive dallying calls */
6545 if (call->state == RX_STATE_DALLY) {
6546 MUTEX_EXIT(&call->lock);
6551 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6552 /* Don't try to send keepalives if there is unacknowledged data */
6553 /* the rexmit code should be good enough, this little hack
6554 * doesn't quite work XXX */
6555 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6557 rxi_ScheduleKeepAliveEvent(call);
6558 MUTEX_EXIT(&call->lock);
6561 /* Does what's on the nameplate. */
6563 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6565 struct rx_call *call = arg1;
6566 struct rx_connection *conn;
6568 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6569 MUTEX_ENTER(&call->lock);
6571 if (event == call->growMTUEvent) {
6572 rxevent_Put(call->growMTUEvent);
6573 call->growMTUEvent = NULL;
6576 if (rxi_CheckCall(call, 0)) {
6577 MUTEX_EXIT(&call->lock);
6581 /* Don't bother with dallying calls */
6582 if (call->state == RX_STATE_DALLY) {
6583 MUTEX_EXIT(&call->lock);
6590 * keep being scheduled, just don't do anything if we're at peak,
6591 * or we're not set up to be properly handled (idle timeout required)
6593 if ((conn->peer->maxPacketSize != 0) &&
6594 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6595 conn->idleDeadDetection)
6596 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6597 rxi_ScheduleGrowMTUEvent(call, 0);
6598 MUTEX_EXIT(&call->lock);
6602 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6604 if (!call->keepAliveEvent) {
6605 struct clock when, now;
6606 clock_GetTime(&now);
6608 when.sec += call->conn->secondsUntilPing;
6609 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6610 call->keepAliveEvent =
6611 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6616 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6618 if (!call->growMTUEvent) {
6619 struct clock when, now;
6621 clock_GetTime(&now);
6624 if (call->conn->secondsUntilPing)
6625 secs = (6*call->conn->secondsUntilPing)-1;
6627 if (call->conn->secondsUntilDead)
6628 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6632 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6633 call->growMTUEvent =
6634 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6638 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6640 rxi_KeepAliveOn(struct rx_call *call)
6642 /* Pretend last packet received was received now--i.e. if another
6643 * packet isn't received within the keep alive time, then the call
6644 * will die; Initialize last send time to the current time--even
6645 * if a packet hasn't been sent yet. This will guarantee that a
6646 * keep-alive is sent within the ping time */
6647 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6648 rxi_ScheduleKeepAliveEvent(call);
6652 * Solely in order that callers not need to include rx_call.h
6655 rx_KeepAliveOff(struct rx_call *call)
6657 rxi_KeepAliveOff(call);
6660 rx_KeepAliveOn(struct rx_call *call)
6662 rxi_KeepAliveOn(call);
6666 rxi_GrowMTUOn(struct rx_call *call)
6668 struct rx_connection *conn = call->conn;
6669 MUTEX_ENTER(&conn->conn_data_lock);
6670 conn->lastPingSizeSer = conn->lastPingSize = 0;
6671 MUTEX_EXIT(&conn->conn_data_lock);
6672 rxi_ScheduleGrowMTUEvent(call, 1);
6675 /* This routine is called to send connection abort messages
6676 * that have been delayed to throttle looping clients. */
6678 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6681 struct rx_connection *conn = arg1;
6684 struct rx_packet *packet;
6686 MUTEX_ENTER(&conn->conn_data_lock);
6687 rxevent_Put(conn->delayedAbortEvent);
6688 conn->delayedAbortEvent = NULL;
6689 error = htonl(conn->error);
6691 MUTEX_EXIT(&conn->conn_data_lock);
6692 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6695 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6696 RX_PACKET_TYPE_ABORT, (char *)&error,
6698 rxi_FreePacket(packet);
6702 /* This routine is called to send call abort messages
6703 * that have been delayed to throttle looping clients. */
6705 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6708 struct rx_call *call = arg1;
6711 struct rx_packet *packet;
6713 MUTEX_ENTER(&call->lock);
6714 rxevent_Put(call->delayedAbortEvent);
6715 call->delayedAbortEvent = NULL;
6716 error = htonl(call->error);
6718 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6721 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6722 (char *)&error, sizeof(error), 0);
6723 rxi_FreePacket(packet);
6725 MUTEX_EXIT(&call->lock);
6726 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6729 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6730 * seconds) to ask the client to authenticate itself. The routine
6731 * issues a challenge to the client, which is obtained from the
6732 * security object associated with the connection */
6734 rxi_ChallengeEvent(struct rxevent *event,
6735 void *arg0, void *arg1, int tries)
6737 struct rx_connection *conn = arg0;
6740 rxevent_Put(conn->challengeEvent);
6741 conn->challengeEvent = NULL;
6744 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6745 struct rx_packet *packet;
6746 struct clock when, now;
6749 /* We've failed to authenticate for too long.
6750 * Reset any calls waiting for authentication;
6751 * they are all in RX_STATE_PRECALL.
6755 MUTEX_ENTER(&conn->conn_call_lock);
6756 for (i = 0; i < RX_MAXCALLS; i++) {
6757 struct rx_call *call = conn->call[i];
6759 MUTEX_ENTER(&call->lock);
6760 if (call->state == RX_STATE_PRECALL) {
6761 rxi_CallError(call, RX_CALL_DEAD);
6762 rxi_SendCallAbort(call, NULL, 0, 0);
6764 MUTEX_EXIT(&call->lock);
6767 MUTEX_EXIT(&conn->conn_call_lock);
6771 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6773 /* If there's no packet available, do this later. */
6774 RXS_GetChallenge(conn->securityObject, conn, packet);
6775 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6776 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6777 rxi_FreePacket(packet);
6779 clock_GetTime(&now);
6781 when.sec += RX_CHALLENGE_TIMEOUT;
6782 conn->challengeEvent =
6783 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6788 /* Call this routine to start requesting the client to authenticate
6789 * itself. This will continue until authentication is established,
6790 * the call times out, or an invalid response is returned. The
6791 * security object associated with the connection is asked to create
6792 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6793 * defined earlier. */
6795 rxi_ChallengeOn(struct rx_connection *conn)
6797 if (!conn->challengeEvent) {
6798 RXS_CreateChallenge(conn->securityObject, conn);
6799 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6804 /* rxi_ComputeRoundTripTime is called with peer locked. */
6805 /* peer may be null */
6807 rxi_ComputeRoundTripTime(struct rx_packet *p,
6808 struct rx_ackPacket *ack,
6809 struct rx_call *call,
6810 struct rx_peer *peer,
6813 struct clock thisRtt, *sentp;
6817 /* If the ACK is delayed, then do nothing */
6818 if (ack->reason == RX_ACK_DELAY)
6821 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6822 * their RTT multiple times, so only include the RTT of the last packet
6824 if (p->flags & RX_JUMBO_PACKET)
6827 /* Use the serial number to determine which transmission the ACK is for,
6828 * and set the sent time to match this. If we have no serial number, then
6829 * only use the ACK for RTT calculations if the packet has not been
6833 serial = ntohl(ack->serial);
6835 if (serial == p->header.serial) {
6836 sentp = &p->timeSent;
6837 } else if (serial == p->firstSerial) {
6838 sentp = &p->firstSent;
6839 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6840 sentp = &p->firstSent;
6844 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6845 sentp = &p->firstSent;
6852 if (clock_Lt(&thisRtt, sentp))
6853 return; /* somebody set the clock back, don't count this time. */
6855 clock_Sub(&thisRtt, sentp);
6856 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6857 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6859 if (clock_IsZero(&thisRtt)) {
6861 * The actual round trip time is shorter than the
6862 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6863 * Since we can't tell which at the moment we will assume 1ms.
