2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
13 #include <afs/param.h>
16 # include "afs/sysincludes.h"
17 # include "afsincludes.h"
22 # ifdef AFS_LINUX20_ENV
23 # include "h/socket.h"
25 # include "netinet/in.h"
27 # include "netinet/ip6.h"
28 # include "inet/common.h"
30 # include "inet/ip_ire.h"
32 # include "afs/afs_args.h"
33 # include "afs/afs_osi.h"
34 # ifdef RX_KERNEL_TRACE
35 # include "rx_kcommon.h"
37 # if defined(AFS_AIX_ENV)
41 # undef RXDEBUG /* turn off debugging */
43 # if defined(AFS_SGI_ENV)
44 # include "sys/debug.h"
47 # include "afs/sysincludes.h"
48 # include "afsincludes.h"
49 # endif /* !UKERNEL */
50 # include "afs/lock.h"
51 # include "rx_kmutex.h"
52 # include "rx_kernel.h"
53 # define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
54 # define AFSOP_STOP_AFS 211 /* Stop AFS process */
55 # define AFSOP_STOP_BKG 212 /* Stop BKG process */
56 extern afs_int32 afs_termState;
58 # include "sys/lockl.h"
59 # include "sys/lock_def.h"
60 # endif /* AFS_AIX41_ENV */
61 # include "afs/rxgen_consts.h"
66 # include <afs/afsutil.h>
67 # include <WINNT\afsreg.h>
76 #include "rx_atomic.h"
77 #include "rx_globals.h"
79 #include "rx_internal.h"
86 #include "rx_packet.h"
88 #include <afs/rxgen_consts.h>
91 #ifdef AFS_PTHREAD_ENV
93 int (*registerProgram) (pid_t, char *) = 0;
94 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
97 int (*registerProgram) (PROCESS, char *) = 0;
98 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
102 /* Local static routines */
103 static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
104 static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
105 struct rx_call *, struct rx_peer *,
107 static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
109 static void rxi_SendDelayedAck(struct rxevent *event, void *call,
110 void *dummy, int dummy2);
111 static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
112 void *dummy, int dummy2);
113 static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
114 void *unused, int unused2);
115 static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
116 void *unused2, int unused3);
117 static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
118 struct rx_packet *packet,
119 int istack, int force);
120 static void rxi_AckAll(struct rx_call *call);
121 static struct rx_connection
122 *rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
123 u_short serviceId, afs_uint32 cid,
124 afs_uint32 epoch, int type, u_int securityIndex);
125 static struct rx_packet
126 *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
127 int istack, osi_socket socket,
128 afs_uint32 host, u_short port, int *tnop,
129 struct rx_call **newcallp);
130 static struct rx_packet
131 *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
133 static struct rx_packet
134 *rxi_ReceiveResponsePacket(struct rx_connection *conn,
135 struct rx_packet *np, int istack);
136 static struct rx_packet
137 *rxi_ReceiveChallengePacket(struct rx_connection *conn,
138 struct rx_packet *np, int istack);
139 static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
140 int *tnop, struct rx_call **newcallp);
141 static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
142 static void rxi_ClearReceiveQueue(struct rx_call *call);
143 static void rxi_ResetCall(struct rx_call *call, int newcall);
144 static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
145 static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
146 static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
147 static void rxi_KeepAliveOn(struct rx_call *call);
148 static void rxi_GrowMTUOn(struct rx_call *call);
149 static void rxi_ChallengeOn(struct rx_connection *conn);
151 #ifdef RX_ENABLE_LOCKS
152 static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
153 static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
155 static int rxi_CheckCall(struct rx_call *call);
158 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
160 rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
161 rx_atomic_t rxi_start_in_error;
163 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
165 /* Constant delay time before sending an acknowledge of the last packet
166 * received. This is to avoid sending an extra acknowledge when the
167 * client is about to make another call, anyway, or the server is
170 * The lastAckDelay may not exceeed 400ms without causing peers to
171 * unecessarily timeout.
173 struct clock rx_lastAckDelay = {0, 400000};
175 /* Constant delay time before sending a soft ack when none was requested.
176 * This is to make sure we send soft acks before the sender times out,
177 * Normally we wait and send a hard ack when the receiver consumes the packet
179 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
180 * will require changes to the peer's RTT calculations.
182 struct clock rx_softAckDelay = {0, 100000};
185 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
186 * currently allocated within rx. This number is used to allocate the
187 * memory required to return the statistics when queried.
188 * Protected by the rx_rpc_stats mutex.
191 static unsigned int rxi_rpc_peer_stat_cnt;
194 * rxi_rpc_process_stat_cnt counts the total number of local process stat
195 * structures currently allocated within rx. The number is used to allocate
196 * the memory required to return the statistics when queried.
197 * Protected by the rx_rpc_stats mutex.
200 static unsigned int rxi_rpc_process_stat_cnt;
203 * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
204 * errors should be reported to the application when a call channel appears busy
205 * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
206 * and there are other call channels in the connection that are not busy.
207 * If 0, we do not return errors upon receiving busy packets; we just keep
208 * trying on the same call channel until we hit a timeout.
210 static afs_int32 rxi_busyChannelError = 0;
212 rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
213 rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
215 #if !defined(offsetof)
216 #include <stddef.h> /* for definition of offsetof() */
219 #ifdef RX_ENABLE_LOCKS
220 afs_kmutex_t rx_atomic_mutex;
223 /* Forward prototypes */
224 static struct rx_call * rxi_NewCall(struct rx_connection *, int);
227 putConnection (struct rx_connection *conn) {
228 MUTEX_ENTER(&rx_refcnt_mutex);
230 MUTEX_EXIT(&rx_refcnt_mutex);
233 #ifdef AFS_PTHREAD_ENV
236 * Use procedural initialization of mutexes/condition variables
240 extern afs_kmutex_t rx_quota_mutex;
241 extern afs_kmutex_t rx_pthread_mutex;
242 extern afs_kmutex_t rx_packets_mutex;
243 extern afs_kmutex_t rx_refcnt_mutex;
244 extern afs_kmutex_t des_init_mutex;
245 extern afs_kmutex_t des_random_mutex;
246 extern afs_kmutex_t rx_clock_mutex;
247 extern afs_kmutex_t rxi_connCacheMutex;
248 extern afs_kmutex_t event_handler_mutex;
249 extern afs_kmutex_t listener_mutex;
250 extern afs_kmutex_t rx_if_init_mutex;
251 extern afs_kmutex_t rx_if_mutex;
253 extern afs_kcondvar_t rx_event_handler_cond;
254 extern afs_kcondvar_t rx_listener_cond;
256 static afs_kmutex_t epoch_mutex;
257 static afs_kmutex_t rx_init_mutex;
258 static afs_kmutex_t rx_debug_mutex;
259 static afs_kmutex_t rx_rpc_stats;
262 rxi_InitPthread(void)
264 MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
265 MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
266 MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
267 MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
268 MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
269 MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
270 MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
271 MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
272 MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
273 MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
274 MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
275 MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
276 MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
277 MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
278 MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
280 CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
281 CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
283 osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
284 osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
286 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
287 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
288 #ifdef RX_ENABLE_LOCKS
291 #endif /* RX_LOCKS_DB */
292 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
293 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
295 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
297 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
299 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
301 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
302 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
303 #endif /* RX_ENABLE_LOCKS */
306 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
307 #define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
309 * The rx_stats_mutex mutex protects the following global variables:
310 * rxi_lowConnRefCount
311 * rxi_lowPeerRefCount
320 * The rx_quota_mutex mutex protects the following global variables:
328 * The rx_freePktQ_lock protects the following global variables:
333 * The rx_packets_mutex mutex protects the following global variables:
341 * The rx_pthread_mutex mutex protects the following global variables:
342 * rxi_fcfs_thread_num
345 #define INIT_PTHREAD_LOCKS
349 /* Variables for handling the minProcs implementation. availProcs gives the
350 * number of threads available in the pool at this moment (not counting dudes
351 * executing right now). totalMin gives the total number of procs required
352 * for handling all minProcs requests. minDeficit is a dynamic variable
353 * tracking the # of procs required to satisfy all of the remaining minProcs
355 * For fine grain locking to work, the quota check and the reservation of
356 * a server thread has to come while rxi_availProcs and rxi_minDeficit
357 * are locked. To this end, the code has been modified under #ifdef
358 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
359 * same time. A new function, ReturnToServerPool() returns the allocation.
361 * A call can be on several queue's (but only one at a time). When
362 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
363 * that no one else is touching the queue. To this end, we store the address
364 * of the queue lock in the call structure (under the call lock) when we
365 * put the call on a queue, and we clear the call_queue_lock when the
366 * call is removed from a queue (once the call lock has been obtained).
367 * This allows rxi_ResetCall to safely synchronize with others wishing
368 * to manipulate the queue.
371 #if defined(RX_ENABLE_LOCKS)
372 static afs_kmutex_t rx_rpc_stats;
375 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
376 ** pretty good that the next packet coming in is from the same connection
377 ** as the last packet, since we're send multiple packets in a transmit window.
379 struct rx_connection *rxLastConn = 0;
381 #ifdef RX_ENABLE_LOCKS
382 /* The locking hierarchy for rx fine grain locking is composed of these
385 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
386 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
387 * call->lock - locks call data fields.
388 * These are independent of each other:
389 * rx_freeCallQueue_lock
394 * serverQueueEntry->lock
395 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
397 * peer->lock - locks peer data fields.
398 * conn_data_lock - that more than one thread is not updating a conn data
399 * field at the same time.
410 * Do we need a lock to protect the peer field in the conn structure?
411 * conn->peer was previously a constant for all intents and so has no
412 * lock protecting this field. The multihomed client delta introduced
413 * a RX code change : change the peer field in the connection structure
414 * to that remote interface from which the last packet for this
415 * connection was sent out. This may become an issue if further changes
418 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
419 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
421 /* rxdb_fileID is used to identify the lock location, along with line#. */
422 static int rxdb_fileID = RXDB_FILE_RX;
423 #endif /* RX_LOCKS_DB */
424 #else /* RX_ENABLE_LOCKS */
425 #define SET_CALL_QUEUE_LOCK(C, L)
426 #define CLEAR_CALL_QUEUE_LOCK(C)
427 #endif /* RX_ENABLE_LOCKS */
428 struct rx_serverQueueEntry *rx_waitForPacket = 0;
429 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
431 /* ------------Exported Interfaces------------- */
433 /* This function allows rxkad to set the epoch to a suitably random number
434 * which rx_NewConnection will use in the future. The principle purpose is to
435 * get rxnull connections to use the same epoch as the rxkad connections do, at
436 * least once the first rxkad connection is established. This is important now
437 * that the host/port addresses aren't used in FindConnection: the uniqueness
438 * of epoch/cid matters and the start time won't do. */
440 #ifdef AFS_PTHREAD_ENV
442 * This mutex protects the following global variables:
446 #define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
447 #define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
451 #endif /* AFS_PTHREAD_ENV */
454 rx_SetEpoch(afs_uint32 epoch)
461 /* Initialize rx. A port number may be mentioned, in which case this
462 * becomes the default port number for any service installed later.
463 * If 0 is provided for the port number, a random port will be chosen
464 * by the kernel. Whether this will ever overlap anything in
465 * /etc/services is anybody's guess... Returns 0 on success, -1 on
470 int rxinit_status = 1;
471 #ifdef AFS_PTHREAD_ENV
473 * This mutex protects the following global variables:
477 #define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
478 #define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
481 #define UNLOCK_RX_INIT
485 rx_InitHost(u_int host, u_int port)
492 char *htable, *ptable;
499 if (rxinit_status == 0) {
500 tmp_status = rxinit_status;
502 return tmp_status; /* Already started; return previous error code. */
508 if (afs_winsockInit() < 0)
514 * Initialize anything necessary to provide a non-premptive threading
517 rxi_InitializeThreadSupport();
520 /* Allocate and initialize a socket for client and perhaps server
523 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
524 if (rx_socket == OSI_NULLSOCKET) {
528 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
531 #endif /* RX_LOCKS_DB */
532 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
533 MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
534 MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
535 MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
536 MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 0);
537 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
538 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
539 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
540 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
542 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
544 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
546 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
548 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
549 #if defined(AFS_HPUX110_ENV)
551 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
552 #endif /* AFS_HPUX110_ENV */
553 #endif /* RX_ENABLE_LOCKS && KERNEL */
556 rx_connDeadTime = 12;
557 rx_tranquil = 0; /* reset flag */
558 rxi_ResetStatistics();
559 htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
560 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
561 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
562 ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
563 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
564 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
566 /* Malloc up a bunch of packets & buffers */
568 queue_Init(&rx_freePacketQueue);
569 rxi_NeedMorePackets = FALSE;
570 rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
572 /* enforce a minimum number of allocated packets */
573 if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
574 rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;
576 /* allocate the initial free packet pool */
577 #ifdef RX_ENABLE_TSFPQ
578 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
579 #else /* RX_ENABLE_TSFPQ */
580 rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2); /* fudge */
581 #endif /* RX_ENABLE_TSFPQ */
588 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
589 tv.tv_sec = clock_now.sec;
590 tv.tv_usec = clock_now.usec;
591 srand((unsigned int)tv.tv_usec);
598 #if defined(KERNEL) && !defined(UKERNEL)
599 /* Really, this should never happen in a real kernel */
602 struct sockaddr_in addr;
604 int addrlen = sizeof(addr);
606 socklen_t addrlen = sizeof(addr);
608 if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
610 osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
613 rx_port = addr.sin_port;
616 rx_stats.minRtt.sec = 9999999;
618 rx_SetEpoch(tv.tv_sec | 0x80000000);
620 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
621 * will provide a randomer value. */
623 MUTEX_ENTER(&rx_quota_mutex);
624 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
625 MUTEX_EXIT(&rx_quota_mutex);
626 /* *Slightly* random start time for the cid. This is just to help
627 * out with the hashing function at the peer */
628 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
629 rx_connHashTable = (struct rx_connection **)htable;
630 rx_peerHashTable = (struct rx_peer **)ptable;
632 rx_hardAckDelay.sec = 0;
633 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
635 rxevent_Init(20, rxi_ReScheduleEvents);
637 /* Initialize various global queues */
638 queue_Init(&rx_idleServerQueue);
639 queue_Init(&rx_incomingCallQueue);
640 queue_Init(&rx_freeCallQueue);
642 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
643 /* Initialize our list of usable IP addresses. */
647 #if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
648 /* Start listener process (exact function is dependent on the
649 * implementation environment--kernel or user space) */
654 tmp_status = rxinit_status = 0;
662 return rx_InitHost(htonl(INADDR_ANY), port);
668 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
669 * maintaing the round trip timer.
674 * Start a new RTT timer for a given call and packet.
676 * There must be no resendEvent already listed for this call, otherwise this
677 * will leak events - intended for internal use within the RTO code only
680 * the RX call to start the timer for
681 * @param[in] lastPacket
682 * a flag indicating whether the last packet has been sent or not
684 * @pre call must be locked before calling this function
688 rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
690 struct clock now, retryTime;
695 clock_Add(&retryTime, &call->rto);
697 /* If we're sending the last packet, and we're the client, then the server
698 * may wait for an additional 400ms before returning the ACK, wait for it
699 * rather than hitting a timeout */
700 if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
701 clock_Addmsec(&retryTime, 400);
703 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
704 call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
709 * Cancel an RTT timer for a given call.
713 * the RX call to cancel the timer for
715 * @pre call must be locked before calling this function
720 rxi_rto_cancel(struct rx_call *call)
722 rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
726 * Tell the RTO timer that we have sent a packet.
728 * If the timer isn't already running, then start it. If the timer is running,
732 * the RX call that the packet has been sent on
733 * @param[in] lastPacket
734 * A flag which is true if this is the last packet for the call
736 * @pre The call must be locked before calling this function
741 rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
743 if (call->resendEvent)
746 rxi_rto_startTimer(call, lastPacket, istack);
750 * Tell the RTO timer that we have received an new ACK message
752 * This function should be called whenever a call receives an ACK that
753 * acknowledges new packets. Whatever happens, we stop the current timer.
754 * If there are unacked packets in the queue which have been sent, then
755 * we restart the timer from now. Otherwise, we leave it stopped.
758 * the RX call that the ACK has been received on
762 rxi_rto_packet_acked(struct rx_call *call, int istack)
764 struct rx_packet *p, *nxp;
766 rxi_rto_cancel(call);
768 if (queue_IsEmpty(&call->tq))
771 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
772 if (p->header.seq > call->tfirst + call->twind)
775 if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
776 rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
784 * Set an initial round trip timeout for a peer connection
786 * @param[in] secs The timeout to set in seconds
790 rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
791 peer->rtt = secs * 8000;
795 * Enables or disables the busy call channel error (RX_CALL_BUSY).
797 * @param[in] onoff Non-zero to enable busy call channel errors.
799 * @pre Neither rx_Init nor rx_InitHost have been called yet
802 rx_SetBusyChannelError(afs_int32 onoff)
804 osi_Assert(rxinit_status != 0);
805 rxi_busyChannelError = onoff ? 1 : 0;
809 * Set a delayed ack event on the specified call for the given time
811 * @param[in] call - the call on which to set the event
812 * @param[in] offset - the delay from now after which the event fires
815 rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
817 struct clock now, when;
821 clock_Add(&when, offset);
823 if (!call->delayedAckEvent
824 || clock_Gt(&call->delayedAckTime, &when)) {
826 rxevent_Cancel(&call->delayedAckEvent, call,
827 RX_CALL_REFCOUNT_DELAY);
828 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
830 call->delayedAckEvent = rxevent_Post(&when, &now,
833 call->delayedAckTime = when;
837 /* called with unincremented nRequestsRunning to see if it is OK to start
838 * a new thread in this service. Could be "no" for two reasons: over the
839 * max quota, or would prevent others from reaching their min quota.
841 #ifdef RX_ENABLE_LOCKS
842 /* This verion of QuotaOK reserves quota if it's ok while the
843 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
846 QuotaOK(struct rx_service *aservice)
848 /* check if over max quota */
849 if (aservice->nRequestsRunning >= aservice->maxProcs) {
853 /* under min quota, we're OK */
854 /* otherwise, can use only if there are enough to allow everyone
855 * to go to their min quota after this guy starts.
858 MUTEX_ENTER(&rx_quota_mutex);
859 if ((aservice->nRequestsRunning < aservice->minProcs)
860 || (rxi_availProcs > rxi_minDeficit)) {
861 aservice->nRequestsRunning++;
862 /* just started call in minProcs pool, need fewer to maintain
864 if (aservice->nRequestsRunning <= aservice->minProcs)
867 MUTEX_EXIT(&rx_quota_mutex);
870 MUTEX_EXIT(&rx_quota_mutex);
876 ReturnToServerPool(struct rx_service *aservice)
878 aservice->nRequestsRunning--;
879 MUTEX_ENTER(&rx_quota_mutex);
880 if (aservice->nRequestsRunning < aservice->minProcs)
883 MUTEX_EXIT(&rx_quota_mutex);
886 #else /* RX_ENABLE_LOCKS */
888 QuotaOK(struct rx_service *aservice)
891 /* under min quota, we're OK */
892 if (aservice->nRequestsRunning < aservice->minProcs)
895 /* check if over max quota */
896 if (aservice->nRequestsRunning >= aservice->maxProcs)
899 /* otherwise, can use only if there are enough to allow everyone
900 * to go to their min quota after this guy starts.