6865 thisRtt.usec = 1000;
6868 if (rx_stats_active) {
6869 MUTEX_ENTER(&rx_stats_mutex);
6870 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6871 rx_stats.minRtt = thisRtt;
6872 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6873 if (thisRtt.sec > 60) {
6874 MUTEX_EXIT(&rx_stats_mutex);
6875 return; /* somebody set the clock ahead */
6877 rx_stats.maxRtt = thisRtt;
6879 clock_Add(&rx_stats.totalRtt, &thisRtt);
6880 rx_atomic_inc(&rx_stats.nRttSamples);
6881 MUTEX_EXIT(&rx_stats_mutex);
6884 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6886 /* Apply VanJacobson round-trip estimations */
6891 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6892 * srtt is stored as fixed point with 3 bits after the binary
6893 * point (i.e., scaled by 8). The following magic is
6894 * equivalent to the smoothing algorithm in rfc793 with an
6895 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6896 * srtt'*8 = rtt + srtt*7
6897 * srtt'*8 = srtt*8 + rtt - srtt
6898 * srtt' = srtt + rtt/8 - srtt/8
6899 * srtt' = srtt + (rtt - srtt)/8
6902 delta = _8THMSEC(&thisRtt) - call->rtt;
6903 call->rtt += (delta >> 3);
6906 * We accumulate a smoothed rtt variance (actually, a smoothed
6907 * mean difference), then set the retransmit timer to smoothed
6908 * rtt + 4 times the smoothed variance (was 2x in van's original
6909 * paper, but 4x works better for me, and apparently for him as
6911 * rttvar is stored as
6912 * fixed point with 2 bits after the binary point (scaled by
6913 * 4). The following is equivalent to rfc793 smoothing with
6914 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6915 * rttvar'*4 = rttvar*3 + |delta|
6916 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6917 * rttvar' = rttvar + |delta|/4 - rttvar/4
6918 * rttvar' = rttvar + (|delta| - rttvar)/4
6919 * This replaces rfc793's wired-in beta.
6920 * dev*4 = dev*4 + (|actual - expected| - dev)
6926 delta -= (call->rtt_dev << 1);
6927 call->rtt_dev += (delta >> 3);
6929 /* I don't have a stored RTT so I start with this value. Since I'm
6930 * probably just starting a call, and will be pushing more data down
6931 * this, I expect congestion to increase rapidly. So I fudge a
6932 * little, and I set deviance to half the rtt. In practice,
6933 * deviance tends to approach something a little less than
6934 * half the smoothed rtt. */
6935 call->rtt = _8THMSEC(&thisRtt) + 8;
6936 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6938 /* the smoothed RTT time is RTT + 4*MDEV
6940 * We allow a user specified minimum to be set for this, to allow clamping
6941 * at a minimum value in the same way as TCP. In addition, we have to allow
6942 * for the possibility that this packet is answered by a delayed ACK, so we
6943 * add on a fixed 200ms to account for that timer expiring.
6946 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6947 rx_minPeerTimeout) + 200;
6948 clock_Zero(&call->rto);
6949 clock_Addmsec(&call->rto, rtt_timeout);
6951 /* Update the peer, so any new calls start with our values */
6952 peer->rtt_dev = call->rtt_dev;
6953 peer->rtt = call->rtt;
6955 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6956 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6960 /* Find all server connections that have not been active for a long time, and
6963 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6966 struct clock now, when;
6967 clock_GetTime(&now);
6969 /* Find server connection structures that haven't been used for
6970 * greater than rx_idleConnectionTime */
6972 struct rx_connection **conn_ptr, **conn_end;
6973 int i, havecalls = 0;
6974 MUTEX_ENTER(&rx_connHashTable_lock);
6975 for (conn_ptr = &rx_connHashTable[0], conn_end =
6976 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6978 struct rx_connection *conn, *next;
6979 struct rx_call *call;
6983 for (conn = *conn_ptr; conn; conn = next) {
6984 /* XXX -- Shouldn't the connection be locked? */
6987 for (i = 0; i < RX_MAXCALLS; i++) {
6988 call = conn->call[i];
6992 code = MUTEX_TRYENTER(&call->lock);
6995 result = rxi_CheckCall(call, 1);
6996 MUTEX_EXIT(&call->lock);
6998 /* If CheckCall freed the call, it might
6999 * have destroyed the connection as well,
7000 * which screws up the linked lists.
7006 if (conn->type == RX_SERVER_CONNECTION) {
7007 /* This only actually destroys the connection if
7008 * there are no outstanding calls */
7009 MUTEX_ENTER(&conn->conn_data_lock);
7010 MUTEX_ENTER(&rx_refcnt_mutex);
7011 if (!havecalls && !conn->refCount
7012 && ((conn->lastSendTime + rx_idleConnectionTime) <
7014 conn->refCount++; /* it will be decr in rx_DestroyConn */
7015 MUTEX_EXIT(&rx_refcnt_mutex);
7016 MUTEX_EXIT(&conn->conn_data_lock);
7017 #ifdef RX_ENABLE_LOCKS
7018 rxi_DestroyConnectionNoLock(conn);
7019 #else /* RX_ENABLE_LOCKS */
7020 rxi_DestroyConnection(conn);
7021 #endif /* RX_ENABLE_LOCKS */
7023 #ifdef RX_ENABLE_LOCKS
7025 MUTEX_EXIT(&rx_refcnt_mutex);
7026 MUTEX_EXIT(&conn->conn_data_lock);
7028 #endif /* RX_ENABLE_LOCKS */
7032 #ifdef RX_ENABLE_LOCKS
7033 while (rx_connCleanup_list) {
7034 struct rx_connection *conn;
7035 conn = rx_connCleanup_list;
7036 rx_connCleanup_list = rx_connCleanup_list->next;
7037 MUTEX_EXIT(&rx_connHashTable_lock);
7038 rxi_CleanupConnection(conn);
7039 MUTEX_ENTER(&rx_connHashTable_lock);
7041 MUTEX_EXIT(&rx_connHashTable_lock);
7042 #endif /* RX_ENABLE_LOCKS */
7045 /* Find any peer structures that haven't been used (haven't had an
7046 * associated connection) for greater than rx_idlePeerTime */
7048 struct rx_peer **peer_ptr, **peer_end;
7052 * Why do we need to hold the rx_peerHashTable_lock across
7053 * the incrementing of peer_ptr since the rx_peerHashTable
7054 * array is not changing? We don't.
7056 * By dropping the lock periodically we can permit other
7057 * activities to be performed while a rxi_ReapConnections
7058 * call is in progress. The goal of reap connections
7059 * is to clean up quickly without causing large amounts
7060 * of contention. Therefore, it is important that global
7061 * mutexes not be held for extended periods of time.
7063 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7064 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7066 struct rx_peer *peer, *next, *prev;
7068 MUTEX_ENTER(&rx_peerHashTable_lock);
7069 for (prev = peer = *peer_ptr; peer; peer = next) {
7071 code = MUTEX_TRYENTER(&peer->peer_lock);
7072 if ((code) && (peer->refCount == 0)
7073 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7074 struct opr_queue *cursor, *store;
7078 * now know that this peer object is one to be
7079 * removed from the hash table. Once it is removed
7080 * it can't be referenced by other threads.
7081 * Lets remove it first and decrement the struct
7082 * nPeerStructs count.
7084 if (peer == *peer_ptr) {
7090 if (rx_stats_active)
7091 rx_atomic_dec(&rx_stats.nPeerStructs);
7094 * Now if we hold references on 'prev' and 'next'
7095 * we can safely drop the rx_peerHashTable_lock
7096 * while we destroy this 'peer' object.
7102 MUTEX_EXIT(&rx_peerHashTable_lock);
7104 MUTEX_EXIT(&peer->peer_lock);
7105 MUTEX_DESTROY(&peer->peer_lock);
7107 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7108 unsigned int num_funcs;
7109 struct rx_interface_stat *rpc_stat
7110 = opr_queue_Entry(cursor, struct rx_interface_stat,
7115 opr_queue_Remove(&rpc_stat->entry);
7116 opr_queue_Remove(&rpc_stat->entryPeers);
7118 num_funcs = rpc_stat->stats[0].func_total;
7120 sizeof(rx_interface_stat_t) +
7121 rpc_stat->stats[0].func_total *
7122 sizeof(rx_function_entry_v1_t);
7124 rxi_Free(rpc_stat, space);
7126 MUTEX_ENTER(&rx_rpc_stats);
7127 rxi_rpc_peer_stat_cnt -= num_funcs;
7128 MUTEX_EXIT(&rx_rpc_stats);
7133 * Regain the rx_peerHashTable_lock and
7134 * decrement the reference count on 'prev'
7137 MUTEX_ENTER(&rx_peerHashTable_lock);
7144 MUTEX_EXIT(&peer->peer_lock);
7149 MUTEX_EXIT(&rx_peerHashTable_lock);
7153 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7154 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7155 * GC, just below. Really, we shouldn't have to keep moving packets from
7156 * one place to another, but instead ought to always know if we can
7157 * afford to hold onto a packet in its particular use. */
7158 MUTEX_ENTER(&rx_freePktQ_lock);
7159 if (rx_waitingForPackets) {
7160 rx_waitingForPackets = 0;
7161 #ifdef RX_ENABLE_LOCKS
7162 CV_BROADCAST(&rx_waitingForPackets_cv);
7164 osi_rxWakeup(&rx_waitingForPackets);
7167 MUTEX_EXIT(&rx_freePktQ_lock);
7170 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7171 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7175 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7176 * rx.h is sort of strange this is better. This is called with a security
7177 * object before it is discarded. Each connection using a security object has
7178 * its own refcount to the object so it won't actually be freed until the last
7179 * connection is destroyed.
7181 * This is the only rxs module call. A hold could also be written but no one
7185 rxs_Release(struct rx_securityClass *aobj)
7187 return RXS_Close(aobj);
7195 #define TRACE_OPTION_RX_DEBUG 16
7203 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7204 0, KEY_QUERY_VALUE, &parmKey);
7205 if (code != ERROR_SUCCESS)
7208 dummyLen = sizeof(TraceOption);
7209 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7210 (BYTE *) &TraceOption, &dummyLen);
7211 if (code == ERROR_SUCCESS) {
7212 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7214 RegCloseKey (parmKey);
7215 #endif /* AFS_NT40_ENV */
7220 rx_DebugOnOff(int on)
7224 rxdebug_active = on;
7230 rx_StatsOnOff(int on)
7232 rx_stats_active = on;
7236 /* Don't call this debugging routine directly; use dpf */
7238 rxi_DebugPrint(char *format, ...)
7247 va_start(ap, format);
7249 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7252 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7254 OutputDebugString(msg);
7260 va_start(ap, format);
7262 clock_GetTime(&now);
7263 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7264 (unsigned int)now.usec);
7265 vfprintf(rx_Log, format, ap);
7273 * This function is used to process the rx_stats structure that is local
7274 * to a process as well as an rx_stats structure received from a remote
7275 * process (via rxdebug). Therefore, it needs to do minimal version
7279 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7280 afs_int32 freePackets, char version)
7284 if (size != sizeof(struct rx_statistics)) {
7286 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7287 size, sizeof(struct rx_statistics));
7290 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7293 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7294 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7295 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7296 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7297 s->specialPktAllocFailures);
7299 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7300 s->receivePktAllocFailures, s->sendPktAllocFailures,
7301 s->specialPktAllocFailures);
7305 " greedy %u, " "bogusReads %u (last from host %x), "
7306 "noPackets %u, " "noBuffers %u, " "selects %u, "
7307 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7308 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7309 s->selects, s->sendSelects);
7311 fprintf(file, " packets read: ");
7312 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7313 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7315 fprintf(file, "\n");
7318 " other read counters: data %u, " "ack %u, " "dup %u "
7319 "spurious %u " "dally %u\n", s->dataPacketsRead,
7320 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7321 s->ignorePacketDally);
7323 fprintf(file, " packets sent: ");
7324 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7325 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7327 fprintf(file, "\n");
7330 " other send counters: ack %u, " "data %u (not resends), "
7331 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7332 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7333 s->dataPacketsPushed, s->ignoreAckedPacket);
7336 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7337 s->netSendFailures, (int)s->fatalErrors);
7339 if (s->nRttSamples) {
7340 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7341 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7343 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7344 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7348 " %d server connections, " "%d client connections, "
7349 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7350 s->nServerConns, s->nClientConns, s->nPeerStructs,
7351 s->nCallStructs, s->nFreeCallStructs);
7353 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7354 fprintf(file, " %d clock updates\n", clock_nUpdates);
7358 /* for backward compatibility */
7360 rx_PrintStats(FILE * file)
7362 MUTEX_ENTER(&rx_stats_mutex);
7363 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7364 sizeof(rx_stats), rx_nFreePackets,
7366 MUTEX_EXIT(&rx_stats_mutex);
7370 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7372 fprintf(file, "Peer %x.%d.\n",
7373 ntohl(peer->host), (int)ntohs(peer->port));
7376 " Rtt %d, " "total sent %d, " "resent %d\n",
7377 peer->rtt, peer->nSent, peer->reSends);
7379 fprintf(file, " Packet size %d\n", peer->ifMTU);
7383 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7385 * This mutex protects the following static variables:
7389 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7390 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7392 #define LOCK_RX_DEBUG
7393 #define UNLOCK_RX_DEBUG
7394 #endif /* AFS_PTHREAD_ENV */
7396 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7398 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7399 u_char type, void *inputData, size_t inputLength,
7400 void *outputData, size_t outputLength)
7402 static afs_int32 counter = 100;
7403 time_t waitTime, waitCount;
7404 struct rx_header theader;
7407 struct timeval tv_now, tv_wake, tv_delta;
7408 struct sockaddr_in taddr, faddr;
7422 tp = &tbuffer[sizeof(struct rx_header)];
7423 taddr.sin_family = AF_INET;
7424 taddr.sin_port = remotePort;
7425 taddr.sin_addr.s_addr = remoteAddr;
7426 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7427 taddr.sin_len = sizeof(struct sockaddr_in);
7430 memset(&theader, 0, sizeof(theader));
7431 theader.epoch = htonl(999);
7433 theader.callNumber = htonl(counter);
7436 theader.type = type;
7437 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7438 theader.serviceId = 0;
7440 memcpy(tbuffer, &theader, sizeof(theader));
7441 memcpy(tp, inputData, inputLength);
7443 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7444 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7446 /* see if there's a packet available */
7447 gettimeofday(&tv_wake, NULL);
7448 tv_wake.tv_sec += waitTime;
7451 FD_SET(socket, &imask);
7452 tv_delta.tv_sec = tv_wake.tv_sec;
7453 tv_delta.tv_usec = tv_wake.tv_usec;
7454 gettimeofday(&tv_now, NULL);
7456 if (tv_delta.tv_usec < tv_now.tv_usec) {
7458 tv_delta.tv_usec += 1000000;
7461 tv_delta.tv_usec -= tv_now.tv_usec;
7463 if (tv_delta.tv_sec < tv_now.tv_sec) {
7467 tv_delta.tv_sec -= tv_now.tv_sec;
7470 code = select(0, &imask, 0, 0, &tv_delta);
7471 #else /* AFS_NT40_ENV */
7472 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7473 #endif /* AFS_NT40_ENV */
7474 if (code == 1 && FD_ISSET(socket, &imask)) {
7475 /* now receive a packet */
7476 faddrLen = sizeof(struct sockaddr_in);
7478 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7479 (struct sockaddr *)&faddr, &faddrLen);
7482 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7483 if (counter == ntohl(theader.callNumber))
7491 /* see if we've timed out */
7499 code -= sizeof(struct rx_header);
7500 if (code > outputLength)
7501 code = outputLength;
7502 memcpy(outputData, tp, code);
7505 #endif /* RXDEBUG */
7508 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7509 afs_uint16 remotePort, struct rx_debugStats * stat,
7510 afs_uint32 * supportedValues)
7512 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7514 struct rx_debugIn in;
7516 *supportedValues = 0;
7517 in.type = htonl(RX_DEBUGI_GETSTATS);
7520 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7521 &in, sizeof(in), stat, sizeof(*stat));
7524 * If the call was successful, fixup the version and indicate
7525 * what contents of the stat structure are valid.
7526 * Also do net to host conversion of fields here.