902 MUTEX_ENTER(&rx_quota_mutex);
903 if (rxi_availProcs > rxi_minDeficit)
905 MUTEX_EXIT(&rx_quota_mutex);
908 #endif /* RX_ENABLE_LOCKS */
911 /* Called by rx_StartServer to start up lwp's to service calls.
912 NExistingProcs gives the number of procs already existing, and which
913 therefore needn't be created. */
915 rxi_StartServerProcs(int nExistingProcs)
917 struct rx_service *service;
922 /* For each service, reserve N processes, where N is the "minimum"
923 * number of processes that MUST be able to execute a request in parallel,
924 * at any time, for that process. Also compute the maximum difference
925 * between any service's maximum number of processes that can run
926 * (i.e. the maximum number that ever will be run, and a guarantee
927 * that this number will run if other services aren't running), and its
928 * minimum number. The result is the extra number of processes that
929 * we need in order to provide the latter guarantee */
930 for (i = 0; i < RX_MAX_SERVICES; i++) {
932 service = rx_services[i];
933 if (service == (struct rx_service *)0)
935 nProcs += service->minProcs;
936 diff = service->maxProcs - service->minProcs;
940 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
941 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
942 for (i = 0; i < nProcs; i++) {
943 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
949 /* This routine is only required on Windows */
951 rx_StartClientThread(void)
953 #ifdef AFS_PTHREAD_ENV
955 pid = pthread_self();
956 #endif /* AFS_PTHREAD_ENV */
958 #endif /* AFS_NT40_ENV */
960 /* This routine must be called if any services are exported. If the
961 * donateMe flag is set, the calling process is donated to the server
964 rx_StartServer(int donateMe)
966 struct rx_service *service;
972 /* Start server processes, if necessary (exact function is dependent
973 * on the implementation environment--kernel or user space). DonateMe
974 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
975 * case, one less new proc will be created rx_StartServerProcs.
977 rxi_StartServerProcs(donateMe);
979 /* count up the # of threads in minProcs, and add set the min deficit to
980 * be that value, too.
982 for (i = 0; i < RX_MAX_SERVICES; i++) {
983 service = rx_services[i];
984 if (service == (struct rx_service *)0)
986 MUTEX_ENTER(&rx_quota_mutex);
987 rxi_totalMin += service->minProcs;
988 /* below works even if a thread is running, since minDeficit would
989 * still have been decremented and later re-incremented.
991 rxi_minDeficit += service->minProcs;
992 MUTEX_EXIT(&rx_quota_mutex);
995 /* Turn on reaping of idle server connections */
996 rxi_ReapConnections(NULL, NULL, NULL, 0);
1001 #ifndef AFS_NT40_ENV
1005 #ifdef AFS_PTHREAD_ENV
1007 pid = afs_pointer_to_int(pthread_self());
1008 #else /* AFS_PTHREAD_ENV */
1010 LWP_CurrentProcess(&pid);
1011 #endif /* AFS_PTHREAD_ENV */
1013 sprintf(name, "srv_%d", ++nProcs);
1014 if (registerProgram)
1015 (*registerProgram) (pid, name);
1017 #endif /* AFS_NT40_ENV */
1018 rx_ServerProc(NULL); /* Never returns */
1020 #ifdef RX_ENABLE_TSFPQ
1021 /* no use leaving packets around in this thread's local queue if
1022 * it isn't getting donated to the server thread pool.
1024 rxi_FlushLocalPacketsTSFPQ();
1025 #endif /* RX_ENABLE_TSFPQ */
1029 /* Create a new client connection to the specified service, using the
1030 * specified security object to implement the security model for this
1032 struct rx_connection *
1033 rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
1034 struct rx_securityClass *securityObject,
1035 int serviceSecurityIndex)
1039 struct rx_connection *conn;
1044 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
1045 "serviceSecurityIndex %d)\n",
1046 ntohl(shost), ntohs(sport), sservice, securityObject,
1047 serviceSecurityIndex));
1049 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
1050 * the case of kmem_alloc? */
1051 conn = rxi_AllocConnection();
1052 #ifdef RX_ENABLE_LOCKS
1053 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
1054 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
1055 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
1058 MUTEX_ENTER(&rx_connHashTable_lock);
1059 cid = (rx_nextCid += RX_MAXCALLS);
1060 conn->type = RX_CLIENT_CONNECTION;
1062 conn->epoch = rx_epoch;
1063 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
1064 conn->serviceId = sservice;
1065 conn->securityObject = securityObject;
1066 conn->securityData = (void *) 0;
1067 conn->securityIndex = serviceSecurityIndex;
1068 rx_SetConnDeadTime(conn, rx_connDeadTime);
1069 rx_SetConnSecondsUntilNatPing(conn, 0);
1070 conn->ackRate = RX_FAST_ACK_RATE;
1071 conn->nSpecific = 0;
1072 conn->specific = NULL;
1073 conn->challengeEvent = NULL;
1074 conn->delayedAbortEvent = NULL;
1075 conn->abortCount = 0;
1077 for (i = 0; i < RX_MAXCALLS; i++) {
1078 conn->twind[i] = rx_initSendWindow;
1079 conn->rwind[i] = rx_initReceiveWindow;
1080 conn->lastBusy[i] = 0;
1083 RXS_NewConnection(securityObject, conn);
1085 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
1087 conn->refCount++; /* no lock required since only this thread knows... */
1088 conn->next = rx_connHashTable[hashindex];
1089 rx_connHashTable[hashindex] = conn;
1090 if (rx_stats_active)
1091 rx_atomic_inc(&rx_stats.nClientConns);
1092 MUTEX_EXIT(&rx_connHashTable_lock);
1098 * Ensure a connection's timeout values are valid.
1100 * @param[in] conn The connection to check
1102 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
1103 * unless idleDeadTime and/or hardDeadTime are not set
1107 rxi_CheckConnTimeouts(struct rx_connection *conn)
1109 /* a connection's timeouts must have the relationship
1110 * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
1111 * total loss of network to a peer may cause an idle timeout instead of a
1112 * dead timeout, simply because the idle timeout gets hit first. Also set
1113 * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
1114 /* this logic is slightly complicated by the fact that
1115 * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
1117 conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
1118 if (conn->idleDeadTime) {
1119 conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
1121 if (conn->hardDeadTime) {
1122 if (conn->idleDeadTime) {
1123 conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
1125 conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
1131 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
1133 /* The idea is to set the dead time to a value that allows several
1134 * keepalives to be dropped without timing out the connection. */
1135 conn->secondsUntilDead = seconds;
1136 rxi_CheckConnTimeouts(conn);
1137 conn->secondsUntilPing = conn->secondsUntilDead / 6;
1141 rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
1143 conn->hardDeadTime = seconds;
1144 rxi_CheckConnTimeouts(conn);
1148 rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
1150 conn->idleDeadTime = seconds;
1151 conn->idleDeadDetection = (seconds ? 1 : 0);
1152 rxi_CheckConnTimeouts(conn);
1155 int rxi_lowPeerRefCount = 0;
1156 int rxi_lowConnRefCount = 0;
1159 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
1160 * NOTE: must not be called with rx_connHashTable_lock held.
1163 rxi_CleanupConnection(struct rx_connection *conn)
1165 /* Notify the service exporter, if requested, that this connection
1166 * is being destroyed */
1167 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
1168 (*conn->service->destroyConnProc) (conn);
1170 /* Notify the security module that this connection is being destroyed */
1171 RXS_DestroyConnection(conn->securityObject, conn);
1173 /* If this is the last connection using the rx_peer struct, set its
1174 * idle time to now. rxi_ReapConnections will reap it if it's still
1175 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
1177 MUTEX_ENTER(&rx_peerHashTable_lock);
1178 if (conn->peer->refCount < 2) {
1179 conn->peer->idleWhen = clock_Sec();
1180 if (conn->peer->refCount < 1) {
1181 conn->peer->refCount = 1;
1182 if (rx_stats_active) {
1183 MUTEX_ENTER(&rx_stats_mutex);
1184 rxi_lowPeerRefCount++;
1185 MUTEX_EXIT(&rx_stats_mutex);
1189 conn->peer->refCount--;
1190 MUTEX_EXIT(&rx_peerHashTable_lock);
1192 if (rx_stats_active)
1194 if (conn->type == RX_SERVER_CONNECTION)
1195 rx_atomic_dec(&rx_stats.nServerConns);
1197 rx_atomic_dec(&rx_stats.nClientConns);
1200 if (conn->specific) {
1202 for (i = 0; i < conn->nSpecific; i++) {
1203 if (conn->specific[i] && rxi_keyCreate_destructor[i])
1204 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
1205 conn->specific[i] = NULL;
1207 free(conn->specific);
1209 conn->specific = NULL;
1210 conn->nSpecific = 0;
1211 #endif /* !KERNEL */
1213 MUTEX_DESTROY(&conn->conn_call_lock);
1214 MUTEX_DESTROY(&conn->conn_data_lock);
1215 CV_DESTROY(&conn->conn_call_cv);
1217 rxi_FreeConnection(conn);
1220 /* Destroy the specified connection */
1222 rxi_DestroyConnection(struct rx_connection *conn)
1224 MUTEX_ENTER(&rx_connHashTable_lock);
1225 rxi_DestroyConnectionNoLock(conn);
1226 /* conn should be at the head of the cleanup list */
1227 if (conn == rx_connCleanup_list) {
1228 rx_connCleanup_list = rx_connCleanup_list->next;
1229 MUTEX_EXIT(&rx_connHashTable_lock);
1230 rxi_CleanupConnection(conn);
1232 #ifdef RX_ENABLE_LOCKS
1234 MUTEX_EXIT(&rx_connHashTable_lock);
1236 #endif /* RX_ENABLE_LOCKS */
1240 rxi_DestroyConnectionNoLock(struct rx_connection *conn)
1242 struct rx_connection **conn_ptr;
1244 struct rx_packet *packet;
1251 MUTEX_ENTER(&conn->conn_data_lock);
1252 MUTEX_ENTER(&rx_refcnt_mutex);
1253 if (conn->refCount > 0)
1256 if (rx_stats_active) {
1257 MUTEX_ENTER(&rx_stats_mutex);
1258 rxi_lowConnRefCount++;
1259 MUTEX_EXIT(&rx_stats_mutex);
1263 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1264 /* Busy; wait till the last guy before proceeding */
1265 MUTEX_EXIT(&rx_refcnt_mutex);
1266 MUTEX_EXIT(&conn->conn_data_lock);
1271 /* If the client previously called rx_NewCall, but it is still
1272 * waiting, treat this as a running call, and wait to destroy the
1273 * connection later when the call completes. */
1274 if ((conn->type == RX_CLIENT_CONNECTION)
1275 && (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
1276 conn->flags |= RX_CONN_DESTROY_ME;
1277 MUTEX_EXIT(&conn->conn_data_lock);
1281 MUTEX_EXIT(&rx_refcnt_mutex);
1282 MUTEX_EXIT(&conn->conn_data_lock);
1284 /* Check for extant references to this connection */
1285 MUTEX_ENTER(&conn->conn_call_lock);
1286 for (i = 0; i < RX_MAXCALLS; i++) {
1287 struct rx_call *call = conn->call[i];
1290 if (conn->type == RX_CLIENT_CONNECTION) {
1291 MUTEX_ENTER(&call->lock);
1292 if (call->delayedAckEvent) {
1293 /* Push the final acknowledgment out now--there
1294 * won't be a subsequent call to acknowledge the
1295 * last reply packets */
1296 rxevent_Cancel(&call->delayedAckEvent, call,
1297 RX_CALL_REFCOUNT_DELAY);
1298 if (call->state == RX_STATE_PRECALL
1299 || call->state == RX_STATE_ACTIVE) {
1300 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1305 MUTEX_EXIT(&call->lock);
1309 MUTEX_EXIT(&conn->conn_call_lock);
1311 #ifdef RX_ENABLE_LOCKS
1313 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1314 MUTEX_EXIT(&conn->conn_data_lock);
1316 /* Someone is accessing a packet right now. */
1320 #endif /* RX_ENABLE_LOCKS */
1323 /* Don't destroy the connection if there are any call
1324 * structures still in use */
1325 MUTEX_ENTER(&conn->conn_data_lock);
1326 conn->flags |= RX_CONN_DESTROY_ME;
1327 MUTEX_EXIT(&conn->conn_data_lock);
1332 if (conn->natKeepAliveEvent) {
1333 rxi_NatKeepAliveOff(conn);
1336 if (conn->delayedAbortEvent) {
1337 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
1338 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1340 MUTEX_ENTER(&conn->conn_data_lock);
1341 rxi_SendConnectionAbort(conn, packet, 0, 1);
1342 MUTEX_EXIT(&conn->conn_data_lock);
1343 rxi_FreePacket(packet);
1347 /* Remove from connection hash table before proceeding */
1349 &rx_connHashTable[CONN_HASH
1350 (peer->host, peer->port, conn->cid, conn->epoch,
1352 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1353 if (*conn_ptr == conn) {
1354 *conn_ptr = conn->next;
1358 /* if the conn that we are destroying was the last connection, then we
1359 * clear rxLastConn as well */
1360 if (rxLastConn == conn)
1363 /* Make sure the connection is completely reset before deleting it. */
1364 /* get rid of pending events that could zap us later */
1365 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
1366 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
1367 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
1369 /* Add the connection to the list of destroyed connections that
1370 * need to be cleaned up. This is necessary to avoid deadlocks
1371 * in the routines we call to inform others that this connection is
1372 * being destroyed. */
1373 conn->next = rx_connCleanup_list;
1374 rx_connCleanup_list = conn;
1377 /* Externally available version */
1379 rx_DestroyConnection(struct rx_connection *conn)
1384 rxi_DestroyConnection(conn);
1389 rx_GetConnection(struct rx_connection *conn)
1394 MUTEX_ENTER(&rx_refcnt_mutex);
1396 MUTEX_EXIT(&rx_refcnt_mutex);
1400 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1401 /* Wait for the transmit queue to no longer be busy.
1402 * requires the call->lock to be held */
1404 rxi_WaitforTQBusy(struct rx_call *call) {
1405 while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
1406 call->flags |= RX_CALL_TQ_WAIT;
1408 #ifdef RX_ENABLE_LOCKS
1409 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1410 CV_WAIT(&call->cv_tq, &call->lock);
1411 #else /* RX_ENABLE_LOCKS */
1412 osi_rxSleep(&call->tq);
1413 #endif /* RX_ENABLE_LOCKS */
1415 if (call->tqWaiters == 0) {
1416 call->flags &= ~RX_CALL_TQ_WAIT;
1423 rxi_WakeUpTransmitQueue(struct rx_call *call)
1425 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
1426 dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
1427 call, call->tqWaiters, call->flags));
1428 #ifdef RX_ENABLE_LOCKS
1429 osirx_AssertMine(&call->lock, "rxi_Start start");
1430 CV_BROADCAST(&call->cv_tq);
1431 #else /* RX_ENABLE_LOCKS */
1432 osi_rxWakeup(&call->tq);
1433 #endif /* RX_ENABLE_LOCKS */
1437 /* Start a new rx remote procedure call, on the specified connection.
1438 * If wait is set to 1, wait for a free call channel; otherwise return
1439 * 0. Maxtime gives the maximum number of seconds this call may take,
1440 * after rx_NewCall returns. After this time interval, a call to any
1441 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1442 * For fine grain locking, we hold the conn_call_lock in order to
1443 * to ensure that we don't get signalle after we found a call in an active
1444 * state and before we go to sleep.
1447 rx_NewCall(struct rx_connection *conn)
1449 int i, wait, ignoreBusy = 1;
1450 struct rx_call *call;
1451 struct clock queueTime;
1452 afs_uint32 leastBusy = 0;
1456 dpf(("rx_NewCall(conn %"AFS_PTR_FMT")\n", conn));
1459 clock_GetTime(&queueTime);
1461 * Check if there are others waiting for a new call.
1462 * If so, let them go first to avoid starving them.
1463 * This is a fairly simple scheme, and might not be
1464 * a complete solution for large numbers of waiters.
1466 * makeCallWaiters keeps track of the number of
1467 * threads waiting to make calls and the
1468 * RX_CONN_MAKECALL_WAITING flag bit is used to
1469 * indicate that there are indeed calls waiting.
1470 * The flag is set when the waiter is incremented.
1471 * It is only cleared when makeCallWaiters is 0.
1472 * This prevents us from accidently destroying the
1473 * connection while it is potentially about to be used.
1475 MUTEX_ENTER(&conn->conn_call_lock);
1476 MUTEX_ENTER(&conn->conn_data_lock);
1477 while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
1478 conn->flags |= RX_CONN_MAKECALL_WAITING;
1479 conn->makeCallWaiters++;
1480 MUTEX_EXIT(&conn->conn_data_lock);
1482 #ifdef RX_ENABLE_LOCKS
1483 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1487 MUTEX_ENTER(&conn->conn_data_lock);
1488 conn->makeCallWaiters--;
1489 if (conn->makeCallWaiters == 0)
1490 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1493 /* We are now the active thread in rx_NewCall */
1494 conn->flags |= RX_CONN_MAKECALL_ACTIVE;
1495 MUTEX_EXIT(&conn->conn_data_lock);
1500 for (i = 0; i < RX_MAXCALLS; i++) {
1501 call = conn->call[i];
1503 if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
1504 /* we're not ignoring busy call slots; only look at the
1505 * call slot that is the "least" busy */
1509 if (call->state == RX_STATE_DALLY) {
1510 MUTEX_ENTER(&call->lock);
1511 if (call->state == RX_STATE_DALLY) {
1512 if (ignoreBusy && conn->lastBusy[i]) {
1513 /* if we're ignoring busy call slots, skip any ones that
1514 * have lastBusy set */
1515 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1516 leastBusy = conn->lastBusy[i];
1518 MUTEX_EXIT(&call->lock);
1523 * We are setting the state to RX_STATE_RESET to
1524 * ensure that no one else will attempt to use this
1525 * call once we drop the conn->conn_call_lock and
1526 * call->lock. We must drop the conn->conn_call_lock
1527 * before calling rxi_ResetCall because the process
1528 * of clearing the transmit queue can block for an
1529 * extended period of time. If we block while holding
1530 * the conn->conn_call_lock, then all rx_EndCall
1531 * processing will block as well. This has a detrimental
1532 * effect on overall system performance.