7530 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7531 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7533 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7534 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7536 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7537 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7539 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7540 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7542 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7543 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7545 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7546 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7548 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7549 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7551 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7552 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7554 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7555 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7557 stat->nFreePackets = ntohl(stat->nFreePackets);
7558 stat->packetReclaims = ntohl(stat->packetReclaims);
7559 stat->callsExecuted = ntohl(stat->callsExecuted);
7560 stat->nWaiting = ntohl(stat->nWaiting);
7561 stat->idleThreads = ntohl(stat->idleThreads);
7562 stat->nWaited = ntohl(stat->nWaited);
7563 stat->nPackets = ntohl(stat->nPackets);
7572 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7573 afs_uint16 remotePort, struct rx_statistics * stat,
7574 afs_uint32 * supportedValues)
7576 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7578 struct rx_debugIn in;
7579 afs_int32 *lp = (afs_int32 *) stat;
7583 * supportedValues is currently unused, but added to allow future
7584 * versioning of this function.
7587 *supportedValues = 0;
7588 in.type = htonl(RX_DEBUGI_RXSTATS);
7590 memset(stat, 0, sizeof(*stat));
7592 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7593 &in, sizeof(in), stat, sizeof(*stat));
7598 * Do net to host conversion here
7601 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7612 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7613 afs_uint16 remotePort, size_t version_length,
7616 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7618 return MakeDebugCall(socket, remoteAddr, remotePort,
7619 RX_PACKET_TYPE_VERSION, a, 1, version,
7627 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7628 afs_uint16 remotePort, afs_int32 * nextConnection,
7629 int allConnections, afs_uint32 debugSupportedValues,
7630 struct rx_debugConn * conn,
7631 afs_uint32 * supportedValues)
7633 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7635 struct rx_debugIn in;
7639 * supportedValues is currently unused, but added to allow future
7640 * versioning of this function.
7643 *supportedValues = 0;
7644 if (allConnections) {
7645 in.type = htonl(RX_DEBUGI_GETALLCONN);
7647 in.type = htonl(RX_DEBUGI_GETCONN);
7649 in.index = htonl(*nextConnection);
7650 memset(conn, 0, sizeof(*conn));
7652 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7653 &in, sizeof(in), conn, sizeof(*conn));
7656 *nextConnection += 1;
7659 * Convert old connection format to new structure.
7662 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7663 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7664 #define MOVEvL(a) (conn->a = vL->a)
7666 /* any old or unrecognized version... */
7667 for (i = 0; i < RX_MAXCALLS; i++) {
7668 MOVEvL(callState[i]);
7669 MOVEvL(callMode[i]);
7670 MOVEvL(callFlags[i]);
7671 MOVEvL(callOther[i]);
7673 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7674 MOVEvL(secStats.type);
7675 MOVEvL(secStats.level);
7676 MOVEvL(secStats.flags);
7677 MOVEvL(secStats.expires);
7678 MOVEvL(secStats.packetsReceived);
7679 MOVEvL(secStats.packetsSent);
7680 MOVEvL(secStats.bytesReceived);
7681 MOVEvL(secStats.bytesSent);
7686 * Do net to host conversion here
7688 * I don't convert host or port since we are most likely
7689 * going to want these in NBO.
7691 conn->cid = ntohl(conn->cid);
7692 conn->serial = ntohl(conn->serial);
7693 for (i = 0; i < RX_MAXCALLS; i++) {
7694 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7696 conn->error = ntohl(conn->error);
7697 conn->secStats.flags = ntohl(conn->secStats.flags);
7698 conn->secStats.expires = ntohl(conn->secStats.expires);
7699 conn->secStats.packetsReceived =
7700 ntohl(conn->secStats.packetsReceived);
7701 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7702 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7703 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7704 conn->epoch = ntohl(conn->epoch);
7705 conn->natMTU = ntohl(conn->natMTU);
7714 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7715 afs_uint16 remotePort, afs_int32 * nextPeer,
7716 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7717 afs_uint32 * supportedValues)
7719 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7721 struct rx_debugIn in;
7724 * supportedValues is currently unused, but added to allow future
7725 * versioning of this function.
7728 *supportedValues = 0;
7729 in.type = htonl(RX_DEBUGI_GETPEER);
7730 in.index = htonl(*nextPeer);
7731 memset(peer, 0, sizeof(*peer));
7733 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7734 &in, sizeof(in), peer, sizeof(*peer));
7740 * Do net to host conversion here
7742 * I don't convert host or port since we are most likely
7743 * going to want these in NBO.
7745 peer->ifMTU = ntohs(peer->ifMTU);
7746 peer->idleWhen = ntohl(peer->idleWhen);
7747 peer->refCount = ntohs(peer->refCount);
7748 peer->rtt = ntohl(peer->rtt);
7749 peer->rtt_dev = ntohl(peer->rtt_dev);
7750 peer->timeout.sec = 0;
7751 peer->timeout.usec = 0;
7752 peer->nSent = ntohl(peer->nSent);
7753 peer->reSends = ntohl(peer->reSends);
7754 peer->natMTU = ntohs(peer->natMTU);
7755 peer->maxMTU = ntohs(peer->maxMTU);
7756 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7757 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7758 peer->MTU = ntohs(peer->MTU);
7759 peer->cwind = ntohs(peer->cwind);
7760 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7761 peer->congestSeq = ntohs(peer->congestSeq);
7762 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7763 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7764 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7765 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7774 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7775 struct rx_debugPeer * peerStats)
7778 afs_int32 error = 1; /* default to "did not succeed" */
7779 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7781 MUTEX_ENTER(&rx_peerHashTable_lock);
7782 for(tp = rx_peerHashTable[hashValue];
7783 tp != NULL; tp = tp->next) {
7784 if (tp->host == peerHost)
7790 MUTEX_EXIT(&rx_peerHashTable_lock);
7794 MUTEX_ENTER(&tp->peer_lock);
7795 peerStats->host = tp->host;
7796 peerStats->port = tp->port;
7797 peerStats->ifMTU = tp->ifMTU;
7798 peerStats->idleWhen = tp->idleWhen;
7799 peerStats->refCount = tp->refCount;
7800 peerStats->burstSize = 0;
7801 peerStats->burst = 0;
7802 peerStats->burstWait.sec = 0;
7803 peerStats->burstWait.usec = 0;
7804 peerStats->rtt = tp->rtt;
7805 peerStats->rtt_dev = tp->rtt_dev;
7806 peerStats->timeout.sec = 0;
7807 peerStats->timeout.usec = 0;
7808 peerStats->nSent = tp->nSent;
7809 peerStats->reSends = tp->reSends;
7810 peerStats->natMTU = tp->natMTU;
7811 peerStats->maxMTU = tp->maxMTU;
7812 peerStats->maxDgramPackets = tp->maxDgramPackets;
7813 peerStats->ifDgramPackets = tp->ifDgramPackets;
7814 peerStats->MTU = tp->MTU;
7815 peerStats->cwind = tp->cwind;
7816 peerStats->nDgramPackets = tp->nDgramPackets;
7817 peerStats->congestSeq = tp->congestSeq;
7818 peerStats->bytesSent.high = tp->bytesSent >> 32;
7819 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7820 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7821 peerStats->bytesReceived.low
7822 = tp->bytesReceived & MAX_AFS_UINT32;
7823 MUTEX_EXIT(&tp->peer_lock);
7825 MUTEX_ENTER(&rx_peerHashTable_lock);
7828 MUTEX_EXIT(&rx_peerHashTable_lock);
7836 struct rx_serverQueueEntry *np;
7839 struct rx_call *call;
7840 struct rx_serverQueueEntry *sq;
7844 if (rxinit_status == 1) {
7846 return; /* Already shutdown. */
7850 #ifndef AFS_PTHREAD_ENV
7851 FD_ZERO(&rx_selectMask);
7852 #endif /* AFS_PTHREAD_ENV */
7853 rxi_dataQuota = RX_MAX_QUOTA;
7854 #ifndef AFS_PTHREAD_ENV
7856 #endif /* AFS_PTHREAD_ENV */
7859 #ifndef AFS_PTHREAD_ENV
7860 #ifndef AFS_USE_GETTIMEOFDAY
7862 #endif /* AFS_USE_GETTIMEOFDAY */
7863 #endif /* AFS_PTHREAD_ENV */
7865 while (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
7866 call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
7867 opr_queue_Remove(&call->entry);
7868 rxi_Free(call, sizeof(struct rx_call));
7871 while (!opr_queue_IsEmpty(&rx_idleServerQueue)) {
7872 sq = opr_queue_First(&rx_idleServerQueue, struct rx_serverQueueEntry,
7874 opr_queue_Remove(&sq->entry);
7879 struct rx_peer **peer_ptr, **peer_end;
7880 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7881 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7883 struct rx_peer *peer, *next;
7885 MUTEX_ENTER(&rx_peerHashTable_lock);
7886 for (peer = *peer_ptr; peer; peer = next) {
7887 struct opr_queue *cursor, *store;
7890 MUTEX_ENTER(&rx_rpc_stats);
7891 MUTEX_ENTER(&peer->peer_lock);
7892 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
7893 unsigned int num_funcs;
7894 struct rx_interface_stat *rpc_stat
7895 = opr_queue_Entry(cursor, struct rx_interface_stat,
7899 opr_queue_Remove(&rpc_stat->entry);
7900 opr_queue_Remove(&rpc_stat->entryPeers);
7901 num_funcs = rpc_stat->stats[0].func_total;
7903 sizeof(rx_interface_stat_t) +
7904 rpc_stat->stats[0].func_total *
7905 sizeof(rx_function_entry_v1_t);
7907 rxi_Free(rpc_stat, space);
7909 /* rx_rpc_stats must be held */
7910 rxi_rpc_peer_stat_cnt -= num_funcs;
7912 MUTEX_EXIT(&peer->peer_lock);
7913 MUTEX_EXIT(&rx_rpc_stats);
7917 if (rx_stats_active)
7918 rx_atomic_dec(&rx_stats.nPeerStructs);
7920 MUTEX_EXIT(&rx_peerHashTable_lock);
7923 for (i = 0; i < RX_MAX_SERVICES; i++) {
7925 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7927 for (i = 0; i < rx_hashTableSize; i++) {
7928 struct rx_connection *tc, *ntc;
7929 MUTEX_ENTER(&rx_connHashTable_lock);
7930 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7932 for (j = 0; j < RX_MAXCALLS; j++) {
7934 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7937 rxi_Free(tc, sizeof(*tc));
7939 MUTEX_EXIT(&rx_connHashTable_lock);
7942 MUTEX_ENTER(&freeSQEList_lock);
7944 while ((np = rx_FreeSQEList)) {
7945 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7946 MUTEX_DESTROY(&np->lock);
7947 rxi_Free(np, sizeof(*np));
7950 MUTEX_EXIT(&freeSQEList_lock);
7951 MUTEX_DESTROY(&freeSQEList_lock);
7952 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7953 MUTEX_DESTROY(&rx_connHashTable_lock);
7954 MUTEX_DESTROY(&rx_peerHashTable_lock);
7955 MUTEX_DESTROY(&rx_serverPool_lock);
7957 osi_Free(rx_connHashTable,
7958 rx_hashTableSize * sizeof(struct rx_connection *));
7959 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7961 UNPIN(rx_connHashTable,
7962 rx_hashTableSize * sizeof(struct rx_connection *));
7963 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7965 rxi_FreeAllPackets();
7967 MUTEX_ENTER(&rx_quota_mutex);
7968 rxi_dataQuota = RX_MAX_QUOTA;
7969 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7970 MUTEX_EXIT(&rx_quota_mutex);
7975 #ifdef RX_ENABLE_LOCKS
7977 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7979 if (!MUTEX_ISMINE(lockaddr))
7980 osi_Panic("Lock not held: %s", msg);
7982 #endif /* RX_ENABLE_LOCKS */
7987 * Routines to implement connection specific data.
7991 rx_KeyCreate(rx_destructor_t rtn)
7994 MUTEX_ENTER(&rxi_keyCreate_lock);
7995 key = rxi_keyCreate_counter++;
7996 rxi_keyCreate_destructor = (rx_destructor_t *)
7997 realloc((void *)rxi_keyCreate_destructor,
7998 (key + 1) * sizeof(rx_destructor_t));
7999 rxi_keyCreate_destructor[key] = rtn;
8000 MUTEX_EXIT(&rxi_keyCreate_lock);
8005 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
8008 MUTEX_ENTER(&conn->conn_data_lock);
8009 if (!conn->specific) {
8010 conn->specific = malloc((key + 1) * sizeof(void *));
8011 for (i = 0; i < key; i++)
8012 conn->specific[i] = NULL;
8013 conn->nSpecific = key + 1;
8014 conn->specific[key] = ptr;
8015 } else if (key >= conn->nSpecific) {
8016 conn->specific = (void **)
8017 realloc(conn->specific, (key + 1) * sizeof(void *));
8018 for (i = conn->nSpecific; i < key; i++)
8019 conn->specific[i] = NULL;
8020 conn->nSpecific = key + 1;
8021 conn->specific[key] = ptr;
8023 if (conn->specific[key] && rxi_keyCreate_destructor[key])
8024 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
8025 conn->specific[key] = ptr;
8027 MUTEX_EXIT(&conn->conn_data_lock);
8031 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
8034 MUTEX_ENTER(&svc->svc_data_lock);
8035 if (!svc->specific) {
8036 svc->specific = malloc((key + 1) * sizeof(void *));
8037 for (i = 0; i < key; i++)
8038 svc->specific[i] = NULL;
8039 svc->nSpecific = key + 1;
8040 svc->specific[key] = ptr;
8041 } else if (key >= svc->nSpecific) {
8042 svc->specific = (void **)
8043 realloc(svc->specific, (key + 1) * sizeof(void *));
8044 for (i = svc->nSpecific; i < key; i++)
8045 svc->specific[i] = NULL;
8046 svc->nSpecific = key + 1;
8047 svc->specific[key] = ptr;
8049 if (svc->specific[key] && rxi_keyCreate_destructor[key])
8050 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
8051 svc->specific[key] = ptr;
8053 MUTEX_EXIT(&svc->svc_data_lock);
8057 rx_GetSpecific(struct rx_connection *conn, int key)
8060 MUTEX_ENTER(&conn->conn_data_lock);
8061 if (key >= conn->nSpecific)
8064 ptr = conn->specific[key];
8065 MUTEX_EXIT(&conn->conn_data_lock);
8070 rx_GetServiceSpecific(struct rx_service *svc, int key)
8073 MUTEX_ENTER(&svc->svc_data_lock);
8074 if (key >= svc->nSpecific)
8077 ptr = svc->specific[key];
8078 MUTEX_EXIT(&svc->svc_data_lock);
8083 #endif /* !KERNEL */
8086 * processStats is a queue used to store the statistics for the local
8087 * process. Its contents are similar to the contents of the rpcStats
8088 * queue on a rx_peer structure, but the actual data stored within
8089 * this queue contains totals across the lifetime of the process (assuming
8090 * the stats have not been reset) - unlike the per peer structures
8091 * which can come and go based upon the peer lifetime.
8094 static struct opr_queue processStats = { &processStats, &processStats };
8097 * peerStats is a queue used to store the statistics for all peer structs.
8098 * Its contents are the union of all the peer rpcStats queues.
8101 static struct opr_queue peerStats = { &peerStats, &peerStats };
8104 * rxi_monitor_processStats is used to turn process wide stat collection
8108 static int rxi_monitor_processStats = 0;
8111 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8114 static int rxi_monitor_peerStats = 0;
8118 rxi_ClearRPCOpStat(rx_function_entry_v1_p rpc_stat)
8120 rpc_stat->invocations = 0;
8121 rpc_stat->bytes_sent = 0;
8122 rpc_stat->bytes_rcvd = 0;
8123 rpc_stat->queue_time_sum.sec = 0;
8124 rpc_stat->queue_time_sum.usec = 0;
8125 rpc_stat->queue_time_sum_sqr.sec = 0;
8126 rpc_stat->queue_time_sum_sqr.usec = 0;
8127 rpc_stat->queue_time_min.sec = 9999999;
8128 rpc_stat->queue_time_min.usec = 9999999;
8129 rpc_stat->queue_time_max.sec = 0;
8130 rpc_stat->queue_time_max.usec = 0;
8131 rpc_stat->execution_time_sum.sec = 0;
8132 rpc_stat->execution_time_sum.usec = 0;
8133 rpc_stat->execution_time_sum_sqr.sec = 0;
8134 rpc_stat->execution_time_sum_sqr.usec = 0;
8135 rpc_stat->execution_time_min.sec = 9999999;
8136 rpc_stat->execution_time_min.usec = 9999999;
8137 rpc_stat->execution_time_max.sec = 0;
8138 rpc_stat->execution_time_max.usec = 0;
8142 * Given all of the information for a particular rpc
8143 * call, find or create (if requested) the stat structure for the rpc.