1534 call->state = RX_STATE_RESET;
1535 (*call->callNumber)++;
1536 MUTEX_EXIT(&conn->conn_call_lock);
1537 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1538 rxi_ResetCall(call, 0);
1539 if (MUTEX_TRYENTER(&conn->conn_call_lock))
1543 * If we failed to be able to safely obtain the
1544 * conn->conn_call_lock we will have to drop the
1545 * call->lock to avoid a deadlock. When the call->lock
1546 * is released the state of the call can change. If it
1547 * is no longer RX_STATE_RESET then some other thread is
1550 MUTEX_EXIT(&call->lock);
1551 MUTEX_ENTER(&conn->conn_call_lock);
1552 MUTEX_ENTER(&call->lock);
1554 if (call->state == RX_STATE_RESET)
1558 * If we get here it means that after dropping
1559 * the conn->conn_call_lock and call->lock that
1560 * the call is no longer ours. If we can't find
1561 * a free call in the remaining slots we should
1562 * not go immediately to RX_CONN_MAKECALL_WAITING
1563 * because by dropping the conn->conn_call_lock
1564 * we have given up synchronization with rx_EndCall.
1565 * Instead, cycle through one more time to see if
1566 * we can find a call that can call our own.
1568 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1571 MUTEX_EXIT(&call->lock);
1574 if (ignoreBusy && conn->lastBusy[i]) {
1575 /* if we're ignoring busy call slots, skip any ones that
1576 * have lastBusy set */
1577 if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
1578 leastBusy = conn->lastBusy[i];
1583 /* rxi_NewCall returns with mutex locked */
1584 call = rxi_NewCall(conn, i);
1585 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1589 if (i < RX_MAXCALLS) {
1590 conn->lastBusy[i] = 0;
1591 call->flags &= ~RX_CALL_PEER_BUSY;
1596 if (leastBusy && ignoreBusy) {
1597 /* we didn't find a useable call slot, but we did see at least one
1598 * 'busy' slot; look again and only use a slot with the 'least
1604 MUTEX_ENTER(&conn->conn_data_lock);
1605 conn->flags |= RX_CONN_MAKECALL_WAITING;
1606 conn->makeCallWaiters++;
1607 MUTEX_EXIT(&conn->conn_data_lock);
1609 #ifdef RX_ENABLE_LOCKS
1610 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1614 MUTEX_ENTER(&conn->conn_data_lock);
1615 conn->makeCallWaiters--;
1616 if (conn->makeCallWaiters == 0)
1617 conn->flags &= ~RX_CONN_MAKECALL_WAITING;
1618 MUTEX_EXIT(&conn->conn_data_lock);
1620 /* Client is initially in send mode */
1621 call->state = RX_STATE_ACTIVE;
1622 call->error = conn->error;
1624 call->mode = RX_MODE_ERROR;
1626 call->mode = RX_MODE_SENDING;
1628 /* remember start time for call in case we have hard dead time limit */
1629 call->queueTime = queueTime;
1630 clock_GetTime(&call->startTime);
1631 call->bytesSent = 0;
1632 call->bytesRcvd = 0;
1634 /* Turn on busy protocol. */
1635 rxi_KeepAliveOn(call);
1637 /* Attempt MTU discovery */
1638 rxi_GrowMTUOn(call);
1641 * We are no longer the active thread in rx_NewCall
1643 MUTEX_ENTER(&conn->conn_data_lock);
1644 conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
1645 MUTEX_EXIT(&conn->conn_data_lock);
1648 * Wake up anyone else who might be giving us a chance to
1649 * run (see code above that avoids resource starvation).
1651 #ifdef RX_ENABLE_LOCKS
1652 CV_BROADCAST(&conn->conn_call_cv);
1656 MUTEX_EXIT(&conn->conn_call_lock);
1658 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1659 if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
1660 osi_Panic("rx_NewCall call about to be used without an empty tq");
1662 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1664 MUTEX_EXIT(&call->lock);
1667 dpf(("rx_NewCall(call %"AFS_PTR_FMT")\n", call));
1672 rxi_HasActiveCalls(struct rx_connection *aconn)
1675 struct rx_call *tcall;
1679 for (i = 0; i < RX_MAXCALLS; i++) {
1680 if ((tcall = aconn->call[i])) {
1681 if ((tcall->state == RX_STATE_ACTIVE)
1682 || (tcall->state == RX_STATE_PRECALL)) {
1693 rxi_GetCallNumberVector(struct rx_connection *aconn,
1694 afs_int32 * aint32s)
1697 struct rx_call *tcall;
1701 MUTEX_ENTER(&aconn->conn_call_lock);
1702 for (i = 0; i < RX_MAXCALLS; i++) {
1703 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1704 aint32s[i] = aconn->callNumber[i] + 1;
1706 aint32s[i] = aconn->callNumber[i];
1708 MUTEX_EXIT(&aconn->conn_call_lock);
1714 rxi_SetCallNumberVector(struct rx_connection *aconn,
1715 afs_int32 * aint32s)
1718 struct rx_call *tcall;
1722 MUTEX_ENTER(&aconn->conn_call_lock);
1723 for (i = 0; i < RX_MAXCALLS; i++) {
1724 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1725 aconn->callNumber[i] = aint32s[i] - 1;
1727 aconn->callNumber[i] = aint32s[i];
1729 MUTEX_EXIT(&aconn->conn_call_lock);
1734 /* Advertise a new service. A service is named locally by a UDP port
1735 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1738 char *serviceName; Name for identification purposes (e.g. the
1739 service name might be used for probing for
1742 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1743 char *serviceName, struct rx_securityClass **securityObjects,
1744 int nSecurityObjects,
1745 afs_int32(*serviceProc) (struct rx_call * acall))
1747 osi_socket socket = OSI_NULLSOCKET;
1748 struct rx_service *tservice;
1754 if (serviceId == 0) {
1756 "rx_NewService: service id for service %s is not non-zero.\n",
1763 "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",
1771 tservice = rxi_AllocService();
1774 #ifdef RX_ENABLE_LOCKS
1775 MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);
1778 for (i = 0; i < RX_MAX_SERVICES; i++) {
1779 struct rx_service *service = rx_services[i];
1781 if (port == service->servicePort && host == service->serviceHost) {
1782 if (service->serviceId == serviceId) {
1783 /* The identical service has already been
1784 * installed; if the caller was intending to
1785 * change the security classes used by this
1786 * service, he/she loses. */
1788 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1789 serviceName, serviceId, service->serviceName);
1791 rxi_FreeService(tservice);
1794 /* Different service, same port: re-use the socket
1795 * which is bound to the same port */
1796 socket = service->socket;
1799 if (socket == OSI_NULLSOCKET) {
1800 /* If we don't already have a socket (from another
1801 * service on same port) get a new one */
1802 socket = rxi_GetHostUDPSocket(host, port);
1803 if (socket == OSI_NULLSOCKET) {
1805 rxi_FreeService(tservice);
1810 service->socket = socket;
1811 service->serviceHost = host;
1812 service->servicePort = port;
1813 service->serviceId = serviceId;
1814 service->serviceName = serviceName;
1815 service->nSecurityObjects = nSecurityObjects;
1816 service->securityObjects = securityObjects;
1817 service->minProcs = 0;
1818 service->maxProcs = 1;
1819 service->idleDeadTime = 60;
1820 service->idleDeadErr = 0;
1821 service->connDeadTime = rx_connDeadTime;
1822 service->executeRequestProc = serviceProc;
1823 service->checkReach = 0;
1824 service->nSpecific = 0;
1825 service->specific = NULL;
1826 rx_services[i] = service; /* not visible until now */
1832 rxi_FreeService(tservice);
1833 (osi_Msg "rx_NewService: cannot support > %d services\n",
1838 /* Set configuration options for all of a service's security objects */
1841 rx_SetSecurityConfiguration(struct rx_service *service,
1842 rx_securityConfigVariables type,
1846 for (i = 0; i<service->nSecurityObjects; i++) {
1847 if (service->securityObjects[i]) {
1848 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1856 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1857 struct rx_securityClass **securityObjects, int nSecurityObjects,
1858 afs_int32(*serviceProc) (struct rx_call * acall))
1860 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1863 /* Generic request processing loop. This routine should be called
1864 * by the implementation dependent rx_ServerProc. If socketp is
1865 * non-null, it will be set to the file descriptor that this thread
1866 * is now listening on. If socketp is null, this routine will never
1869 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1871 struct rx_call *call;
1873 struct rx_service *tservice = NULL;
1880 call = rx_GetCall(threadID, tservice, socketp);
1881 if (socketp && *socketp != OSI_NULLSOCKET) {
1882 /* We are now a listener thread */
1888 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1889 #ifdef RX_ENABLE_LOCKS
1891 #endif /* RX_ENABLE_LOCKS */
1892 afs_termState = AFSOP_STOP_AFS;
1893 afs_osi_Wakeup(&afs_termState);
1894 #ifdef RX_ENABLE_LOCKS
1896 #endif /* RX_ENABLE_LOCKS */
1901 /* if server is restarting( typically smooth shutdown) then do not
1902 * allow any new calls.
1905 if (rx_tranquil && (call != NULL)) {
1909 MUTEX_ENTER(&call->lock);
1911 rxi_CallError(call, RX_RESTARTING);
1912 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1914 MUTEX_EXIT(&call->lock);
1919 tservice = call->conn->service;
1921 if (tservice->beforeProc)
1922 (*tservice->beforeProc) (call);
1924 code = tservice->executeRequestProc(call);
1926 if (tservice->afterProc)
1927 (*tservice->afterProc) (call, code);
1929 rx_EndCall(call, code);
1931 if (tservice->postProc)
1932 (*tservice->postProc) (code);
1934 if (rx_stats_active) {
1935 MUTEX_ENTER(&rx_stats_mutex);
1937 MUTEX_EXIT(&rx_stats_mutex);
1944 rx_WakeupServerProcs(void)
1946 struct rx_serverQueueEntry *np, *tqp;
1950 MUTEX_ENTER(&rx_serverPool_lock);
1952 #ifdef RX_ENABLE_LOCKS
1953 if (rx_waitForPacket)
1954 CV_BROADCAST(&rx_waitForPacket->cv);
1955 #else /* RX_ENABLE_LOCKS */
1956 if (rx_waitForPacket)
1957 osi_rxWakeup(rx_waitForPacket);
1958 #endif /* RX_ENABLE_LOCKS */
1959 MUTEX_ENTER(&freeSQEList_lock);
1960 for (np = rx_FreeSQEList; np; np = tqp) {
1961 tqp = *(struct rx_serverQueueEntry **)np;
1962 #ifdef RX_ENABLE_LOCKS
1963 CV_BROADCAST(&np->cv);
1964 #else /* RX_ENABLE_LOCKS */
1966 #endif /* RX_ENABLE_LOCKS */
1968 MUTEX_EXIT(&freeSQEList_lock);
1969 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1970 #ifdef RX_ENABLE_LOCKS
1971 CV_BROADCAST(&np->cv);
1972 #else /* RX_ENABLE_LOCKS */
1974 #endif /* RX_ENABLE_LOCKS */
1976 MUTEX_EXIT(&rx_serverPool_lock);
1981 * One thing that seems to happen is that all the server threads get
1982 * tied up on some empty or slow call, and then a whole bunch of calls
1983 * arrive at once, using up the packet pool, so now there are more
1984 * empty calls. The most critical resources here are server threads
1985 * and the free packet pool. The "doreclaim" code seems to help in
1986 * general. I think that eventually we arrive in this state: there
1987 * are lots of pending calls which do have all their packets present,
1988 * so they won't be reclaimed, are multi-packet calls, so they won't
1989 * be scheduled until later, and thus are tying up most of the free
1990 * packet pool for a very long time.
1992 * 1. schedule multi-packet calls if all the packets are present.
1993 * Probably CPU-bound operation, useful to return packets to pool.
1994 * Do what if there is a full window, but the last packet isn't here?
1995 * 3. preserve one thread which *only* runs "best" calls, otherwise
1996 * it sleeps and waits for that type of call.
1997 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1998 * the current dataquota business is badly broken. The quota isn't adjusted
1999 * to reflect how many packets are presently queued for a running call.
2000 * So, when we schedule a queued call with a full window of packets queued
2001 * up for it, that *should* free up a window full of packets for other 2d-class
2002 * calls to be able to use from the packet pool. But it doesn't.
2004 * NB. Most of the time, this code doesn't run -- since idle server threads
2005 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
2006 * as a new call arrives.
2008 /* Sleep until a call arrives. Returns a pointer to the call, ready
2009 * for an rx_Read. */
2010 #ifdef RX_ENABLE_LOCKS
2012 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2014 struct rx_serverQueueEntry *sq;
2015 struct rx_call *call = (struct rx_call *)0;
2016 struct rx_service *service = NULL;
2018 MUTEX_ENTER(&freeSQEList_lock);
2020 if ((sq = rx_FreeSQEList)) {
2021 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2022 MUTEX_EXIT(&freeSQEList_lock);
2023 } else { /* otherwise allocate a new one and return that */
2024 MUTEX_EXIT(&freeSQEList_lock);
2025 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2026 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2027 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2030 MUTEX_ENTER(&rx_serverPool_lock);
2031 if (cur_service != NULL) {
2032 ReturnToServerPool(cur_service);
2035 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2036 struct rx_call *tcall, *ncall, *choice2 = NULL;
2038 /* Scan for eligible incoming calls. A call is not eligible
2039 * if the maximum number of calls for its service type are
2040 * already executing */
2041 /* One thread will process calls FCFS (to prevent starvation),
2042 * while the other threads may run ahead looking for calls which
2043 * have all their input data available immediately. This helps
2044 * keep threads from blocking, waiting for data from the client. */
2045 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2046 service = tcall->conn->service;
2047 if (!QuotaOK(service)) {
2050 MUTEX_ENTER(&rx_pthread_mutex);
2051 if (tno == rxi_fcfs_thread_num
2052 || queue_IsLast(&rx_incomingCallQueue, tcall)) {
2053 MUTEX_EXIT(&rx_pthread_mutex);
2054 /* If we're the fcfs thread , then we'll just use
2055 * this call. If we haven't been able to find an optimal
2056 * choice, and we're at the end of the list, then use a
2057 * 2d choice if one has been identified. Otherwise... */
2058 call = (choice2 ? choice2 : tcall);
2059 service = call->conn->service;
2061 MUTEX_EXIT(&rx_pthread_mutex);
2062 if (!queue_IsEmpty(&tcall->rq)) {
2063 struct rx_packet *rp;
2064 rp = queue_First(&tcall->rq, rx_packet);
2065 if (rp->header.seq == 1) {
2067 || (rp->header.flags & RX_LAST_PACKET)) {
2069 } else if (rxi_2dchoice && !choice2
2070 && !(tcall->flags & RX_CALL_CLEARED)
2071 && (tcall->rprev > rxi_HardAckRate)) {
2081 ReturnToServerPool(service);
2088 MUTEX_EXIT(&rx_serverPool_lock);
2089 MUTEX_ENTER(&call->lock);
2091 if (call->flags & RX_CALL_WAIT_PROC) {
2092 call->flags &= ~RX_CALL_WAIT_PROC;
2093 rx_atomic_dec(&rx_nWaiting);
2096 if (call->state != RX_STATE_PRECALL || call->error) {
2097 MUTEX_EXIT(&call->lock);
2098 MUTEX_ENTER(&rx_serverPool_lock);
2099 ReturnToServerPool(service);
2104 if (queue_IsEmpty(&call->rq)
2105 || queue_First(&call->rq, rx_packet)->header.seq != 1)
2106 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2108 CLEAR_CALL_QUEUE_LOCK(call);
2111 /* If there are no eligible incoming calls, add this process
2112 * to the idle server queue, to wait for one */
2116 *socketp = OSI_NULLSOCKET;
2118 sq->socketp = socketp;
2119 queue_Append(&rx_idleServerQueue, sq);
2120 #ifndef AFS_AIX41_ENV
2121 rx_waitForPacket = sq;
2123 rx_waitingForPacket = sq;
2124 #endif /* AFS_AIX41_ENV */
2126 CV_WAIT(&sq->cv, &rx_serverPool_lock);
2128 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2129 MUTEX_EXIT(&rx_serverPool_lock);
2130 return (struct rx_call *)0;
2133 } while (!(call = sq->newcall)
2134 && !(socketp && *socketp != OSI_NULLSOCKET));
2135 MUTEX_EXIT(&rx_serverPool_lock);
2137 MUTEX_ENTER(&call->lock);
2143 MUTEX_ENTER(&freeSQEList_lock);
2144 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2145 rx_FreeSQEList = sq;
2146 MUTEX_EXIT(&freeSQEList_lock);
2149 clock_GetTime(&call->startTime);
2150 call->state = RX_STATE_ACTIVE;
2151 call->mode = RX_MODE_RECEIVING;
2152 #ifdef RX_KERNEL_TRACE
2153 if (ICL_SETACTIVE(afs_iclSetp)) {
2154 int glockOwner = ISAFS_GLOCK();
2157 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2158 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2165 rxi_calltrace(RX_CALL_START, call);
2166 dpf(("rx_GetCall(port=%d, service=%d) ==> call %"AFS_PTR_FMT"\n",
2167 call->conn->service->servicePort, call->conn->service->serviceId,
2170 MUTEX_EXIT(&call->lock);
2171 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
2173 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2178 #else /* RX_ENABLE_LOCKS */
2180 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
2182 struct rx_serverQueueEntry *sq;
2183 struct rx_call *call = (struct rx_call *)0, *choice2;
2184 struct rx_service *service = NULL;
2188 MUTEX_ENTER(&freeSQEList_lock);
2190 if ((sq = rx_FreeSQEList)) {
2191 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
2192 MUTEX_EXIT(&freeSQEList_lock);
2193 } else { /* otherwise allocate a new one and return that */
2194 MUTEX_EXIT(&freeSQEList_lock);
2195 sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
2196 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
2197 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
2199 MUTEX_ENTER(&sq->lock);
2201 if (cur_service != NULL) {
2202 cur_service->nRequestsRunning--;
2203 MUTEX_ENTER(&rx_quota_mutex);
2204 if (cur_service->nRequestsRunning < cur_service->minProcs)
2207 MUTEX_EXIT(&rx_quota_mutex);
2209 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
2210 struct rx_call *tcall, *ncall;
2211 /* Scan for eligible incoming calls. A call is not eligible
2212 * if the maximum number of calls for its service type are
2213 * already executing */
2214 /* One thread will process calls FCFS (to prevent starvation),
2215 * while the other threads may run ahead looking for calls which
2216 * have all their input data available immediately. This helps
2217 * keep threads from blocking, waiting for data from the client. */
2218 choice2 = (struct rx_call *)0;
2219 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
2220 service = tcall->conn->service;
2221 if (QuotaOK(service)) {
2222 MUTEX_ENTER(&rx_pthread_mutex);
2223 if (tno == rxi_fcfs_thread_num
2224 || !tcall->queue_item_header.next) {
2225 MUTEX_EXIT(&rx_pthread_mutex);
2226 /* If we're the fcfs thread, then we'll just use
2227 * this call. If we haven't been able to find an optimal
2228 * choice, and we're at the end of the list, then use a
2229 * 2d choice if one has been identified. Otherwise... */
2230 call = (choice2 ? choice2 : tcall);
2231 service = call->conn->service;
2233 MUTEX_EXIT(&rx_pthread_mutex);
2234 if (!queue_IsEmpty(&tcall->rq)) {
2235 struct rx_packet *rp;
2236 rp = queue_First(&tcall->rq, rx_packet);
2237 if (rp->header.seq == 1
2239 || (rp->header.flags & RX_LAST_PACKET))) {
2241 } else if (rxi_2dchoice && !choice2
2242 && !(tcall->flags & RX_CALL_CLEARED)
2243 && (tcall->rprev > rxi_HardAckRate)) {
2257 /* we can't schedule a call if there's no data!!! */
2258 /* send an ack if there's no data, if we're missing the
2259 * first packet, or we're missing something between first
2260 * and last -- there's a "hole" in the incoming data. */
2261 if (queue_IsEmpty(&call->rq)
2262 || queue_First(&call->rq, rx_packet)->header.seq != 1
2263 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
2264 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
2266 call->flags &= (~RX_CALL_WAIT_PROC);
2267 service->nRequestsRunning++;
2268 /* just started call in minProcs pool, need fewer to maintain
2270 MUTEX_ENTER(&rx_quota_mutex);
2271 if (service->nRequestsRunning <= service->minProcs)
2274 MUTEX_EXIT(&rx_quota_mutex);
2275 rx_atomic_dec(&rx_nWaiting);
2276 /* MUTEX_EXIT(&call->lock); */
2278 /* If there are no eligible incoming calls, add this process
2279 * to the idle server queue, to wait for one */
2282 *socketp = OSI_NULLSOCKET;
2284 sq->socketp = socketp;
2285 queue_Append(&rx_idleServerQueue, sq);
2289 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
2291 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
2292 return (struct rx_call *)0;
2295 } while (!(call = sq->newcall)
2296 && !(socketp && *socketp != OSI_NULLSOCKET));
2298 MUTEX_EXIT(&sq->lock);
2300 MUTEX_ENTER(&freeSQEList_lock);
2301 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
2302 rx_FreeSQEList = sq;
2303 MUTEX_EXIT(&freeSQEList_lock);
2306 clock_GetTime(&call->startTime);
2307 call->state = RX_STATE_ACTIVE;
2308 call->mode = RX_MODE_RECEIVING;
2309 #ifdef RX_KERNEL_TRACE
2310 if (ICL_SETACTIVE(afs_iclSetp)) {
2311 int glockOwner = ISAFS_GLOCK();
2314 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
2315 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2322 rxi_calltrace(RX_CALL_START, call);
2323 dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
2324 call->conn->service->servicePort, call->conn->service->serviceId,
2327 dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
2334 #endif /* RX_ENABLE_LOCKS */
2338 /* Establish a procedure to be called when a packet arrives for a
2339 * call. This routine will be called at most once after each call,
2340 * and will also be called if there is an error condition on the or
2341 * the call is complete. Used by multi rx to build a selection
2342 * function which determines which of several calls is likely to be a
2343 * good one to read from.
2344 * NOTE: the way this is currently implemented it is probably only a
2345 * good idea to (1) use it immediately after a newcall (clients only)
2346 * and (2) only use it once. Other uses currently void your warranty
2349 rx_SetArrivalProc(struct rx_call *call,
2350 void (*proc) (struct rx_call * call,
2353 void * handle, int arg)
2355 call->arrivalProc = proc;
2356 call->arrivalProcHandle = handle;
2357 call->arrivalProcArg = arg;
2360 /* Call is finished (possibly prematurely). Return rc to the peer, if
2361 * appropriate, and return the final error code from the conversation
2365 rx_EndCall(struct rx_call *call, afs_int32 rc)
2367 struct rx_connection *conn = call->conn;
2371 dpf(("rx_EndCall(call %"AFS_PTR_FMT" rc %d error %d abortCode %d)\n",
2372 call, rc, call->error, call->abortCode));
2375 MUTEX_ENTER(&call->lock);
2377 if (rc == 0 && call->error == 0) {
2378 call->abortCode = 0;
2379 call->abortCount = 0;
2382 call->arrivalProc = (void (*)())0;
2383 if (rc && call->error == 0) {
2384 rxi_CallError(call, rc);
2385 call->mode = RX_MODE_ERROR;
2386 /* Send an abort message to the peer if this error code has
2387 * only just been set. If it was set previously, assume the
2388 * peer has already been sent the error code or will request it
2390 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
2392 if (conn->type == RX_SERVER_CONNECTION) {
2393 /* Make sure reply or at least dummy reply is sent */
2394 if (call->mode == RX_MODE_RECEIVING) {
2395 MUTEX_EXIT(&call->lock);
2396 rxi_WriteProc(call, 0, 0);
2397 MUTEX_ENTER(&call->lock);
2399 if (call->mode == RX_MODE_SENDING) {
2400 MUTEX_EXIT(&call->lock);
2401 rxi_FlushWrite(call);
2402 MUTEX_ENTER(&call->lock);
2404 rxi_calltrace(RX_CALL_END, call);
2405 /* Call goes to hold state until reply packets are acknowledged */
2406 if (call->tfirst + call->nSoftAcked < call->tnext) {
2407 call->state = RX_STATE_HOLD;
2409 call->state = RX_STATE_DALLY;
2410 rxi_ClearTransmitQueue(call, 0);
2411 rxi_rto_cancel(call);
2412 rxevent_Cancel(&call->keepAliveEvent, call,
2413 RX_CALL_REFCOUNT_ALIVE);
2415 } else { /* Client connection */
2417 /* Make sure server receives input packets, in the case where
2418 * no reply arguments are expected */
2419 if ((call->mode == RX_MODE_SENDING)
2420 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2421 MUTEX_EXIT(&call->lock);
2422 (void)rxi_ReadProc(call, &dummy, 1);
2423 MUTEX_ENTER(&call->lock);
2426 /* If we had an outstanding delayed ack, be nice to the server
2427 * and force-send it now.
2429 if (call->delayedAckEvent) {
2430 rxevent_Cancel(&call->delayedAckEvent, call,
2431 RX_CALL_REFCOUNT_DELAY);
2432 rxi_SendDelayedAck(NULL, call, NULL, 0);
2435 /* We need to release the call lock since it's lower than the
2436 * conn_call_lock and we don't want to hold the conn_call_lock
2437 * over the rx_ReadProc call. The conn_call_lock needs to be held
2438 * here for the case where rx_NewCall is perusing the calls on
2439 * the connection structure. We don't want to signal until
2440 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2441 * have checked this call, found it active and by the time it
2442 * goes to sleep, will have missed the signal.
2444 MUTEX_EXIT(&call->lock);
2445 MUTEX_ENTER(&conn->conn_call_lock);
2446 MUTEX_ENTER(&call->lock);
2448 if (!(call->flags & RX_CALL_PEER_BUSY)) {
2449 conn->lastBusy[call->channel] = 0;
2452 MUTEX_ENTER(&conn->conn_data_lock);
2453 conn->flags |= RX_CONN_BUSY;
2454 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2455 MUTEX_EXIT(&conn->conn_data_lock);
2456 #ifdef RX_ENABLE_LOCKS
2457 CV_BROADCAST(&conn->conn_call_cv);
2462 #ifdef RX_ENABLE_LOCKS
2464 MUTEX_EXIT(&conn->conn_data_lock);
2466 #endif /* RX_ENABLE_LOCKS */
2467 call->state = RX_STATE_DALLY;
2469 error = call->error;
2471 /* currentPacket, nLeft, and NFree must be zeroed here, because
2472 * ResetCall cannot: ResetCall may be called at splnet(), in the
2473 * kernel version, and may interrupt the macros rx_Read or
2474 * rx_Write, which run at normal priority for efficiency. */
2475 if (call->currentPacket) {
2476 #ifdef RX_TRACK_PACKETS
2477 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2479 rxi_FreePacket(call->currentPacket);
2480 call->currentPacket = (struct rx_packet *)0;
2483 call->nLeft = call->nFree = call->curlen = 0;
2485 /* Free any packets from the last call to ReadvProc/WritevProc */
2486 #ifdef RXDEBUG_PACKET
2488 #endif /* RXDEBUG_PACKET */
2489 rxi_FreePackets(0, &call->iovq);
2490 MUTEX_EXIT(&call->lock);
2492 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2493 if (conn->type == RX_CLIENT_CONNECTION) {
2494 MUTEX_ENTER(&conn->conn_data_lock);
2495 conn->flags &= ~RX_CONN_BUSY;
2496 MUTEX_EXIT(&conn->conn_data_lock);
2497 MUTEX_EXIT(&conn->conn_call_lock);
2501 * Map errors to the local host's errno.h format.
2503 error = ntoh_syserr_conv(error);
2507 #if !defined(KERNEL)
2509 /* Call this routine when shutting down a server or client (especially
2510 * clients). This will allow Rx to gracefully garbage collect server
2511 * connections, and reduce the number of retries that a server might
2512 * make to a dead client.
2513 * This is not quite right, since some calls may still be ongoing and
2514 * we can't lock them to destroy them. */
2518 struct rx_connection **conn_ptr, **conn_end;
2522 if (rxinit_status == 1) {
2524 return; /* Already shutdown. */
2526 rxi_DeleteCachedConnections();
2527 if (rx_connHashTable) {
2528 MUTEX_ENTER(&rx_connHashTable_lock);
2529 for (conn_ptr = &rx_connHashTable[0], conn_end =
2530 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2532 struct rx_connection *conn, *next;
2533 for (conn = *conn_ptr; conn; conn = next) {
2535 if (conn->type == RX_CLIENT_CONNECTION) {
2536 MUTEX_ENTER(&rx_refcnt_mutex);
2538 MUTEX_EXIT(&rx_refcnt_mutex);
2539 #ifdef RX_ENABLE_LOCKS
2540 rxi_DestroyConnectionNoLock(conn);
2541 #else /* RX_ENABLE_LOCKS */
2542 rxi_DestroyConnection(conn);
2543 #endif /* RX_ENABLE_LOCKS */
2547 #ifdef RX_ENABLE_LOCKS
2548 while (rx_connCleanup_list) {
2549 struct rx_connection *conn;
2550 conn = rx_connCleanup_list;
2551 rx_connCleanup_list = rx_connCleanup_list->next;
2552 MUTEX_EXIT(&rx_connHashTable_lock);
2553 rxi_CleanupConnection(conn);
2554 MUTEX_ENTER(&rx_connHashTable_lock);
2556 MUTEX_EXIT(&rx_connHashTable_lock);
2557 #endif /* RX_ENABLE_LOCKS */
2562 afs_winsockCleanup();
2570 /* if we wakeup packet waiter too often, can get in loop with two
2571 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2573 rxi_PacketsUnWait(void)
2575 if (!rx_waitingForPackets) {
2579 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2580 return; /* still over quota */
2583 rx_waitingForPackets = 0;
2584 #ifdef RX_ENABLE_LOCKS
2585 CV_BROADCAST(&rx_waitingForPackets_cv);
2587 osi_rxWakeup(&rx_waitingForPackets);
2593 /* ------------------Internal interfaces------------------------- */
2595 /* Return this process's service structure for the
2596 * specified socket and service */
2597 static struct rx_service *
2598 rxi_FindService(osi_socket socket, u_short serviceId)
2600 struct rx_service **sp;
2601 for (sp = &rx_services[0]; *sp; sp++) {
2602 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2608 #ifdef RXDEBUG_PACKET
2609 #ifdef KDUMP_RX_LOCK
2610 static struct rx_call_rx_lock *rx_allCallsp = 0;
2612 static struct rx_call *rx_allCallsp = 0;
2614 #endif /* RXDEBUG_PACKET */
2616 /* Allocate a call structure, for the indicated channel of the
2617 * supplied connection. The mode and state of the call must be set by
2618 * the caller. Returns the call with mutex locked. */
2619 static struct rx_call *
2620 rxi_NewCall(struct rx_connection *conn, int channel)
2622 struct rx_call *call;
2623 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2624 struct rx_call *cp; /* Call pointer temp */
2625 struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2626 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2628 dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", channel %d)\n", conn, channel));
2630 /* Grab an existing call structure, or allocate a new one.
2631 * Existing call structures are assumed to have been left reset by
2633 MUTEX_ENTER(&rx_freeCallQueue_lock);
2635 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2637 * EXCEPT that the TQ might not yet be cleared out.
2638 * Skip over those with in-use TQs.
2641 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2642 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2648 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2649 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2650 call = queue_First(&rx_freeCallQueue, rx_call);
2651 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2653 if (rx_stats_active)
2654 rx_atomic_dec(&rx_stats.nFreeCallStructs);
2655 MUTEX_EXIT(&rx_freeCallQueue_lock);
2656 MUTEX_ENTER(&call->lock);
2657 CLEAR_CALL_QUEUE_LOCK(call);
2658 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2659 /* Now, if TQ wasn't cleared earlier, do it now. */
2660 rxi_WaitforTQBusy(call);
2661 if (call->flags & RX_CALL_TQ_CLEARME) {
2662 rxi_ClearTransmitQueue(call, 1);
2663 /*queue_Init(&call->tq);*/
2665 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2666 /* Bind the call to its connection structure */
2668 rxi_ResetCall(call, 1);
2671 call = rxi_Alloc(sizeof(struct rx_call));
2672 #ifdef RXDEBUG_PACKET
2673 call->allNextp = rx_allCallsp;
2674 rx_allCallsp = call;
2676 rx_atomic_inc_and_read(&rx_stats.nCallStructs);
2677 #else /* RXDEBUG_PACKET */
2678 rx_atomic_inc(&rx_stats.nCallStructs);
2679 #endif /* RXDEBUG_PACKET */
2681 MUTEX_EXIT(&rx_freeCallQueue_lock);
2682 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2683 MUTEX_ENTER(&call->lock);
2684 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2685 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2686 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2688 /* Initialize once-only items */
2689 queue_Init(&call->tq);
2690 queue_Init(&call->rq);
2691 queue_Init(&call->iovq);
2692 #ifdef RXDEBUG_PACKET
2693 call->rqc = call->tqc = call->iovqc = 0;
2694 #endif /* RXDEBUG_PACKET */
2695 /* Bind the call to its connection structure (prereq for reset) */
2697 rxi_ResetCall(call, 1);
2699 call->channel = channel;
2700 call->callNumber = &conn->callNumber[channel];
2701 call->rwind = conn->rwind[channel];
2702 call->twind = conn->twind[channel];
2703 /* Note that the next expected call number is retained (in
2704 * conn->callNumber[i]), even if we reallocate the call structure
2706 conn->call[channel] = call;
2707 /* if the channel's never been used (== 0), we should start at 1, otherwise
2708 * the call number is valid from the last time this channel was used */
2709 if (*call->callNumber == 0)
2710 *call->callNumber = 1;
2715 /* A call has been inactive long enough that so we can throw away
2716 * state, including the call structure, which is placed on the call
2719 * call->lock amd rx_refcnt_mutex are held upon entry.
2720 * haveCTLock is set when called from rxi_ReapConnections.
2722 * return 1 if the call is freed, 0 if not.
2725 rxi_FreeCall(struct rx_call *call, int haveCTLock)
2727 int channel = call->channel;
2728 struct rx_connection *conn = call->conn;
2729 u_char state = call->state;
2732 * We are setting the state to RX_STATE_RESET to
2733 * ensure that no one else will attempt to use this
2734 * call once we drop the refcnt lock. We must drop
2735 * the refcnt lock before calling rxi_ResetCall
2736 * because it cannot be held across acquiring the
2737 * freepktQ lock. NewCall does the same.
2739 call->state = RX_STATE_RESET;
2740 MUTEX_EXIT(&rx_refcnt_mutex);
2741 rxi_ResetCall(call, 0);
2743 if (MUTEX_TRYENTER(&conn->conn_call_lock))
2745 if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
2746 (*call->callNumber)++;
2748 if (call->conn->call[channel] == call)
2749 call->conn->call[channel] = 0;
2750 MUTEX_EXIT(&conn->conn_call_lock);
2753 * We couldn't obtain the conn_call_lock so we can't
2754 * disconnect the call from the connection. Set the
2755 * call state to dally so that the call can be reused.
2757 MUTEX_ENTER(&rx_refcnt_mutex);
2758 call->state = RX_STATE_DALLY;
2762 MUTEX_ENTER(&rx_freeCallQueue_lock);
2763 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2764 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2765 /* A call may be free even though its transmit queue is still in use.
2766 * Since we search the call list from head to tail, put busy calls at
2767 * the head of the list, and idle calls at the tail.
2769 if (call->flags & RX_CALL_TQ_BUSY)
2770 queue_Prepend(&rx_freeCallQueue, call);
2772 queue_Append(&rx_freeCallQueue, call);
2773 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2774 queue_Append(&rx_freeCallQueue, call);
2775 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2776 if (rx_stats_active)
2777 rx_atomic_inc(&rx_stats.nFreeCallStructs);
2778 MUTEX_EXIT(&rx_freeCallQueue_lock);
2780 /* Destroy the connection if it was previously slated for
2781 * destruction, i.e. the Rx client code previously called
2782 * rx_DestroyConnection (client connections), or
2783 * rxi_ReapConnections called the same routine (server
2784 * connections). Only do this, however, if there are no
2785 * outstanding calls. Note that for fine grain locking, there appears
2786 * to be a deadlock in that rxi_FreeCall has a call locked and
2787 * DestroyConnectionNoLock locks each call in the conn. But note a
2788 * few lines up where we have removed this call from the conn.
2789 * If someone else destroys a connection, they either have no
2790 * call lock held or are going through this section of code.