8146 * the queue of stats that will be updated with the new value
8148 * @param rxInterface
8149 * a unique number that identifies the rpc interface
8152 * the total number of functions in this interface. this is only
8153 * required if create is true
8156 * if true, this invocation was made to a server
8159 * the ip address of the remote host. this is only required if create
8160 * and addToPeerList are true
8163 * the port of the remote host. this is only required if create
8164 * and addToPeerList are true
8166 * @param addToPeerList
8167 * if != 0, add newly created stat to the global peer list
8170 * if a new stats structure is allocated, the counter will
8171 * be updated with the new number of allocated stat structures.
8172 * only required if create is true
8175 * if no stats structure exists, allocate one
8179 static rx_interface_stat_p
8180 rxi_FindRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8181 afs_uint32 totalFunc, int isServer, afs_uint32 remoteHost,
8182 afs_uint32 remotePort, int addToPeerList,
8183 unsigned int *counter, int create)
8185 rx_interface_stat_p rpc_stat = NULL;
8186 struct opr_queue *cursor;
8189 * See if there's already a structure for this interface
8192 for (opr_queue_Scan(stats, cursor)) {
8193 rpc_stat = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8195 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8196 && (rpc_stat->stats[0].remote_is_server == isServer))
8200 /* if they didn't ask us to create, we're done */
8202 if (opr_queue_IsEnd(stats, cursor))
8208 /* can't proceed without these */
8209 if (!totalFunc || !counter)
8213 * Didn't find a match so allocate a new structure and add it to the
8217 if (opr_queue_IsEnd(stats, cursor) || (rpc_stat == NULL)
8218 || (rpc_stat->stats[0].interfaceId != rxInterface)
8219 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8224 sizeof(rx_interface_stat_t) +
8225 totalFunc * sizeof(rx_function_entry_v1_t);
8227 rpc_stat = rxi_Alloc(space);
8228 if (rpc_stat == NULL)
8231 *counter += totalFunc;
8232 for (i = 0; i < totalFunc; i++) {
8233 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8234 rpc_stat->stats[i].remote_peer = remoteHost;
8235 rpc_stat->stats[i].remote_port = remotePort;
8236 rpc_stat->stats[i].remote_is_server = isServer;
8237 rpc_stat->stats[i].interfaceId = rxInterface;
8238 rpc_stat->stats[i].func_total = totalFunc;
8239 rpc_stat->stats[i].func_index = i;
8241 opr_queue_Prepend(stats, &rpc_stat->entry);
8242 if (addToPeerList) {
8243 opr_queue_Prepend(&peerStats, &rpc_stat->entryPeers);
8250 rx_ClearProcessRPCStats(afs_int32 rxInterface)
8252 rx_interface_stat_p rpc_stat;
8255 if (rxInterface == -1)
8258 MUTEX_ENTER(&rx_rpc_stats);
8259 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8262 totalFunc = rpc_stat->stats[0].func_total;
8263 for (i = 0; i < totalFunc; i++)
8264 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8266 MUTEX_EXIT(&rx_rpc_stats);
8271 rx_ClearPeerRPCStats(afs_int32 rxInterface, afs_uint32 peerHost, afs_uint16 peerPort)
8273 rx_interface_stat_p rpc_stat;
8275 struct rx_peer * peer;
8277 if (rxInterface == -1)
8280 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8284 MUTEX_ENTER(&rx_rpc_stats);
8285 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8288 totalFunc = rpc_stat->stats[0].func_total;
8289 for (i = 0; i < totalFunc; i++)
8290 rxi_ClearRPCOpStat(&(rpc_stat->stats[i]));
8292 MUTEX_EXIT(&rx_rpc_stats);
8297 rx_CopyProcessRPCStats(afs_uint64 op)
8299 rx_interface_stat_p rpc_stat;
8300 rx_function_entry_v1_p rpcop_stat =
8301 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8302 int currentFunc = (op & MAX_AFS_UINT32);
8303 afs_int32 rxInterface = (op >> 32);
8305 if (!rxi_monitor_processStats)
8308 if (rxInterface == -1)
8311 if (rpcop_stat == NULL)
8314 MUTEX_ENTER(&rx_rpc_stats);
8315 rpc_stat = rxi_FindRpcStat(&processStats, rxInterface, 0, 0,
8318 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8319 sizeof(rx_function_entry_v1_t));
8320 MUTEX_EXIT(&rx_rpc_stats);
8322 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8329 rx_CopyPeerRPCStats(afs_uint64 op, afs_uint32 peerHost, afs_uint16 peerPort)
8331 rx_interface_stat_p rpc_stat;
8332 rx_function_entry_v1_p rpcop_stat =
8333 rxi_Alloc(sizeof(rx_function_entry_v1_t));
8334 int currentFunc = (op & MAX_AFS_UINT32);
8335 afs_int32 rxInterface = (op >> 32);
8336 struct rx_peer *peer;
8338 if (!rxi_monitor_peerStats)
8341 if (rxInterface == -1)
8344 if (rpcop_stat == NULL)
8347 peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
8351 MUTEX_ENTER(&rx_rpc_stats);
8352 rpc_stat = rxi_FindRpcStat(&peer->rpcStats, rxInterface, 0, 1,
8355 memcpy(rpcop_stat, &(rpc_stat->stats[currentFunc]),
8356 sizeof(rx_function_entry_v1_t));
8357 MUTEX_EXIT(&rx_rpc_stats);
8359 rxi_Free(rpcop_stat, sizeof(rx_function_entry_v1_t));
8366 rx_ReleaseRPCStats(void *stats)
8369 rxi_Free(stats, sizeof(rx_function_entry_v1_t));
8373 * Given all of the information for a particular rpc
8374 * call, create (if needed) and update the stat totals for the rpc.
8377 * the queue of stats that will be updated with the new value
8379 * @param rxInterface
8380 * a unique number that identifies the rpc interface
8382 * @param currentFunc
8383 * the index of the function being invoked
8386 * the total number of functions in this interface
8389 * the amount of time this function waited for a thread
8392 * the amount of time this function invocation took to execute
8395 * the number bytes sent by this invocation
8398 * the number bytes received by this invocation
8401 * if true, this invocation was made to a server
8404 * the ip address of the remote host
8407 * the port of the remote host
8409 * @param addToPeerList
8410 * if != 0, add newly created stat to the global peer list
8413 * if a new stats structure is allocated, the counter will
8414 * be updated with the new number of allocated stat structures
8419 rxi_AddRpcStat(struct opr_queue *stats, afs_uint32 rxInterface,
8420 afs_uint32 currentFunc, afs_uint32 totalFunc,
8421 struct clock *queueTime, struct clock *execTime,
8422 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8423 afs_uint32 remoteHost, afs_uint32 remotePort,
8424 int addToPeerList, unsigned int *counter)
8427 rx_interface_stat_p rpc_stat;
8429 rpc_stat = rxi_FindRpcStat(stats, rxInterface, totalFunc, isServer,
8430 remoteHost, remotePort, addToPeerList, counter,
8438 * Increment the stats for this function
8441 rpc_stat->stats[currentFunc].invocations++;
8442 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8443 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8444 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8445 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8446 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8447 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8449 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8450 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8452 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8453 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8455 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8456 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8458 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8459 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8467 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8468 afs_uint32 currentFunc, afs_uint32 totalFunc,
8469 struct clock *queueTime, struct clock *execTime,
8470 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8474 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8477 MUTEX_ENTER(&rx_rpc_stats);
8479 if (rxi_monitor_peerStats) {
8480 MUTEX_ENTER(&peer->peer_lock);
8481 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8482 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8483 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8484 MUTEX_EXIT(&peer->peer_lock);
8487 if (rxi_monitor_processStats) {
8488 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8489 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8490 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8493 MUTEX_EXIT(&rx_rpc_stats);
8497 * Increment the times and count for a particular rpc function.
8499 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8500 * call rx_RecordCallStatistics instead, so the public version of this
8501 * function is left purely for legacy callers.
8504 * The peer who invoked the rpc
8506 * @param rxInterface
8507 * A unique number that identifies the rpc interface
8509 * @param currentFunc
8510 * The index of the function being invoked
8513 * The total number of functions in this interface
8516 * The amount of time this function waited for a thread
8519 * The amount of time this function invocation took to execute
8522 * The number bytes sent by this invocation
8525 * The number bytes received by this invocation
8528 * If true, this invocation was made to a server
8532 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8533 afs_uint32 currentFunc, afs_uint32 totalFunc,
8534 struct clock *queueTime, struct clock *execTime,
8535 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8541 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8542 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8544 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8545 queueTime, execTime, sent64, rcvd64,
8552 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8556 * IN callerVersion - the rpc stat version of the caller.
8558 * IN count - the number of entries to marshall.
8560 * IN stats - pointer to stats to be marshalled.
8562 * OUT ptr - Where to store the marshalled data.
8569 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8570 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8576 * We only support the first version
8578 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8579 *(ptr++) = stats->remote_peer;
8580 *(ptr++) = stats->remote_port;
8581 *(ptr++) = stats->remote_is_server;
8582 *(ptr++) = stats->interfaceId;
8583 *(ptr++) = stats->func_total;
8584 *(ptr++) = stats->func_index;
8585 *(ptr++) = stats->invocations >> 32;
8586 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8587 *(ptr++) = stats->bytes_sent >> 32;
8588 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8589 *(ptr++) = stats->bytes_rcvd >> 32;
8590 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8591 *(ptr++) = stats->queue_time_sum.sec;
8592 *(ptr++) = stats->queue_time_sum.usec;
8593 *(ptr++) = stats->queue_time_sum_sqr.sec;
8594 *(ptr++) = stats->queue_time_sum_sqr.usec;
8595 *(ptr++) = stats->queue_time_min.sec;
8596 *(ptr++) = stats->queue_time_min.usec;
8597 *(ptr++) = stats->queue_time_max.sec;
8598 *(ptr++) = stats->queue_time_max.usec;
8599 *(ptr++) = stats->execution_time_sum.sec;
8600 *(ptr++) = stats->execution_time_sum.usec;
8601 *(ptr++) = stats->execution_time_sum_sqr.sec;
8602 *(ptr++) = stats->execution_time_sum_sqr.usec;
8603 *(ptr++) = stats->execution_time_min.sec;
8604 *(ptr++) = stats->execution_time_min.usec;
8605 *(ptr++) = stats->execution_time_max.sec;
8606 *(ptr++) = stats->execution_time_max.usec;
8612 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8617 * IN callerVersion - the rpc stat version of the caller
8619 * OUT myVersion - the rpc stat version of this function
8621 * OUT clock_sec - local time seconds
8623 * OUT clock_usec - local time microseconds
8625 * OUT allocSize - the number of bytes allocated to contain stats
8627 * OUT statCount - the number stats retrieved from this process.
8629 * OUT stats - the actual stats retrieved from this process.
8633 * Returns void. If successful, stats will != NULL.
8637 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8638 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8639 size_t * allocSize, afs_uint32 * statCount,
8640 afs_uint32 ** stats)
8650 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8653 * Check to see if stats are enabled
8656 MUTEX_ENTER(&rx_rpc_stats);
8657 if (!rxi_monitor_processStats) {
8658 MUTEX_EXIT(&rx_rpc_stats);
8662 clock_GetTime(&now);
8663 *clock_sec = now.sec;
8664 *clock_usec = now.usec;
8667 * Allocate the space based upon the caller version
8669 * If the client is at an older version than we are,
8670 * we return the statistic data in the older data format, but
8671 * we still return our version number so the client knows we
8672 * are maintaining more data than it can retrieve.
8675 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8676 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8677 *statCount = rxi_rpc_process_stat_cnt;
8680 * This can't happen yet, but in the future version changes
8681 * can be handled by adding additional code here
8685 if (space > (size_t) 0) {
8687 ptr = *stats = rxi_Alloc(space);
8690 struct opr_queue *cursor;
8692 for (opr_queue_Scan(&processStats, cursor)) {
8693 struct rx_interface_stat *rpc_stat =
8694 opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8696 * Copy the data based upon the caller version
8698 rx_MarshallProcessRPCStats(callerVersion,
8699 rpc_stat->stats[0].func_total,
8700 rpc_stat->stats, &ptr);
8706 MUTEX_EXIT(&rx_rpc_stats);
8711 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8715 * IN callerVersion - the rpc stat version of the caller
8717 * OUT myVersion - the rpc stat version of this function
8719 * OUT clock_sec - local time seconds
8721 * OUT clock_usec - local time microseconds
8723 * OUT allocSize - the number of bytes allocated to contain stats
8725 * OUT statCount - the number of stats retrieved from the individual
8728 * OUT stats - the actual stats retrieved from the individual peer structures.
8732 * Returns void. If successful, stats will != NULL.
8736 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8737 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8738 size_t * allocSize, afs_uint32 * statCount,
8739 afs_uint32 ** stats)
8749 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8752 * Check to see if stats are enabled
8755 MUTEX_ENTER(&rx_rpc_stats);
8756 if (!rxi_monitor_peerStats) {
8757 MUTEX_EXIT(&rx_rpc_stats);
8761 clock_GetTime(&now);
8762 *clock_sec = now.sec;
8763 *clock_usec = now.usec;
8766 * Allocate the space based upon the caller version
8768 * If the client is at an older version than we are,
8769 * we return the statistic data in the older data format, but
8770 * we still return our version number so the client knows we
8771 * are maintaining more data than it can retrieve.
8774 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8775 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8776 *statCount = rxi_rpc_peer_stat_cnt;
8779 * This can't happen yet, but in the future version changes
8780 * can be handled by adding additional code here
8784 if (space > (size_t) 0) {
8786 ptr = *stats = rxi_Alloc(space);
8789 struct opr_queue *cursor;
8791 for (opr_queue_Scan(&peerStats, cursor)) {
8792 struct rx_interface_stat *rpc_stat
8793 = opr_queue_Entry(cursor, struct rx_interface_stat,
8797 * Copy the data based upon the caller version
8799 rx_MarshallProcessRPCStats(callerVersion,
8800 rpc_stat->stats[0].func_total,
8801 rpc_stat->stats, &ptr);
8807 MUTEX_EXIT(&rx_rpc_stats);
8812 * rx_FreeRPCStats - free memory allocated by
8813 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8817 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8818 * rx_RetrievePeerRPCStats
8820 * IN allocSize - the number of bytes in stats.
8828 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8830 rxi_Free(stats, allocSize);
8834 * rx_queryProcessRPCStats - see if process rpc stat collection is
8835 * currently enabled.