2792 MUTEX_ENTER(&conn->conn_data_lock);
2793 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2794 MUTEX_ENTER(&rx_refcnt_mutex);
2796 MUTEX_EXIT(&rx_refcnt_mutex);
2797 MUTEX_EXIT(&conn->conn_data_lock);
2798 #ifdef RX_ENABLE_LOCKS
2800 rxi_DestroyConnectionNoLock(conn);
2802 rxi_DestroyConnection(conn);
2803 #else /* RX_ENABLE_LOCKS */
2804 rxi_DestroyConnection(conn);
2805 #endif /* RX_ENABLE_LOCKS */
2807 MUTEX_EXIT(&conn->conn_data_lock);
2809 MUTEX_ENTER(&rx_refcnt_mutex);
2813 rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
2814 rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);
2817 rxi_Alloc(size_t size)
2821 if (rx_stats_active) {
2822 rx_atomic_add(&rxi_Allocsize, (int) size);
2823 rx_atomic_inc(&rxi_Alloccnt);
2827 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2828 afs_osi_Alloc_NoSleep(size);
2833 osi_Panic("rxi_Alloc error");
2839 rxi_Free(void *addr, size_t size)
2841 if (rx_stats_active) {
2842 rx_atomic_sub(&rxi_Allocsize, (int) size);
2843 rx_atomic_dec(&rxi_Alloccnt);
2845 osi_Free(addr, size);
2849 rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
2851 struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
2852 struct rx_peer *next = NULL;
2856 MUTEX_ENTER(&rx_peerHashTable_lock);
2858 peer_ptr = &rx_peerHashTable[0];
2859 peer_end = &rx_peerHashTable[rx_hashTableSize];
2862 for ( ; peer_ptr < peer_end; peer_ptr++) {
2865 for ( ; peer; peer = next) {
2867 if (host == peer->host)
2872 hashIndex = PEER_HASH(host, port);
2873 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2874 if ((peer->host == host) && (peer->port == port))
2879 MUTEX_ENTER(&rx_peerHashTable_lock);
2884 MUTEX_EXIT(&rx_peerHashTable_lock);
2886 MUTEX_ENTER(&peer->peer_lock);
2887 /* We don't handle dropping below min, so don't */
2888 mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
2889 peer->ifMTU=MIN(mtu, peer->ifMTU);
2890 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2891 /* if we tweaked this down, need to tune our peer MTU too */
2892 peer->MTU = MIN(peer->MTU, peer->natMTU);
2893 /* if we discovered a sub-1500 mtu, degrade */
2894 if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
2895 peer->maxDgramPackets = 1;
2896 /* We no longer have valid peer packet information */
2897 if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
2898 peer->maxPacketSize = 0;
2899 MUTEX_EXIT(&peer->peer_lock);
2901 MUTEX_ENTER(&rx_peerHashTable_lock);
2903 if (host && !port) {
2905 /* pick up where we left off */
2909 MUTEX_EXIT(&rx_peerHashTable_lock);
2912 /* Find the peer process represented by the supplied (host,port)
2913 * combination. If there is no appropriate active peer structure, a
2914 * new one will be allocated and initialized
2915 * The origPeer, if set, is a pointer to a peer structure on which the
2916 * refcount will be be decremented. This is used to replace the peer
2917 * structure hanging off a connection structure */
2919 rxi_FindPeer(afs_uint32 host, u_short port,
2920 struct rx_peer *origPeer, int create)
2924 hashIndex = PEER_HASH(host, port);
2925 MUTEX_ENTER(&rx_peerHashTable_lock);
2926 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2927 if ((pp->host == host) && (pp->port == port))
2932 pp = rxi_AllocPeer(); /* This bzero's *pp */
2933 pp->host = host; /* set here or in InitPeerParams is zero */
2935 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2936 queue_Init(&pp->rpcStats);
2937 pp->next = rx_peerHashTable[hashIndex];
2938 rx_peerHashTable[hashIndex] = pp;
2939 rxi_InitPeerParams(pp);
2940 if (rx_stats_active)
2941 rx_atomic_inc(&rx_stats.nPeerStructs);
2948 origPeer->refCount--;
2949 MUTEX_EXIT(&rx_peerHashTable_lock);
2954 /* Find the connection at (host, port) started at epoch, and with the
2955 * given connection id. Creates the server connection if necessary.
2956 * The type specifies whether a client connection or a server
2957 * connection is desired. In both cases, (host, port) specify the
2958 * peer's (host, pair) pair. Client connections are not made
2959 * automatically by this routine. The parameter socket gives the
2960 * socket descriptor on which the packet was received. This is used,
2961 * in the case of server connections, to check that *new* connections
2962 * come via a valid (port, serviceId). Finally, the securityIndex
2963 * parameter must match the existing index for the connection. If a
2964 * server connection is created, it will be created using the supplied
2965 * index, if the index is valid for this service */
2966 static struct rx_connection *
2967 rxi_FindConnection(osi_socket socket, afs_uint32 host,
2968 u_short port, u_short serviceId, afs_uint32 cid,
2969 afs_uint32 epoch, int type, u_int securityIndex)
2971 int hashindex, flag, i;
2972 struct rx_connection *conn;
2973 hashindex = CONN_HASH(host, port, cid, epoch, type);
2974 MUTEX_ENTER(&rx_connHashTable_lock);
2975 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2976 rx_connHashTable[hashindex],
2979 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2980 && (epoch == conn->epoch)) {
2981 struct rx_peer *pp = conn->peer;
2982 if (securityIndex != conn->securityIndex) {
2983 /* this isn't supposed to happen, but someone could forge a packet
2984 * like this, and there seems to be some CM bug that makes this
2985 * happen from time to time -- in which case, the fileserver
2987 MUTEX_EXIT(&rx_connHashTable_lock);
2988 return (struct rx_connection *)0;
2990 if (pp->host == host && pp->port == port)
2992 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2994 /* So what happens when it's a callback connection? */
2995 if ( /*type == RX_CLIENT_CONNECTION && */
2996 (conn->epoch & 0x80000000))
3000 /* the connection rxLastConn that was used the last time is not the
3001 ** one we are looking for now. Hence, start searching in the hash */
3003 conn = rx_connHashTable[hashindex];
3008 struct rx_service *service;
3009 if (type == RX_CLIENT_CONNECTION) {
3010 MUTEX_EXIT(&rx_connHashTable_lock);
3011 return (struct rx_connection *)0;
3013 service = rxi_FindService(socket, serviceId);
3014 if (!service || (securityIndex >= service->nSecurityObjects)
3015 || (service->securityObjects[securityIndex] == 0)) {
3016 MUTEX_EXIT(&rx_connHashTable_lock);
3017 return (struct rx_connection *)0;
3019 conn = rxi_AllocConnection(); /* This bzero's the connection */
3020 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
3021 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
3022 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
3023 conn->next = rx_connHashTable[hashindex];
3024 rx_connHashTable[hashindex] = conn;
3025 conn->peer = rxi_FindPeer(host, port, 0, 1);
3026 conn->type = RX_SERVER_CONNECTION;
3027 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
3028 conn->epoch = epoch;
3029 conn->cid = cid & RX_CIDMASK;
3030 conn->ackRate = RX_FAST_ACK_RATE;
3031 conn->service = service;
3032 conn->serviceId = serviceId;
3033 conn->securityIndex = securityIndex;
3034 conn->securityObject = service->securityObjects[securityIndex];
3035 conn->nSpecific = 0;
3036 conn->specific = NULL;
3037 rx_SetConnDeadTime(conn, service->connDeadTime);
3038 conn->idleDeadTime = service->idleDeadTime;
3039 conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
3040 for (i = 0; i < RX_MAXCALLS; i++) {
3041 conn->twind[i] = rx_initSendWindow;
3042 conn->rwind[i] = rx_initReceiveWindow;
3044 /* Notify security object of the new connection */
3045 RXS_NewConnection(conn->securityObject, conn);
3046 /* XXXX Connection timeout? */
3047 if (service->newConnProc)
3048 (*service->newConnProc) (conn);
3049 if (rx_stats_active)
3050 rx_atomic_inc(&rx_stats.nServerConns);
3053 MUTEX_ENTER(&rx_refcnt_mutex);
3055 MUTEX_EXIT(&rx_refcnt_mutex);
3057 rxLastConn = conn; /* store this connection as the last conn used */
3058 MUTEX_EXIT(&rx_connHashTable_lock);
3063 * Timeout a call on a busy call channel if appropriate.
3065 * @param[in] call The busy call.
3067 * @pre 'call' is marked as busy (namely,
3068 * call->conn->lastBusy[call->channel] != 0)
3070 * @pre call->lock is held
3071 * @pre rxi_busyChannelError is nonzero
3073 * @note call->lock is dropped and reacquired
3076 rxi_CheckBusy(struct rx_call *call)
3078 struct rx_connection *conn = call->conn;
3079 int channel = call->channel;
3080 int freechannel = 0;
3082 afs_uint32 callNumber;
3084 MUTEX_EXIT(&call->lock);
3086 MUTEX_ENTER(&conn->conn_call_lock);
3087 callNumber = *call->callNumber;
3089 /* Are there any other call slots on this conn that we should try? Look for
3090 * slots that are empty and are either non-busy, or were marked as busy
3091 * longer than conn->secondsUntilDead seconds before this call started. */
3093 for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
3095 /* only look at channels that aren't us */
3099 if (conn->lastBusy[i]) {
3100 /* if this channel looked busy too recently, don't look at it */
3101 if (conn->lastBusy[i] >= call->startTime.sec) {
3104 if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
3109 if (conn->call[i]) {
3110 struct rx_call *tcall = conn->call[i];
3111 MUTEX_ENTER(&tcall->lock);
3112 if (tcall->state == RX_STATE_DALLY) {
3115 MUTEX_EXIT(&tcall->lock);
3121 MUTEX_ENTER(&call->lock);
3123 /* Since the call->lock and conn->conn_call_lock have been released it is
3124 * possible that (1) the call may no longer be busy and/or (2) the call may
3125 * have been reused by another waiting thread. Therefore, we must confirm
3126 * that the call state has not changed when deciding whether or not to
3127 * force this application thread to retry by forcing a Timeout error. */
3129 if (freechannel && *call->callNumber == callNumber &&
3130 (call->flags & RX_CALL_PEER_BUSY)) {
3131 /* Since 'freechannel' is set, there exists another channel in this
3132 * rx_conn that the application thread might be able to use. We know
3133 * that we have the correct call since callNumber is unchanged, and we
3134 * know that the call is still busy. So, set the call error state to
3135 * rxi_busyChannelError so the application can retry the request,
3136 * presumably on a less-busy call channel. */
3138 rxi_CallError(call, RX_CALL_BUSY);
3140 MUTEX_EXIT(&conn->conn_call_lock);
3143 /* There are two packet tracing routines available for testing and monitoring
3144 * Rx. One is called just after every packet is received and the other is
3145 * called just before every packet is sent. Received packets, have had their
3146 * headers decoded, and packets to be sent have not yet had their headers
3147 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
3148 * containing the network address. Both can be modified. The return value, if
3149 * non-zero, indicates that the packet should be dropped. */
3151 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
3152 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
3154 /* A packet has been received off the interface. Np is the packet, socket is
3155 * the socket number it was received from (useful in determining which service
3156 * this packet corresponds to), and (host, port) reflect the host,port of the
3157 * sender. This call returns the packet to the caller if it is finished with
3158 * it, rather than de-allocating it, just as a small performance hack */
3161 rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
3162 afs_uint32 host, u_short port, int *tnop,
3163 struct rx_call **newcallp)
3165 struct rx_call *call;
3166 struct rx_connection *conn;
3168 afs_uint32 currentCallNumber;
3173 struct rx_packet *tnp;
3176 /* We don't print out the packet until now because (1) the time may not be
3177 * accurate enough until now in the lwp implementation (rx_Listener only gets
3178 * the time after the packet is read) and (2) from a protocol point of view,
3179 * this is the first time the packet has been seen */
3180 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
3181 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
3182 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %"AFS_PTR_FMT"\n",
3183 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
3184 np->header.epoch, np->header.cid, np->header.callNumber,
3185 np->header.seq, np->header.flags, np));
3188 /* Account for connectionless packets */
3189 if (rx_stats_active &&
3190 ((np->header.type == RX_PACKET_TYPE_VERSION) ||
3191 (np->header.type == RX_PACKET_TYPE_DEBUG))) {
3192 struct rx_peer *peer;
3194 /* Try to look up the peer structure, but don't create one */
3195 peer = rxi_FindPeer(host, port, 0, 0);
3197 /* Since this may not be associated with a connection, it may have
3198 * no refCount, meaning we could race with ReapConnections
3201 if (peer && (peer->refCount > 0)) {
3202 MUTEX_ENTER(&peer->peer_lock);
3203 peer->bytesReceived += np->length;
3204 MUTEX_EXIT(&peer->peer_lock);
3208 if (np->header.type == RX_PACKET_TYPE_VERSION) {
3209 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
3212 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
3213 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
3216 /* If an input tracer function is defined, call it with the packet and
3217 * network address. Note this function may modify its arguments. */
3218 if (rx_justReceived) {
3219 struct sockaddr_in addr;
3221 addr.sin_family = AF_INET;
3222 addr.sin_port = port;
3223 addr.sin_addr.s_addr = host;
3224 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
3225 addr.sin_len = sizeof(addr);
3226 #endif /* AFS_OSF_ENV */
3227 drop = (*rx_justReceived) (np, &addr);
3228 /* drop packet if return value is non-zero */
3231 port = addr.sin_port; /* in case fcn changed addr */
3232 host = addr.sin_addr.s_addr;
3236 /* If packet was not sent by the client, then *we* must be the client */
3237 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
3238 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
3240 /* Find the connection (or fabricate one, if we're the server & if
3241 * necessary) associated with this packet */
3243 rxi_FindConnection(socket, host, port, np->header.serviceId,
3244 np->header.cid, np->header.epoch, type,
3245 np->header.securityIndex);
3247 /* To avoid having 2 connections just abort at each other,
3248 don't abort an abort. */
3249 if (!conn && (np->header.type != RX_PACKET_TYPE_ABORT)) {
3250 rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
3255 /* If we're doing statistics, then account for the incoming packet */
3256 if (rx_stats_active) {
3257 MUTEX_ENTER(&conn->peer->peer_lock);
3258 conn->peer->bytesReceived += np->length;
3259 MUTEX_EXIT(&conn->peer->peer_lock);
3262 /* If the connection is in an error state, send an abort packet and ignore
3263 * the incoming packet */
3265 /* Don't respond to an abort packet--we don't want loops! */
3266 MUTEX_ENTER(&conn->conn_data_lock);
3267 if (np->header.type != RX_PACKET_TYPE_ABORT)
3268 np = rxi_SendConnectionAbort(conn, np, 1, 0);
3269 putConnection(conn);
3270 MUTEX_EXIT(&conn->conn_data_lock);
3274 /* Check for connection-only requests (i.e. not call specific). */
3275 if (np->header.callNumber == 0) {
3276 switch (np->header.type) {
3277 case RX_PACKET_TYPE_ABORT: {
3278 /* What if the supplied error is zero? */
3279 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
3280 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
3281 rxi_ConnectionError(conn, errcode);
3282 putConnection(conn);
3285 case RX_PACKET_TYPE_CHALLENGE:
3286 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
3287 putConnection(conn);
3289 case RX_PACKET_TYPE_RESPONSE:
3290 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
3291 putConnection(conn);
3293 case RX_PACKET_TYPE_PARAMS:
3294 case RX_PACKET_TYPE_PARAMS + 1:
3295 case RX_PACKET_TYPE_PARAMS + 2:
3296 /* ignore these packet types for now */
3297 putConnection(conn);
3301 /* Should not reach here, unless the peer is broken: send an
3303 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
3304 MUTEX_ENTER(&conn->conn_data_lock);
3305 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
3306 putConnection(conn);
3307 MUTEX_EXIT(&conn->conn_data_lock);
3312 channel = np->header.cid & RX_CHANNELMASK;
3313 MUTEX_ENTER(&conn->conn_call_lock);
3314 call = conn->call[channel];
3317 MUTEX_ENTER(&call->lock);
3318 currentCallNumber = conn->callNumber[channel];
3319 MUTEX_EXIT(&conn->conn_call_lock);
3320 } else if (type == RX_SERVER_CONNECTION) { /* No call allocated */
3321 call = conn->call[channel];
3323 MUTEX_ENTER(&call->lock);
3324 currentCallNumber = conn->callNumber[channel];
3325 MUTEX_EXIT(&conn->conn_call_lock);
3327 call = rxi_NewCall(conn, channel); /* returns locked call */
3328 *call->callNumber = currentCallNumber = np->header.callNumber;
3329 MUTEX_EXIT(&conn->conn_call_lock);
3331 if (np->header.callNumber == 0)
3332 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3333 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3334 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3335 np->header.flags, np, np->length));
3337 call->state = RX_STATE_PRECALL;
3338 clock_GetTime(&call->queueTime);
3339 call->bytesSent = 0;
3340 call->bytesRcvd = 0;
3342 * If the number of queued calls exceeds the overload
3343 * threshold then abort this call.
3345 if ((rx_BusyThreshold > 0) &&
3346 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3347 struct rx_packet *tp;
3349 rxi_CallError(call, rx_BusyError);
3350 tp = rxi_SendCallAbort(call, np, 1, 0);
3351 MUTEX_EXIT(&call->lock);
3352 putConnection(conn);
3353 if (rx_stats_active)
3354 rx_atomic_inc(&rx_stats.nBusies);
3357 rxi_KeepAliveOn(call);
3359 } else { /* RX_CLIENT_CONNECTION and No call allocated */
3360 /* This packet can't be for this call. If the new call address is
3361 * 0 then no call is running on this channel. If there is a call
3362 * then, since this is a client connection we're getting data for
3363 * it must be for the previous call.
3365 MUTEX_EXIT(&conn->conn_call_lock);
3366 if (rx_stats_active)
3367 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3368 putConnection(conn);
3372 /* There is a non-NULL locked call at this point */
3373 if (type == RX_SERVER_CONNECTION) { /* We're the server */
3374 if (np->header.callNumber < currentCallNumber) {
3375 MUTEX_EXIT(&call->lock);
3376 if (rx_stats_active)
3377 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3378 putConnection(conn);
3380 } else if (np->header.callNumber != currentCallNumber) {
3381 /* Wait until the transmit queue is idle before deciding
3382 * whether to reset the current call. Chances are that the
3383 * call will be in ether DALLY or HOLD state once the TQ_BUSY
3386 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3387 if (call->state == RX_STATE_ACTIVE) {
3388 rxi_WaitforTQBusy(call);
3390 * If we entered error state while waiting,
3391 * must call rxi_CallError to permit rxi_ResetCall
3392 * to processed when the tqWaiter count hits zero.
3395 rxi_CallError(call, call->error);
3396 MUTEX_EXIT(&call->lock);
3397 putConnection(conn);
3401 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3402 /* If the new call cannot be taken right now send a busy and set
3403 * the error condition in this call, so that it terminates as
3404 * quickly as possible */
3405 if (call->state == RX_STATE_ACTIVE) {
3406 struct rx_packet *tp;
3408 rxi_CallError(call, RX_CALL_DEAD);
3409 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
3411 MUTEX_EXIT(&call->lock);
3412 putConnection(conn);
3415 rxi_ResetCall(call, 0);
3417 * The conn_call_lock is not held but no one else should be
3418 * using this call channel while we are processing this incoming
3419 * packet. This assignment should be safe.
3421 *call->callNumber = np->header.callNumber;
3423 if (np->header.callNumber == 0)
3424 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" len %d\n",
3425 np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
3426 np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
3427 np->header.flags, np, np->length));
3429 call->state = RX_STATE_PRECALL;
3430 clock_GetTime(&call->queueTime);
3431 call->bytesSent = 0;
3432 call->bytesRcvd = 0;
3434 * If the number of queued calls exceeds the overload
3435 * threshold then abort this call.
3437 if ((rx_BusyThreshold > 0) &&
3438 (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
3439 struct rx_packet *tp;
3441 rxi_CallError(call, rx_BusyError);
3442 tp = rxi_SendCallAbort(call, np, 1, 0);
3443 MUTEX_EXIT(&call->lock);
3444 putConnection(conn);
3445 if (rx_stats_active)
3446 rx_atomic_inc(&rx_stats.nBusies);
3449 rxi_KeepAliveOn(call);
3451 /* Continuing call; do nothing here. */
3453 } else { /* we're the client */
3454 /* Ignore all incoming acknowledgements for calls in DALLY state */
3455 if ((call->state == RX_STATE_DALLY)
3456 && (np->header.type == RX_PACKET_TYPE_ACK)) {
3457 if (rx_stats_active)
3458 rx_atomic_inc(&rx_stats.ignorePacketDally);
3459 MUTEX_EXIT(&call->lock);
3460 putConnection(conn);
3464 /* Ignore anything that's not relevant to the current call. If there
3465 * isn't a current call, then no packet is relevant. */
3466 if (np->header.callNumber != currentCallNumber) {
3467 if (rx_stats_active)
3468 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3469 MUTEX_EXIT(&call->lock);
3470 putConnection(conn);
3473 /* If the service security object index stamped in the packet does not
3474 * match the connection's security index, ignore the packet */
3475 if (np->header.securityIndex != conn->securityIndex) {
3476 MUTEX_EXIT(&call->lock);
3477 putConnection(conn);
3481 /* If we're receiving the response, then all transmit packets are
3482 * implicitly acknowledged. Get rid of them. */
3483 if (np->header.type == RX_PACKET_TYPE_DATA) {
3484 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3485 /* XXX Hack. Because we must release the global rx lock when
3486 * sending packets (osi_NetSend) we drop all acks while we're
3487 * traversing the tq in rxi_Start sending packets out because
3488 * packets may move to the freePacketQueue as result of being here!
3489 * So we drop these packets until we're safely out of the
3490 * traversing. Really ugly!
3491 * For fine grain RX locking, we set the acked field in the
3492 * packets and let rxi_Start remove them from the transmit queue.
3494 if (call->flags & RX_CALL_TQ_BUSY) {
3495 #ifdef RX_ENABLE_LOCKS
3496 rxi_SetAcksInTransmitQueue(call);
3498 putConnection(conn);
3499 return np; /* xmitting; drop packet */
3502 rxi_ClearTransmitQueue(call, 0);
3504 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
3505 rxi_ClearTransmitQueue(call, 0);
3506 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3508 if (np->header.type == RX_PACKET_TYPE_ACK) {
3509 /* now check to see if this is an ack packet acknowledging that the
3510 * server actually *lost* some hard-acked data. If this happens we
3511 * ignore this packet, as it may indicate that the server restarted in
3512 * the middle of a call. It is also possible that this is an old ack
3513 * packet. We don't abort the connection in this case, because this
3514 * *might* just be an old ack packet. The right way to detect a server
3515 * restart in the midst of a call is to notice that the server epoch
3517 /* XXX I'm not sure this is exactly right, since tfirst **IS**
3518 * XXX unacknowledged. I think that this is off-by-one, but
3519 * XXX I don't dare change it just yet, since it will
3520 * XXX interact badly with the server-restart detection
3521 * XXX code in receiveackpacket. */
3522 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
3523 if (rx_stats_active)
3524 rx_atomic_inc(&rx_stats.spuriousPacketsRead);
3525 MUTEX_EXIT(&call->lock);
3526 putConnection(conn);
3530 } /* else not a data packet */
3533 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
3534 /* Set remote user defined status from packet */
3535 call->remoteStatus = np->header.userStatus;
3537 /* Now do packet type-specific processing */
3538 switch (np->header.type) {
3539 case RX_PACKET_TYPE_DATA:
3540 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
3543 case RX_PACKET_TYPE_ACK:
3544 /* Respond immediately to ack packets requesting acknowledgement
3546 if (np->header.flags & RX_REQUEST_ACK) {
3548 (void)rxi_SendCallAbort(call, 0, 1, 0);
3550 (void)rxi_SendAck(call, 0, np->header.serial,
3551 RX_ACK_PING_RESPONSE, 1);
3553 np = rxi_ReceiveAckPacket(call, np, 1);
3555 case RX_PACKET_TYPE_ABORT: {
3556 /* An abort packet: reset the call, passing the error up to the user. */
3557 /* What if error is zero? */
3558 /* What if the error is -1? the application will treat it as a timeout. */
3559 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3560 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
3561 rxi_CallError(call, errdata);
3562 MUTEX_EXIT(&call->lock);
3563 putConnection(conn);
3564 return np; /* xmitting; drop packet */
3566 case RX_PACKET_TYPE_BUSY: {
3567 struct clock busyTime;
3569 clock_GetTime(&busyTime);
3571 MUTEX_EXIT(&call->lock);
3573 MUTEX_ENTER(&conn->conn_call_lock);
3574 MUTEX_ENTER(&call->lock);
3575 conn->lastBusy[call->channel] = busyTime.sec;
3576 call->flags |= RX_CALL_PEER_BUSY;
3577 MUTEX_EXIT(&call->lock);
3578 MUTEX_EXIT(&conn->conn_call_lock);
3580 putConnection(conn);
3584 case RX_PACKET_TYPE_ACKALL:
3585 /* All packets acknowledged, so we can drop all packets previously
3586 * readied for sending */
3587 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3588 /* XXX Hack. We because we can't release the global rx lock when
3589 * sending packets (osi_NetSend) we drop all ack pkts while we're
3590 * traversing the tq in rxi_Start sending packets out because
3591 * packets may move to the freePacketQueue as result of being
3592 * here! So we drop these packets until we're safely out of the
3593 * traversing. Really ugly!
3594 * For fine grain RX locking, we set the acked field in the packets
3595 * and let rxi_Start remove the packets from the transmit queue.
3597 if (call->flags & RX_CALL_TQ_BUSY) {
3598 #ifdef RX_ENABLE_LOCKS
3599 rxi_SetAcksInTransmitQueue(call);
3601 #else /* RX_ENABLE_LOCKS */
3602 MUTEX_EXIT(&call->lock);
3603 putConnection(conn);
3604 return np; /* xmitting; drop packet */
3605 #endif /* RX_ENABLE_LOCKS */
3607 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3608 rxi_ClearTransmitQueue(call, 0);
3611 /* Should not reach here, unless the peer is broken: send an abort
3613 rxi_CallError(call, RX_PROTOCOL_ERROR);
3614 np = rxi_SendCallAbort(call, np, 1, 0);
3617 /* Note when this last legitimate packet was received, for keep-alive
3618 * processing. Note, we delay getting the time until now in the hope that
3619 * the packet will be delivered to the user before any get time is required
3620 * (if not, then the time won't actually be re-evaluated here). */
3621 call->lastReceiveTime = clock_Sec();
3622 /* we've received a legit packet, so the channel is not busy */
3623 call->flags &= ~RX_CALL_PEER_BUSY;
3624 MUTEX_EXIT(&call->lock);
3625 putConnection(conn);
3629 /* return true if this is an "interesting" connection from the point of view
3630 of someone trying to debug the system */
3632 rxi_IsConnInteresting(struct rx_connection *aconn)
3635 struct rx_call *tcall;
3637 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3640 for (i = 0; i < RX_MAXCALLS; i++) {
3641 tcall = aconn->call[i];
3643 if ((tcall->state == RX_STATE_PRECALL)
3644 || (tcall->state == RX_STATE_ACTIVE))
3646 if ((tcall->mode == RX_MODE_SENDING)
3647 || (tcall->mode == RX_MODE_RECEIVING))
3655 /* if this is one of the last few packets AND it wouldn't be used by the
3656 receiving call to immediately satisfy a read request, then drop it on
3657 the floor, since accepting it might prevent a lock-holding thread from
3658 making progress in its reading. If a call has been cleared while in
3659 the precall state then ignore all subsequent packets until the call
3660 is assigned to a thread. */
3663 TooLow(struct rx_packet *ap, struct rx_call *acall)
3667 MUTEX_ENTER(&rx_quota_mutex);
3668 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3669 && (acall->state == RX_STATE_PRECALL))
3670 || ((rx_nFreePackets < rxi_dataQuota + 2)
3671 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3672 && (acall->flags & RX_CALL_READER_WAIT)))) {
3675 MUTEX_EXIT(&rx_quota_mutex);
3681 * Clear the attach wait flag on a connection and proceed.
3683 * Any processing waiting for a connection to be attached should be
3684 * unblocked. We clear the flag and do any other needed tasks.
3687 * the conn to unmark waiting for attach
3689 * @pre conn's conn_data_lock must be locked before calling this function
3693 rxi_ConnClearAttachWait(struct rx_connection *conn)
3695 /* Indicate that rxi_CheckReachEvent is no longer running by
3696 * clearing the flag. Must be atomic under conn_data_lock to
3697 * avoid a new call slipping by: rxi_CheckConnReach holds
3698 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3700 conn->flags &= ~RX_CONN_ATTACHWAIT;
3701 if (conn->flags & RX_CONN_NAT_PING) {
3702 conn->flags &= ~RX_CONN_NAT_PING;
3703 rxi_ScheduleNatKeepAliveEvent(conn);
3708 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
3710 struct rx_connection *conn = arg1;
3711 struct rx_call *acall = arg2;
3712 struct rx_call *call = acall;
3713 struct clock when, now;
3716 MUTEX_ENTER(&conn->conn_data_lock);
3719 rxevent_Put(conn->checkReachEvent);
3720 conn->checkReachEvent = NULL;
3723 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3725 putConnection(conn);
3727 MUTEX_EXIT(&conn->conn_data_lock);
3731 MUTEX_ENTER(&conn->conn_call_lock);
3732 MUTEX_ENTER(&conn->conn_data_lock);
3733 for (i = 0; i < RX_MAXCALLS; i++) {
3734 struct rx_call *tc = conn->call[i];
3735 if (tc && tc->state == RX_STATE_PRECALL) {
3741 rxi_ConnClearAttachWait(conn);
3742 MUTEX_EXIT(&conn->conn_data_lock);
3743 MUTEX_EXIT(&conn->conn_call_lock);
3748 MUTEX_ENTER(&call->lock);
3749 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3751 MUTEX_EXIT(&call->lock);
3753 clock_GetTime(&now);
3755 when.sec += RX_CHECKREACH_TIMEOUT;
3756 MUTEX_ENTER(&conn->conn_data_lock);
3757 if (!conn->checkReachEvent) {
3758 MUTEX_ENTER(&rx_refcnt_mutex);
3760 MUTEX_EXIT(&rx_refcnt_mutex);
3761 conn->checkReachEvent = rxevent_Post(&when, &now,
3762 rxi_CheckReachEvent, conn,
3765 MUTEX_EXIT(&conn->conn_data_lock);
3771 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3773 struct rx_service *service = conn->service;
3774 struct rx_peer *peer = conn->peer;
3775 afs_uint32 now, lastReach;
3777 if (service->checkReach == 0)
3781 MUTEX_ENTER(&peer->peer_lock);
3782 lastReach = peer->lastReachTime;
3783 MUTEX_EXIT(&peer->peer_lock);
3784 if (now - lastReach < RX_CHECKREACH_TTL)
3787 MUTEX_ENTER(&conn->conn_data_lock);
3788 if (conn->flags & RX_CONN_ATTACHWAIT) {
3789 MUTEX_EXIT(&conn->conn_data_lock);
3792 conn->flags |= RX_CONN_ATTACHWAIT;
3793 MUTEX_EXIT(&conn->conn_data_lock);
3794 if (!conn->checkReachEvent)
3795 rxi_CheckReachEvent(NULL, conn, call, 0);
3800 /* try to attach call, if authentication is complete */
3802 TryAttach(struct rx_call *acall, osi_socket socket,
3803 int *tnop, struct rx_call **newcallp,
3806 struct rx_connection *conn = acall->conn;
3808 if (conn->type == RX_SERVER_CONNECTION
3809 && acall->state == RX_STATE_PRECALL) {
3810 /* Don't attach until we have any req'd. authentication. */
3811 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3812 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3813 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3814 /* Note: this does not necessarily succeed; there
3815 * may not any proc available
3818 rxi_ChallengeOn(acall->conn);
3823 /* A data packet has been received off the interface. This packet is
3824 * appropriate to the call (the call is in the right state, etc.). This
3825 * routine can return a packet to the caller, for re-use */
3827 static struct rx_packet *
3828 rxi_ReceiveDataPacket(struct rx_call *call,
3829 struct rx_packet *np, int istack,
3830 osi_socket socket, afs_uint32 host, u_short port,
3831 int *tnop, struct rx_call **newcallp)
3833 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3838 afs_uint32 serial=0, flags=0;
3840 struct rx_packet *tnp;
3841 if (rx_stats_active)
3842 rx_atomic_inc(&rx_stats.dataPacketsRead);
3845 /* If there are no packet buffers, drop this new packet, unless we can find
3846 * packet buffers from inactive calls */
3848 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3849 MUTEX_ENTER(&rx_freePktQ_lock);
3850 rxi_NeedMorePackets = TRUE;
3851 MUTEX_EXIT(&rx_freePktQ_lock);
3852 if (rx_stats_active)
3853 rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
3854 rxi_calltrace(RX_TRACE_DROP, call);
3855 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
3856 /* We used to clear the receive queue here, in an attempt to free
3857 * packets. However this is unsafe if the queue has received a
3858 * soft ACK for the final packet */
3859 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
3865 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3866 * packet is one of several packets transmitted as a single
3867 * datagram. Do not send any soft or hard acks until all packets
3868 * in a jumbogram have been processed. Send negative acks right away.
3870 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3871 /* tnp is non-null when there are more packets in the
3872 * current jumbo gram */
3879 seq = np->header.seq;
3880 serial = np->header.serial;
3881 flags = np->header.flags;
3883 /* If the call is in an error state, send an abort message */
3885 return rxi_SendCallAbort(call, np, istack, 0);
3887 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3888 * AFS 3.5 jumbogram. */
3889 if (flags & RX_JUMBO_PACKET) {
3890 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3895 if (np->header.spare != 0) {
3896 MUTEX_ENTER(&call->conn->conn_data_lock);
3897 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3898 MUTEX_EXIT(&call->conn->conn_data_lock);
3901 /* The usual case is that this is the expected next packet */
3902 if (seq == call->rnext) {
3904 /* Check to make sure it is not a duplicate of one already queued */
3905 if (queue_IsNotEmpty(&call->rq)
3906 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3907 if (rx_stats_active)
3908 rx_atomic_inc(&rx_stats.dupPacketsRead);
3909 dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
3910 rxevent_Cancel(&call->delayedAckEvent, call,
3911 RX_CALL_REFCOUNT_DELAY);
3912 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3918 /* It's the next packet. Stick it on the receive queue
3919 * for this call. Set newPackets to make sure we wake
3920 * the reader once all packets have been processed */
3921 #ifdef RX_TRACK_PACKETS
3922 np->flags |= RX_PKTFLAG_RQ;
3924 queue_Prepend(&call->rq, np);
3925 #ifdef RXDEBUG_PACKET
3927 #endif /* RXDEBUG_PACKET */
3929 np = NULL; /* We can't use this anymore */
3932 /* If an ack is requested then set a flag to make sure we
3933 * send an acknowledgement for this packet */
3934 if (flags & RX_REQUEST_ACK) {
3935 ackNeeded = RX_ACK_REQUESTED;
3938 /* Keep track of whether we have received the last packet */
3939 if (flags & RX_LAST_PACKET) {
3940 call->flags |= RX_CALL_HAVE_LAST;
3944 /* Check whether we have all of the packets for this call */
3945 if (call->flags & RX_CALL_HAVE_LAST) {
3946 afs_uint32 tseq; /* temporary sequence number */
3947 struct rx_packet *tp; /* Temporary packet pointer */
3948 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3950 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
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 rx_packet *tp; /* Temporary packet pointer */
3992 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3993 int missing; /* Are any predecessors missing? */
3995 /* If the new packet's sequence number has been sent to the
3996 * application already, then this is a duplicate */
3997 if (seq < call->rnext) {
3998 if (rx_stats_active)
3999 rx_atomic_inc(&rx_stats.dupPacketsRead);
4000 rxevent_Cancel(&call->delayedAckEvent, call,
4001 RX_CALL_REFCOUNT_DELAY);
4002 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
4008 /* If the sequence number is greater than what can be
4009 * accomodated by the current window, then send a negative
4010 * acknowledge and drop the packet */
4011 if ((call->rnext + call->rwind) <= seq) {
4012 rxevent_Cancel(&call->delayedAckEvent, call,
4013 RX_CALL_REFCOUNT_DELAY);
4014 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
4021 /* Look for the packet in the queue of old received packets */
4022 for (prev = call->rnext - 1, missing =
4023 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4024 /*Check for duplicate packet */
4025 if (seq == tp->header.seq) {
4026 if (rx_stats_active)
4027 rx_atomic_inc(&rx_stats.dupPacketsRead);
4028 rxevent_Cancel(&call->delayedAckEvent, call,
4029 RX_CALL_REFCOUNT_DELAY);
4030 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
4036 /* If we find a higher sequence packet, break out and
4037 * insert the new packet here. */
4038 if (seq < tp->header.seq)
4040 /* Check for missing packet */
4041 if (tp->header.seq != prev + 1) {
4045 prev = tp->header.seq;
4048 /* Keep track of whether we have received the last packet. */
4049 if (flags & RX_LAST_PACKET) {
4050 call->flags |= RX_CALL_HAVE_LAST;
4053 /* It's within the window: add it to the the receive queue.