8841 * Returns 0 if stats are not enabled != 0 otherwise
8845 rx_queryProcessRPCStats(void)
8848 MUTEX_ENTER(&rx_rpc_stats);
8849 rc = rxi_monitor_processStats;
8850 MUTEX_EXIT(&rx_rpc_stats);
8855 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8861 * Returns 0 if stats are not enabled != 0 otherwise
8865 rx_queryPeerRPCStats(void)
8868 MUTEX_ENTER(&rx_rpc_stats);
8869 rc = rxi_monitor_peerStats;
8870 MUTEX_EXIT(&rx_rpc_stats);
8875 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8885 rx_enableProcessRPCStats(void)
8887 MUTEX_ENTER(&rx_rpc_stats);
8888 rx_enable_stats = 1;
8889 rxi_monitor_processStats = 1;
8890 MUTEX_EXIT(&rx_rpc_stats);
8894 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8904 rx_enablePeerRPCStats(void)
8906 MUTEX_ENTER(&rx_rpc_stats);
8907 rx_enable_stats = 1;
8908 rxi_monitor_peerStats = 1;
8909 MUTEX_EXIT(&rx_rpc_stats);
8913 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8923 rx_disableProcessRPCStats(void)
8925 struct opr_queue *cursor, *store;
8928 MUTEX_ENTER(&rx_rpc_stats);
8931 * Turn off process statistics and if peer stats is also off, turn
8935 rxi_monitor_processStats = 0;
8936 if (rxi_monitor_peerStats == 0) {
8937 rx_enable_stats = 0;
8940 for (opr_queue_ScanSafe(&processStats, cursor, store)) {
8941 unsigned int num_funcs = 0;
8942 struct rx_interface_stat *rpc_stat
8943 = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
8945 opr_queue_Remove(&rpc_stat->entry);
8947 num_funcs = rpc_stat->stats[0].func_total;
8949 sizeof(rx_interface_stat_t) +
8950 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8952 rxi_Free(rpc_stat, space);
8953 rxi_rpc_process_stat_cnt -= num_funcs;
8955 MUTEX_EXIT(&rx_rpc_stats);
8959 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8969 rx_disablePeerRPCStats(void)
8971 struct rx_peer **peer_ptr, **peer_end;
8975 * Turn off peer statistics and if process stats is also off, turn
8979 rxi_monitor_peerStats = 0;
8980 if (rxi_monitor_processStats == 0) {
8981 rx_enable_stats = 0;
8984 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8985 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8987 struct rx_peer *peer, *next, *prev;
8989 MUTEX_ENTER(&rx_peerHashTable_lock);
8990 MUTEX_ENTER(&rx_rpc_stats);
8991 for (prev = peer = *peer_ptr; peer; peer = next) {
8993 code = MUTEX_TRYENTER(&peer->peer_lock);
8996 struct opr_queue *cursor, *store;
8998 if (prev == *peer_ptr) {
9009 MUTEX_EXIT(&rx_peerHashTable_lock);
9011 for (opr_queue_ScanSafe(&peer->rpcStats, cursor, store)) {
9012 unsigned int num_funcs = 0;
9013 struct rx_interface_stat *rpc_stat
9014 = opr_queue_Entry(cursor, struct rx_interface_stat,
9017 opr_queue_Remove(&rpc_stat->entry);
9018 opr_queue_Remove(&rpc_stat->entryPeers);
9019 num_funcs = rpc_stat->stats[0].func_total;
9021 sizeof(rx_interface_stat_t) +
9022 rpc_stat->stats[0].func_total *
9023 sizeof(rx_function_entry_v1_t);
9025 rxi_Free(rpc_stat, space);
9026 rxi_rpc_peer_stat_cnt -= num_funcs;
9028 MUTEX_EXIT(&peer->peer_lock);
9030 MUTEX_ENTER(&rx_peerHashTable_lock);
9040 MUTEX_EXIT(&rx_rpc_stats);
9041 MUTEX_EXIT(&rx_peerHashTable_lock);
9046 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
9051 * IN clearFlag - flag indicating which stats to clear
9059 rx_clearProcessRPCStats(afs_uint32 clearFlag)
9061 struct opr_queue *cursor;
9063 MUTEX_ENTER(&rx_rpc_stats);
9065 for (opr_queue_Scan(&processStats, cursor)) {
9066 unsigned int num_funcs = 0, i;
9067 struct rx_interface_stat *rpc_stat
9068 = opr_queue_Entry(rpc_stat, struct rx_interface_stat, entry);
9070 num_funcs = rpc_stat->stats[0].func_total;
9071 for (i = 0; i < num_funcs; i++) {
9072 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9073 rpc_stat->stats[i].invocations = 0;
9075 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9076 rpc_stat->stats[i].bytes_sent = 0;
9078 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9079 rpc_stat->stats[i].bytes_rcvd = 0;
9081 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9082 rpc_stat->stats[i].queue_time_sum.sec = 0;
9083 rpc_stat->stats[i].queue_time_sum.usec = 0;
9085 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9086 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9087 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9089 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9090 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9091 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9093 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9094 rpc_stat->stats[i].queue_time_max.sec = 0;
9095 rpc_stat->stats[i].queue_time_max.usec = 0;
9097 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9098 rpc_stat->stats[i].execution_time_sum.sec = 0;
9099 rpc_stat->stats[i].execution_time_sum.usec = 0;
9101 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9102 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9103 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9105 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9106 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9107 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9109 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9110 rpc_stat->stats[i].execution_time_max.sec = 0;
9111 rpc_stat->stats[i].execution_time_max.usec = 0;
9116 MUTEX_EXIT(&rx_rpc_stats);
9120 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
9125 * IN clearFlag - flag indicating which stats to clear
9133 rx_clearPeerRPCStats(afs_uint32 clearFlag)
9135 struct opr_queue *cursor;
9137 MUTEX_ENTER(&rx_rpc_stats);
9139 for (opr_queue_Scan(&peerStats, cursor)) {
9140 unsigned int num_funcs, i;
9141 struct rx_interface_stat *rpc_stat
9142 = opr_queue_Entry(cursor, struct rx_interface_stat, entryPeers);
9144 num_funcs = rpc_stat->stats[0].func_total;
9145 for (i = 0; i < num_funcs; i++) {
9146 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
9147 rpc_stat->stats[i].invocations = 0;
9149 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
9150 rpc_stat->stats[i].bytes_sent = 0;
9152 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
9153 rpc_stat->stats[i].bytes_rcvd = 0;
9155 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
9156 rpc_stat->stats[i].queue_time_sum.sec = 0;
9157 rpc_stat->stats[i].queue_time_sum.usec = 0;
9159 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
9160 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
9161 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
9163 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
9164 rpc_stat->stats[i].queue_time_min.sec = 9999999;
9165 rpc_stat->stats[i].queue_time_min.usec = 9999999;
9167 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
9168 rpc_stat->stats[i].queue_time_max.sec = 0;
9169 rpc_stat->stats[i].queue_time_max.usec = 0;
9171 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
9172 rpc_stat->stats[i].execution_time_sum.sec = 0;
9173 rpc_stat->stats[i].execution_time_sum.usec = 0;
9175 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
9176 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
9177 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
9179 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
9180 rpc_stat->stats[i].execution_time_min.sec = 9999999;
9181 rpc_stat->stats[i].execution_time_min.usec = 9999999;
9183 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
9184 rpc_stat->stats[i].execution_time_max.sec = 0;
9185 rpc_stat->stats[i].execution_time_max.usec = 0;
9190 MUTEX_EXIT(&rx_rpc_stats);
9194 * rxi_rxstat_userok points to a routine that returns 1 if the caller
9195 * is authorized to enable/disable/clear RX statistics.
9197 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
9200 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
9202 rxi_rxstat_userok = proc;
9206 rx_RxStatUserOk(struct rx_call *call)
9208 if (!rxi_rxstat_userok)
9210 return rxi_rxstat_userok(call);
9215 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
9216 * function in the MSVC runtime DLL (msvcrt.dll).
9218 * Note: the system serializes calls to this function.
9221 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
9222 DWORD reason, /* reason function is being called */
9223 LPVOID reserved) /* reserved for future use */
9226 case DLL_PROCESS_ATTACH:
9227 /* library is being attached to a process */
9231 case DLL_PROCESS_DETACH:
9238 #endif /* AFS_NT40_ENV */
9241 int rx_DumpCalls(FILE *outputFile, char *cookie)
9243 #ifdef RXDEBUG_PACKET
9244 #ifdef KDUMP_RX_LOCK
9245 struct rx_call_rx_lock *c;
9252 #define RXDPRINTF sprintf
9253 #define RXDPRINTOUT output
9255 #define RXDPRINTF fprintf
9256 #define RXDPRINTOUT outputFile
9259 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9261 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9264 for (c = rx_allCallsp; c; c = c->allNextp) {
9265 u_short rqc, tqc, iovqc;
9267 MUTEX_ENTER(&c->lock);
9268 rqc = opr_queue_Count(&c->rq);
9269 tqc = opr_queue_Count(&c->tq);
9270 iovqc = opr_queue_Count(&c->app.iovq);
9272 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, "
9273 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9274 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9275 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9276 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9277 #ifdef RX_ENABLE_LOCKS
9280 #ifdef RX_REFCOUNT_CHECK
9281 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9282 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9285 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,
9286 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9287 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9288 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9289 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9290 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9291 #ifdef RX_ENABLE_LOCKS
9292 , (afs_uint32)c->refCount
9294 #ifdef RX_REFCOUNT_CHECK
9295 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9298 MUTEX_EXIT(&c->lock);
9301 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9304 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9306 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9308 #endif /* RXDEBUG_PACKET */