4054 * tp is left by the previous loop either pointing at the
4055 * packet before which to insert the new packet, or at the
4056 * queue head if the queue is empty or the packet should be
4058 #ifdef RX_TRACK_PACKETS
4059 np->flags |= RX_PKTFLAG_RQ;
4061 #ifdef RXDEBUG_PACKET
4063 #endif /* RXDEBUG_PACKET */
4064 queue_InsertBefore(tp, np);
4068 /* Check whether we have all of the packets for this call */
4069 if ((call->flags & RX_CALL_HAVE_LAST)
4070 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
4071 afs_uint32 tseq; /* temporary sequence number */
4074 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
4075 if (tseq != tp->header.seq)
4077 if (tp->header.flags & RX_LAST_PACKET) {
4078 call->flags |= RX_CALL_RECEIVE_DONE;
4085 /* We need to send an ack of the packet is out of sequence,
4086 * or if an ack was requested by the peer. */
4087 if (seq != prev + 1 || missing) {
4088 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
4089 } else if (flags & RX_REQUEST_ACK) {
4090 ackNeeded = RX_ACK_REQUESTED;
4093 /* Acknowledge the last packet for each call */
4094 if (flags & RX_LAST_PACKET) {
4105 * If the receiver is waiting for an iovec, fill the iovec
4106 * using the data from the receive queue */
4107 if (call->flags & RX_CALL_IOVEC_WAIT) {
4108 didHardAck = rxi_FillReadVec(call, serial);
4109 /* the call may have been aborted */
4118 /* Wakeup the reader if any */
4119 if ((call->flags & RX_CALL_READER_WAIT)
4120 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
4121 || (call->iovNext >= call->iovMax)
4122 || (call->flags & RX_CALL_RECEIVE_DONE))) {
4123 call->flags &= ~RX_CALL_READER_WAIT;
4124 #ifdef RX_ENABLE_LOCKS
4125 CV_BROADCAST(&call->cv_rq);
4127 osi_rxWakeup(&call->rq);
4133 * Send an ack when requested by the peer, or once every
4134 * rxi_SoftAckRate packets until the last packet has been
4135 * received. Always send a soft ack for the last packet in
4136 * the server's reply. */
4138 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4139 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
4140 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
4141 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4142 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
4143 } else if (call->nSoftAcks) {
4144 if (haveLast && !(flags & RX_CLIENT_INITIATED))
4145 rxi_PostDelayedAckEvent(call, &rx_lastAckDelay);
4147 rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
4148 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
4149 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4156 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
4158 struct rx_peer *peer = conn->peer;
4160 MUTEX_ENTER(&peer->peer_lock);
4161 peer->lastReachTime = clock_Sec();
4162 MUTEX_EXIT(&peer->peer_lock);
4164 MUTEX_ENTER(&conn->conn_data_lock);
4165 if (conn->flags & RX_CONN_ATTACHWAIT) {
4168 rxi_ConnClearAttachWait(conn);
4169 MUTEX_EXIT(&conn->conn_data_lock);
4171 for (i = 0; i < RX_MAXCALLS; i++) {
4172 struct rx_call *call = conn->call[i];
4175 MUTEX_ENTER(&call->lock);
4176 /* tnop can be null if newcallp is null */
4177 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
4179 MUTEX_EXIT(&call->lock);
4183 MUTEX_EXIT(&conn->conn_data_lock);
4186 #if defined(RXDEBUG) && defined(AFS_NT40_ENV)
4188 rx_ack_reason(int reason)
4191 case RX_ACK_REQUESTED:
4193 case RX_ACK_DUPLICATE:
4195 case RX_ACK_OUT_OF_SEQUENCE:
4197 case RX_ACK_EXCEEDS_WINDOW:
4199 case RX_ACK_NOSPACE:
4203 case RX_ACK_PING_RESPONSE:
4216 /* The real smarts of the whole thing. */
4217 static struct rx_packet *
4218 rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
4221 struct rx_ackPacket *ap;
4223 struct rx_packet *tp;
4224 struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
4225 struct rx_connection *conn = call->conn;
4226 struct rx_peer *peer = conn->peer;
4227 struct clock now; /* Current time, for RTT calculations */
4235 int newAckCount = 0;
4236 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
4237 int pktsize = 0; /* Set if we need to update the peer mtu */
4238 int conn_data_locked = 0;
4240 if (rx_stats_active)
4241 rx_atomic_inc(&rx_stats.ackPacketsRead);
4242 ap = (struct rx_ackPacket *)rx_DataOf(np);
4243 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
4245 return np; /* truncated ack packet */
4247 /* depends on ack packet struct */
4248 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
4249 first = ntohl(ap->firstPacket);
4250 prev = ntohl(ap->previousPacket);
4251 serial = ntohl(ap->serial);
4253 /* Ignore ack packets received out of order */
4254 if (first < call->tfirst ||
4255 (first == call->tfirst && prev < call->tprev)) {
4261 if (np->header.flags & RX_SLOW_START_OK) {
4262 call->flags |= RX_CALL_SLOW_START_OK;
4265 if (ap->reason == RX_ACK_PING_RESPONSE)
4266 rxi_UpdatePeerReach(conn, call);
4268 if (conn->lastPacketSizeSeq) {
4269 MUTEX_ENTER(&conn->conn_data_lock);
4270 conn_data_locked = 1;
4271 if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
4272 pktsize = conn->lastPacketSize;
4273 conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
4276 if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
4277 if (!conn_data_locked) {
4278 MUTEX_ENTER(&conn->conn_data_lock);
4279 conn_data_locked = 1;
4281 if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
4282 /* process mtu ping ack */
4283 pktsize = conn->lastPingSize;
4284 conn->lastPingSizeSer = conn->lastPingSize = 0;
4288 if (conn_data_locked) {
4289 MUTEX_EXIT(&conn->conn_data_lock);
4290 conn_data_locked = 0;
4294 if (rxdebug_active) {
4298 len = _snprintf(msg, sizeof(msg),
4299 "tid[%d] RACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4300 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4301 ntohl(ap->serial), ntohl(ap->previousPacket),
4302 (unsigned int)np->header.seq, ntohl(ap->firstPacket),
4303 ap->nAcks, ntohs(ap->bufferSpace) );
4307 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
4308 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4312 OutputDebugString(msg);
4314 #else /* AFS_NT40_ENV */
4317 "RACK: reason %x previous %u seq %u serial %u first %u",
4318 ap->reason, ntohl(ap->previousPacket),
4319 (unsigned int)np->header.seq, (unsigned int)serial,
4320 ntohl(ap->firstPacket));
4323 for (offset = 0; offset < nAcks; offset++)
4324 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4329 #endif /* AFS_NT40_ENV */
4332 MUTEX_ENTER(&peer->peer_lock);
4335 * Start somewhere. Can't assume we can send what we can receive,
4336 * but we are clearly receiving.
4338 if (!peer->maxPacketSize)
4339 peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
4341 if (pktsize > peer->maxPacketSize) {
4342 peer->maxPacketSize = pktsize;
4343 if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
4344 peer->ifMTU=pktsize-RX_IPUDP_SIZE;
4345 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
4346 rxi_ScheduleGrowMTUEvent(call, 1);
4351 clock_GetTime(&now);
4353 /* The transmit queue splits into 4 sections.
4355 * The first section is packets which have now been acknowledged
4356 * by a window size change in the ack. These have reached the
4357 * application layer, and may be discarded. These are packets
4358 * with sequence numbers < ap->firstPacket.
4360 * The second section is packets which have sequence numbers in
4361 * the range ap->firstPacket to ap->firstPacket + ap->nAcks. The
4362 * contents of the packet's ack array determines whether these
4363 * packets are acknowledged or not.
4365 * The third section is packets which fall above the range
4366 * addressed in the ack packet. These have not yet been received
4369 * The four section is packets which have not yet been transmitted.
4370 * These packets will have a header.serial of 0.
4373 /* First section - implicitly acknowledged packets that can be
4377 tp = queue_First(&call->tq, rx_packet);
4378 while(!queue_IsEnd(&call->tq, tp) && tp->header.seq < first) {
4379 struct rx_packet *next;
4381 next = queue_Next(tp, rx_packet);
4382 call->tfirst = tp->header.seq + 1;
4384 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4386 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4389 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4390 /* XXX Hack. Because we have to release the global rx lock when sending
4391 * packets (osi_NetSend) we drop all acks while we're traversing the tq
4392 * in rxi_Start sending packets out because packets may move to the
4393 * freePacketQueue as result of being here! So we drop these packets until
4394 * we're safely out of the traversing. Really ugly!
4395 * To make it even uglier, if we're using fine grain locking, we can
4396 * set the ack bits in the packets and have rxi_Start remove the packets
4397 * when it's done transmitting.
4399 if (call->flags & RX_CALL_TQ_BUSY) {
4400 #ifdef RX_ENABLE_LOCKS
4401 tp->flags |= RX_PKTFLAG_ACKED;
4402 call->flags |= RX_CALL_TQ_SOME_ACKED;
4403 #else /* RX_ENABLE_LOCKS */
4405 #endif /* RX_ENABLE_LOCKS */
4407 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4410 #ifdef RX_TRACK_PACKETS
4411 tp->flags &= ~RX_PKTFLAG_TQ;
4413 #ifdef RXDEBUG_PACKET
4415 #endif /* RXDEBUG_PACKET */
4416 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
4421 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
4423 /* Second section of the queue - packets for which we are receiving
4426 * Go through the explicit acks/nacks and record the results in
4427 * the waiting packets. These are packets that can't be released
4428 * yet, even with a positive acknowledge. This positive
4429 * acknowledge only means the packet has been received by the
4430 * peer, not that it will be retained long enough to be sent to
4431 * the peer's upper level. In addition, reset the transmit timers
4432 * of any missing packets (those packets that must be missing
4433 * because this packet was out of sequence) */
4435 call->nSoftAcked = 0;
4437 while (!queue_IsEnd(&call->tq, tp) && tp->header.seq < first + nAcks) {
4438 /* Set the acknowledge flag per packet based on the
4439 * information in the ack packet. An acknowlegded packet can
4440 * be downgraded when the server has discarded a packet it
4441 * soacked previously, or when an ack packet is received
4442 * out of sequence. */
4443 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
4444 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4446 tp->flags |= RX_PKTFLAG_ACKED;
4447 rxi_ComputeRoundTripTime(tp, ap, call, peer, &now);
4454 } else /* RX_ACK_TYPE_NACK */ {
4455 tp->flags &= ~RX_PKTFLAG_ACKED;
4459 tp = queue_Next(tp, rx_packet);
4462 /* We don't need to take any action with the 3rd or 4th section in the
4463 * queue - they're not addressed by the contents of this ACK packet.
4466 /* If the window has been extended by this acknowledge packet,
4467 * then wakeup a sender waiting in alloc for window space, or try
4468 * sending packets now, if he's been sitting on packets due to
4469 * lack of window space */
4470 if (call->tnext < (call->tfirst + call->twind)) {
4471 #ifdef RX_ENABLE_LOCKS
4472 CV_SIGNAL(&call->cv_twind);
4474 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
4475 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
4476 osi_rxWakeup(&call->twind);
4479 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
4480 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
4484 /* if the ack packet has a receivelen field hanging off it,
4485 * update our state */
4486 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
4489 /* If the ack packet has a "recommended" size that is less than
4490 * what I am using now, reduce my size to match */
4491 rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
4492 (int)sizeof(afs_int32), &tSize);
4493 tSize = (afs_uint32) ntohl(tSize);
4494 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
4496 /* Get the maximum packet size to send to this peer */
4497 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
4499 tSize = (afs_uint32) ntohl(tSize);
4500 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
4501 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
4503 /* sanity check - peer might have restarted with different params.
4504 * If peer says "send less", dammit, send less... Peer should never
4505 * be unable to accept packets of the size that prior AFS versions would
4506 * send without asking. */
4507 if (peer->maxMTU != tSize) {
4508 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
4510 peer->maxMTU = tSize;
4511 peer->MTU = MIN(tSize, peer->MTU);
4512 call->MTU = MIN(call->MTU, tSize);
4515 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
4518 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4519 (int)sizeof(afs_int32), &tSize);
4520 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
4521 if (tSize < call->twind) { /* smaller than our send */
4522 call->twind = tSize; /* window, we must send less... */
4523 call->ssthresh = MIN(call->twind, call->ssthresh);
4524 call->conn->twind[call->channel] = call->twind;
4527 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
4528 * network MTU confused with the loopback MTU. Calculate the
4529 * maximum MTU here for use in the slow start code below.
4531 /* Did peer restart with older RX version? */
4532 if (peer->maxDgramPackets > 1) {
4533 peer->maxDgramPackets = 1;
4535 } else if (np->length >=
4536 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
4539 rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
4540 sizeof(afs_int32), &tSize);
4541 tSize = (afs_uint32) ntohl(tSize);
4543 * As of AFS 3.5 we set the send window to match the receive window.
4545 if (tSize < call->twind) {
4546 call->twind = tSize;
4547 call->conn->twind[call->channel] = call->twind;
4548 call->ssthresh = MIN(call->twind, call->ssthresh);
4549 } else if (tSize > call->twind) {
4550 call->twind = tSize;
4551 call->conn->twind[call->channel] = call->twind;
4555 * As of AFS 3.5, a jumbogram is more than one fixed size
4556 * packet transmitted in a single UDP datagram. If the remote
4557 * MTU is smaller than our local MTU then never send a datagram
4558 * larger than the natural MTU.
4561 rx_AckDataSize(ap->nAcks) + 3 * (int)sizeof(afs_int32),
4562 (int)sizeof(afs_int32), &tSize);
4563 maxDgramPackets = (afs_uint32) ntohl(tSize);
4564 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
4566 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
4567 if (maxDgramPackets > 1) {
4568 peer->maxDgramPackets = maxDgramPackets;
4569 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4571 peer->maxDgramPackets = 1;
4572 call->MTU = peer->natMTU;
4574 } else if (peer->maxDgramPackets > 1) {
4575 /* Restarted with lower version of RX */
4576 peer->maxDgramPackets = 1;
4578 } else if (peer->maxDgramPackets > 1
4579 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
4580 /* Restarted with lower version of RX */
4581 peer->maxMTU = OLD_MAX_PACKET_SIZE;
4582 peer->natMTU = OLD_MAX_PACKET_SIZE;
4583 peer->MTU = OLD_MAX_PACKET_SIZE;
4584 peer->maxDgramPackets = 1;
4585 peer->nDgramPackets = 1;
4587 call->MTU = OLD_MAX_PACKET_SIZE;
4592 * Calculate how many datagrams were successfully received after
4593 * the first missing packet and adjust the negative ack counter
4598 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4599 if (call->nNacks < nNacked) {
4600 call->nNacks = nNacked;
4603 call->nAcks += newAckCount;
4607 /* If the packet contained new acknowledgements, rather than just
4608 * being a duplicate of one we have previously seen, then we can restart
4611 if (newAckCount > 0)
4612 rxi_rto_packet_acked(call, istack);
4614 if (call->flags & RX_CALL_FAST_RECOVER) {
4615 if (newAckCount == 0) {
4616 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4618 call->flags &= ~RX_CALL_FAST_RECOVER;
4619 call->cwind = call->nextCwind;
4620 call->nextCwind = 0;
4623 call->nCwindAcks = 0;
4624 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4625 /* Three negative acks in a row trigger congestion recovery */
4626 call->flags |= RX_CALL_FAST_RECOVER;
4627 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4629 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4630 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4631 call->nextCwind = call->ssthresh;
4634 peer->MTU = call->MTU;
4635 peer->cwind = call->nextCwind;
4636 peer->nDgramPackets = call->nDgramPackets;
4638 call->congestSeq = peer->congestSeq;
4640 /* Reset the resend times on the packets that were nacked
4641 * so we will retransmit as soon as the window permits
4644 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4646 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4647 tp->flags &= ~RX_PKTFLAG_SENT;
4649 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4654 /* If cwind is smaller than ssthresh, then increase
4655 * the window one packet for each ack we receive (exponential
4657 * If cwind is greater than or equal to ssthresh then increase
4658 * the congestion window by one packet for each cwind acks we
4659 * receive (linear growth). */
4660 if (call->cwind < call->ssthresh) {
4662 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4663 call->nCwindAcks = 0;
4665 call->nCwindAcks += newAckCount;
4666 if (call->nCwindAcks >= call->cwind) {
4667 call->nCwindAcks = 0;
4668 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4672 * If we have received several acknowledgements in a row then
4673 * it is time to increase the size of our datagrams
4675 if ((int)call->nAcks > rx_nDgramThreshold) {
4676 if (peer->maxDgramPackets > 1) {
4677 if (call->nDgramPackets < peer->maxDgramPackets) {
4678 call->nDgramPackets++;
4680 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4681 } else if (call->MTU < peer->maxMTU) {
4682 /* don't upgrade if we can't handle it */
4683 if ((call->nDgramPackets == 1) && (call->MTU >= peer->ifMTU))
4684 call->MTU = peer->ifMTU;
4686 call->MTU += peer->natMTU;
4687 call->MTU = MIN(call->MTU, peer->maxMTU);
4694 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4696 /* Servers need to hold the call until all response packets have
4697 * been acknowledged. Soft acks are good enough since clients
4698 * are not allowed to clear their receive queues. */
4699 if (call->state == RX_STATE_HOLD
4700 && call->tfirst + call->nSoftAcked >= call->tnext) {
4701 call->state = RX_STATE_DALLY;
4702 rxi_ClearTransmitQueue(call, 0);
4703 rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4704 } else if (!queue_IsEmpty(&call->tq)) {
4705 rxi_Start(call, istack);
4710 /* Received a response to a challenge packet */
4711 static struct rx_packet *
4712 rxi_ReceiveResponsePacket(struct rx_connection *conn,
4713 struct rx_packet *np, int istack)
4717 /* Ignore the packet if we're the client */
4718 if (conn->type == RX_CLIENT_CONNECTION)
4721 /* If already authenticated, ignore the packet (it's probably a retry) */
4722 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4725 /* Otherwise, have the security object evaluate the response packet */
4726 error = RXS_CheckResponse(conn->securityObject, conn, np);
4728 /* If the response is invalid, reset the connection, sending
4729 * an abort to the peer */
4733 rxi_ConnectionError(conn, error);
4734 MUTEX_ENTER(&conn->conn_data_lock);
4735 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4736 MUTEX_EXIT(&conn->conn_data_lock);
4739 /* If the response is valid, any calls waiting to attach
4740 * servers can now do so */
4743 for (i = 0; i < RX_MAXCALLS; i++) {
4744 struct rx_call *call = conn->call[i];
4746 MUTEX_ENTER(&call->lock);
4747 if (call->state == RX_STATE_PRECALL)
4748 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4749 /* tnop can be null if newcallp is null */
4750 MUTEX_EXIT(&call->lock);
4754 /* Update the peer reachability information, just in case
4755 * some calls went into attach-wait while we were waiting
4756 * for authentication..
4758 rxi_UpdatePeerReach(conn, NULL);
4763 /* A client has received an authentication challenge: the security
4764 * object is asked to cough up a respectable response packet to send
4765 * back to the server. The server is responsible for retrying the
4766 * challenge if it fails to get a response. */
4768 static struct rx_packet *
4769 rxi_ReceiveChallengePacket(struct rx_connection *conn,
4770 struct rx_packet *np, int istack)
4774 /* Ignore the challenge if we're the server */
4775 if (conn->type == RX_SERVER_CONNECTION)
4778 /* Ignore the challenge if the connection is otherwise idle; someone's
4779 * trying to use us as an oracle. */
4780 if (!rxi_HasActiveCalls(conn))
4783 /* Send the security object the challenge packet. It is expected to fill
4784 * in the response. */
4785 error = RXS_GetResponse(conn->securityObject, conn, np);
4787 /* If the security object is unable to return a valid response, reset the
4788 * connection and send an abort to the peer. Otherwise send the response
4789 * packet to the peer connection. */
4791 rxi_ConnectionError(conn, error);
4792 MUTEX_ENTER(&conn->conn_data_lock);
4793 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4794 MUTEX_EXIT(&conn->conn_data_lock);
4796 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4797 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4803 /* Find an available server process to service the current request in
4804 * the given call structure. If one isn't available, queue up this
4805 * call so it eventually gets one */
4807 rxi_AttachServerProc(struct rx_call *call,
4808 osi_socket socket, int *tnop,
4809 struct rx_call **newcallp)
4811 struct rx_serverQueueEntry *sq;
4812 struct rx_service *service = call->conn->service;
4815 /* May already be attached */
4816 if (call->state == RX_STATE_ACTIVE)
4819 MUTEX_ENTER(&rx_serverPool_lock);
4821 haveQuota = QuotaOK(service);
4822 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4823 /* If there are no processes available to service this call,
4824 * put the call on the incoming call queue (unless it's
4825 * already on the queue).
4827 #ifdef RX_ENABLE_LOCKS
4829 ReturnToServerPool(service);
4830 #endif /* RX_ENABLE_LOCKS */
4832 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4833 call->flags |= RX_CALL_WAIT_PROC;
4834 rx_atomic_inc(&rx_nWaiting);
4835 rx_atomic_inc(&rx_nWaited);
4836 rxi_calltrace(RX_CALL_ARRIVAL, call);
4837 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4838 queue_Append(&rx_incomingCallQueue, call);
4841 sq = queue_Last(&rx_idleServerQueue, rx_serverQueueEntry);
4843 /* If hot threads are enabled, and both newcallp and sq->socketp
4844 * are non-null, then this thread will process the call, and the
4845 * idle server thread will start listening on this threads socket.
4848 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4851 *sq->socketp = socket;
4852 clock_GetTime(&call->startTime);
4853 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4857 if (call->flags & RX_CALL_WAIT_PROC) {
4858 /* Conservative: I don't think this should happen */
4859 call->flags &= ~RX_CALL_WAIT_PROC;
4860 rx_atomic_dec(&rx_nWaiting);
4861 if (queue_IsOnQueue(call)) {
4865 call->state = RX_STATE_ACTIVE;
4866 call->mode = RX_MODE_RECEIVING;
4867 #ifdef RX_KERNEL_TRACE
4869 int glockOwner = ISAFS_GLOCK();
4872 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4873 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4879 if (call->flags & RX_CALL_CLEARED) {
4880 /* send an ack now to start the packet flow up again */
4881 call->flags &= ~RX_CALL_CLEARED;
4882 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4884 #ifdef RX_ENABLE_LOCKS
4887 service->nRequestsRunning++;
4888 MUTEX_ENTER(&rx_quota_mutex);
4889 if (service->nRequestsRunning <= service->minProcs)
4892 MUTEX_EXIT(&rx_quota_mutex);
4896 MUTEX_EXIT(&rx_serverPool_lock);
4899 /* Delay the sending of an acknowledge event for a short while, while
4900 * a new call is being prepared (in the case of a client) or a reply
4901 * is being prepared (in the case of a server). Rather than sending
4902 * an ack packet, an ACKALL packet is sent. */
4904 rxi_AckAll(struct rx_call *call)
4906 rxi_SendSpecial(call, call->conn, NULL, RX_PACKET_TYPE_ACKALL,
4908 call->flags |= RX_CALL_ACKALL_SENT;
4912 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
4915 struct rx_call *call = arg1;
4916 #ifdef RX_ENABLE_LOCKS
4918 MUTEX_ENTER(&call->lock);
4919 if (event == call->delayedAckEvent) {
4920 rxevent_Put(call->delayedAckEvent);
4921 call->delayedAckEvent = NULL;
4923 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4925 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4927 MUTEX_EXIT(&call->lock);
4928 #else /* RX_ENABLE_LOCKS */
4930 rxevent_Put(call->delayedAckEvent);
4931 call->delayedAckEvent = NULL;
4933 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4934 #endif /* RX_ENABLE_LOCKS */
4938 #ifdef RX_ENABLE_LOCKS
4939 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4940 * clearing them out.
4943 rxi_SetAcksInTransmitQueue(struct rx_call *call)
4945 struct rx_packet *p, *tp;
4948 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4949 p->flags |= RX_PKTFLAG_ACKED;
4953 call->flags |= RX_CALL_TQ_CLEARME;
4954 call->flags |= RX_CALL_TQ_SOME_ACKED;
4957 rxi_rto_cancel(call);
4959 call->tfirst = call->tnext;
4960 call->nSoftAcked = 0;
4962 if (call->flags & RX_CALL_FAST_RECOVER) {
4963 call->flags &= ~RX_CALL_FAST_RECOVER;
4964 call->cwind = call->nextCwind;
4965 call->nextCwind = 0;
4968 CV_SIGNAL(&call->cv_twind);
4970 #endif /* RX_ENABLE_LOCKS */
4972 /* Clear out the transmit queue for the current call (all packets have
4973 * been received by peer) */
4975 rxi_ClearTransmitQueue(struct rx_call *call, int force)
4977 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4978 struct rx_packet *p, *tp;
4980 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4982 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4983 p->flags |= RX_PKTFLAG_ACKED;
4987 call->flags |= RX_CALL_TQ_CLEARME;
4988 call->flags |= RX_CALL_TQ_SOME_ACKED;
4991 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4992 #ifdef RXDEBUG_PACKET
4994 #endif /* RXDEBUG_PACKET */
4995 rxi_FreePackets(0, &call->tq);
4996 rxi_WakeUpTransmitQueue(call);
4997 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4998 call->flags &= ~RX_CALL_TQ_CLEARME;
5000 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5002 rxi_rto_cancel(call);
5003 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
5004 call->nSoftAcked = 0;
5006 if (call->flags & RX_CALL_FAST_RECOVER) {
5007 call->flags &= ~RX_CALL_FAST_RECOVER;
5008 call->cwind = call->nextCwind;
5010 #ifdef RX_ENABLE_LOCKS
5011 CV_SIGNAL(&call->cv_twind);
5013 osi_rxWakeup(&call->twind);
5018 rxi_ClearReceiveQueue(struct rx_call *call)
5020 if (queue_IsNotEmpty(&call->rq)) {
5023 count = rxi_FreePackets(0, &call->rq);
5024 rx_packetReclaims += count;
5025 #ifdef RXDEBUG_PACKET
5027 if ( call->rqc != 0 )
5028 dpf(("rxi_ClearReceiveQueue call %"AFS_PTR_FMT" rqc %u != 0\n", call, call->rqc));
5030 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
5032 if (call->state == RX_STATE_PRECALL) {
5033 call->flags |= RX_CALL_CLEARED;
5037 /* Send an abort packet for the specified call */
5038 static struct rx_packet *
5039 rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
5040 int istack, int force)
5042 afs_int32 error, cerror;
5043 struct clock when, now;
5048 switch (call->error) {
5051 cerror = RX_CALL_TIMEOUT;
5054 cerror = call->error;
5057 /* Clients should never delay abort messages */
5058 if (rx_IsClientConn(call->conn))
5061 if (call->abortCode != cerror) {
5062 call->abortCode = cerror;
5063 call->abortCount = 0;
5066 if (force || rxi_callAbortThreshhold == 0
5067 || call->abortCount < rxi_callAbortThreshhold) {
5068 if (call->delayedAbortEvent) {
5069 rxevent_Cancel(&call->delayedAbortEvent, call,
5070 RX_CALL_REFCOUNT_ABORT);
5072 error = htonl(cerror);
5075 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5076 (char *)&error, sizeof(error), istack);
5077 } else if (!call->delayedAbortEvent) {
5078 clock_GetTime(&now);
5080 clock_Addmsec(&when, rxi_callAbortDelay);
5081 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
5082 call->delayedAbortEvent =
5083 rxevent_Post(&when, &now, rxi_SendDelayedCallAbort, call, 0, 0);
5088 /* Send an abort packet for the specified connection. Packet is an
5089 * optional pointer to a packet that can be used to send the abort.
5090 * Once the number of abort messages reaches the threshhold, an
5091 * event is scheduled to send the abort. Setting the force flag
5092 * overrides sending delayed abort messages.
5094 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
5095 * to send the abort packet.
5098 rxi_SendConnectionAbort(struct rx_connection *conn,
5099 struct rx_packet *packet, int istack, int force)
5102 struct clock when, now;
5107 /* Clients should never delay abort messages */
5108 if (rx_IsClientConn(conn))
5111 if (force || rxi_connAbortThreshhold == 0
5112 || conn->abortCount < rxi_connAbortThreshhold) {
5114 rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
5115 error = htonl(conn->error);
5117 MUTEX_EXIT(&conn->conn_data_lock);
5119 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5120 RX_PACKET_TYPE_ABORT, (char *)&error,
5121 sizeof(error), istack);
5122 MUTEX_ENTER(&conn->conn_data_lock);
5123 } else if (!conn->delayedAbortEvent) {
5124 clock_GetTime(&now);
5126 clock_Addmsec(&when, rxi_connAbortDelay);
5127 conn->delayedAbortEvent =
5128 rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
5133 /* Associate an error all of the calls owned by a connection. Called
5134 * with error non-zero. This is only for really fatal things, like
5135 * bad authentication responses. The connection itself is set in
5136 * error at this point, so that future packets received will be
5139 rxi_ConnectionError(struct rx_connection *conn,
5145 dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
5147 MUTEX_ENTER(&conn->conn_data_lock);
5148 rxevent_Cancel(&conn->challengeEvent, NULL, 0);
5149 rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
5150 if (conn->checkReachEvent) {
5151 rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
5152 conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
5153 putConnection(conn);
5155 MUTEX_EXIT(&conn->conn_data_lock);
5156 for (i = 0; i < RX_MAXCALLS; i++) {
5157 struct rx_call *call = conn->call[i];
5159 MUTEX_ENTER(&call->lock);
5160 rxi_CallError(call, error);
5161 MUTEX_EXIT(&call->lock);
5164 conn->error = error;
5165 if (rx_stats_active)
5166 rx_atomic_inc(&rx_stats.fatalErrors);
5171 * Interrupt an in-progress call with the specified error and wakeup waiters.
5173 * @param[in] call The call to interrupt
5174 * @param[in] error The error code to send to the peer
5177 rx_InterruptCall(struct rx_call *call, afs_int32 error)
5179 MUTEX_ENTER(&call->lock);
5180 rxi_CallError(call, error);
5181 rxi_SendCallAbort(call, NULL, 0, 1);
5182 MUTEX_EXIT(&call->lock);
5186 rxi_CallError(struct rx_call *call, afs_int32 error)
5189 osirx_AssertMine(&call->lock, "rxi_CallError");
5191 dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
5193 error = call->error;
5195 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5196 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
5197 rxi_ResetCall(call, 0);
5200 rxi_ResetCall(call, 0);
5202 call->error = error;
5205 /* Reset various fields in a call structure, and wakeup waiting
5206 * processes. Some fields aren't changed: state & mode are not
5207 * touched (these must be set by the caller), and bufptr, nLeft, and
5208 * nFree are not reset, since these fields are manipulated by
5209 * unprotected macros, and may only be reset by non-interrupting code.
5213 rxi_ResetCall(struct rx_call *call, int newcall)
5216 struct rx_peer *peer;
5217 struct rx_packet *packet;
5219 osirx_AssertMine(&call->lock, "rxi_ResetCall");
5221 dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
5223 /* Notify anyone who is waiting for asynchronous packet arrival */
5224 if (call->arrivalProc) {
5225 (*call->arrivalProc) (call, call->arrivalProcHandle,
5226 call->arrivalProcArg);
5227 call->arrivalProc = (void (*)())0;
5231 rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
5233 if (call->delayedAbortEvent) {
5234 rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
5235 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5237 rxi_SendCallAbort(call, packet, 0, 1);
5238 rxi_FreePacket(packet);
5243 * Update the peer with the congestion information in this call
5244 * so other calls on this connection can pick up where this call
5245 * left off. If the congestion sequence numbers don't match then
5246 * another call experienced a retransmission.
5248 peer = call->conn->peer;
5249 MUTEX_ENTER(&peer->peer_lock);
5251 if (call->congestSeq == peer->congestSeq) {
5252 peer->cwind = MAX(peer->cwind, call->cwind);
5253 peer->MTU = MAX(peer->MTU, call->MTU);
5254 peer->nDgramPackets =
5255 MAX(peer->nDgramPackets, call->nDgramPackets);
5258 call->abortCode = 0;
5259 call->abortCount = 0;
5261 if (peer->maxDgramPackets > 1) {
5262 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
5264 call->MTU = peer->MTU;
5266 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
5267 call->ssthresh = rx_maxSendWindow;
5268 call->nDgramPackets = peer->nDgramPackets;
5269 call->congestSeq = peer->congestSeq;
5270 call->rtt = peer->rtt;
5271 call->rtt_dev = peer->rtt_dev;
5272 clock_Zero(&call->rto);
5273 clock_Addmsec(&call->rto,
5274 MAX(((call->rtt >> 3) + call->rtt_dev), rx_minPeerTimeout) + 200);
5275 MUTEX_EXIT(&peer->peer_lock);
5277 flags = call->flags;
5278 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5279 rxi_WaitforTQBusy(call);
5280 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5282 rxi_ClearTransmitQueue(call, 1);
5283 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
5284 dpf(("rcall %"AFS_PTR_FMT" has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5288 if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
5289 /* The call channel is still busy; resetting the call doesn't change
5290 * that. However, if 'newcall' is set, we are processing a call
5291 * structure that has either been recycled from the free list, or has
5292 * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
5293 * 'newcall' is set, since it describes a completely different call
5294 * channel which we do not care about. */
5295 call->flags |= RX_CALL_PEER_BUSY;
5298 rxi_ClearReceiveQueue(call);
5299 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
5303 call->twind = call->conn->twind[call->channel];
5304 call->rwind = call->conn->rwind[call->channel];
5305 call->nSoftAcked = 0;
5306 call->nextCwind = 0;
5309 call->nCwindAcks = 0;
5310 call->nSoftAcks = 0;
5311 call->nHardAcks = 0;
5313 call->tfirst = call->rnext = call->tnext = 1;
5316 call->lastAcked = 0;
5317 call->localStatus = call->remoteStatus = 0;
5319 if (flags & RX_CALL_READER_WAIT) {
5320 #ifdef RX_ENABLE_LOCKS
5321 CV_BROADCAST(&call->cv_rq);
5323 osi_rxWakeup(&call->rq);
5326 if (flags & RX_CALL_WAIT_PACKETS) {
5327 MUTEX_ENTER(&rx_freePktQ_lock);
5328 rxi_PacketsUnWait(); /* XXX */
5329 MUTEX_EXIT(&rx_freePktQ_lock);
5331 #ifdef RX_ENABLE_LOCKS
5332 CV_SIGNAL(&call->cv_twind);
5334 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
5335 osi_rxWakeup(&call->twind);
5338 if (flags & RX_CALL_WAIT_PROC) {
5339 rx_atomic_dec(&rx_nWaiting);
5341 #ifdef RX_ENABLE_LOCKS
5342 /* The following ensures that we don't mess with any queue while some
5343 * other thread might also be doing so. The call_queue_lock field is
5344 * is only modified under the call lock. If the call is in the process
5345 * of being removed from a queue, the call is not locked until the
5346 * the queue lock is dropped and only then is the call_queue_lock field
5347 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
5348 * Note that any other routine which removes a call from a queue has to
5349 * obtain the queue lock before examing the queue and removing the call.
5351 if (call->call_queue_lock) {
5352 MUTEX_ENTER(call->call_queue_lock);
5353 if (queue_IsOnQueue(call)) {
5356 MUTEX_EXIT(call->call_queue_lock);
5357 CLEAR_CALL_QUEUE_LOCK(call);
5359 #else /* RX_ENABLE_LOCKS */
5360 if (queue_IsOnQueue(call)) {
5363 #endif /* RX_ENABLE_LOCKS */
5365 rxi_KeepAliveOff(call);
5366 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5369 /* Send an acknowledge for the indicated packet (seq,serial) of the
5370 * indicated call, for the indicated reason (reason). This
5371 * acknowledge will specifically acknowledge receiving the packet, and
5372 * will also specify which other packets for this call have been
5373 * received. This routine returns the packet that was used to the
5374 * caller. The caller is responsible for freeing it or re-using it.
5375 * This acknowledgement also returns the highest sequence number
5376 * actually read out by the higher level to the sender; the sender
5377 * promises to keep around packets that have not been read by the
5378 * higher level yet (unless, of course, the sender decides to abort
5379 * the call altogether). Any of p, seq, serial, pflags, or reason may
5380 * be set to zero without ill effect. That is, if they are zero, they
5381 * will not convey any information.
5382 * NOW there is a trailer field, after the ack where it will safely be
5383 * ignored by mundanes, which indicates the maximum size packet this
5384 * host can swallow. */
5386 struct rx_packet *optionalPacket; use to send ack (or null)
5387 int seq; Sequence number of the packet we are acking
5388 int serial; Serial number of the packet
5389 int pflags; Flags field from packet header
5390 int reason; Reason an acknowledge was prompted
5394 rxi_SendAck(struct rx_call *call,
5395 struct rx_packet *optionalPacket, int serial, int reason,
5398 struct rx_ackPacket *ap;
5399 struct rx_packet *rqp;
5400 struct rx_packet *nxp; /* For queue_Scan */
5401 struct rx_packet *p;
5404 afs_uint32 padbytes = 0;
5405 #ifdef RX_ENABLE_TSFPQ
5406 struct rx_ts_info_t * rx_ts_info;
5410 * Open the receive window once a thread starts reading packets
5412 if (call->rnext > 1) {
5413 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
5416 /* Don't attempt to grow MTU if this is a critical ping */
5417 if (reason == RX_ACK_MTU) {
5418 /* keep track of per-call attempts, if we're over max, do in small
5419 * otherwise in larger? set a size to increment by, decrease
5422 if (call->conn->peer->maxPacketSize &&
5423 (call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
5425 padbytes = call->conn->peer->maxPacketSize+16;
5427 padbytes = call->conn->peer->maxMTU + 128;
5429 /* do always try a minimum size ping */
5430 padbytes = MAX(padbytes, RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE+4);
5432 /* subtract the ack payload */
5433 padbytes -= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32));
5434 reason = RX_ACK_PING;
5437 call->nHardAcks = 0;
5438 call->nSoftAcks = 0;
5439 if (call->rnext > call->lastAcked)
5440 call->lastAcked = call->rnext;
5444 rx_computelen(p, p->length); /* reset length, you never know */
5445 } /* where that's been... */
5446 #ifdef RX_ENABLE_TSFPQ
5448 RX_TS_INFO_GET(rx_ts_info);
5449 if ((p = rx_ts_info->local_special_packet)) {
5450 rx_computelen(p, p->length);
5451 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5452 rx_ts_info->local_special_packet = p;
5453 } else { /* We won't send the ack, but don't panic. */
5454 return optionalPacket;
5458 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
5459 /* We won't send the ack, but don't panic. */
5460 return optionalPacket;
5465 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
5468 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
5469 #ifndef RX_ENABLE_TSFPQ
5470 if (!optionalPacket)
5473 return optionalPacket;
5475 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
5476 if (rx_Contiguous(p) < templ) {
5477 #ifndef RX_ENABLE_TSFPQ
5478 if (!optionalPacket)
5481 return optionalPacket;
5486 /* MTUXXX failing to send an ack is very serious. We should */
5487 /* try as hard as possible to send even a partial ack; it's */
5488 /* better than nothing. */
5489 ap = (struct rx_ackPacket *)rx_DataOf(p);
5490 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
5491 ap->reason = reason;
5493 /* The skew computation used to be bogus, I think it's better now. */
5494 /* We should start paying attention to skew. XXX */
5495 ap->serial = htonl(serial);
5496 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
5499 * First packet not yet forwarded to reader. When ACKALL has been
5500 * sent the peer has been told that all received packets will be
5501 * delivered to the reader. The value 'rnext' is used internally
5502 * to refer to the next packet in the receive queue that must be
5503 * delivered to the reader. From the perspective of the peer it
5504 * already has so report the last sequence number plus one if there
5505 * are packets in the receive queue awaiting processing.
5507 if ((call->flags & RX_CALL_ACKALL_SENT) &&
5508 !queue_IsEmpty(&call->rq)) {
5509 ap->firstPacket = htonl(queue_Last(&call->rq, rx_packet)->header.seq + 1);
5511 ap->firstPacket = htonl(call->rnext);
5513 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
5515 /* No fear of running out of ack packet here because there can only be at most
5516 * one window full of unacknowledged packets. The window size must be constrained
5517 * to be less than the maximum ack size, of course. Also, an ack should always
5518 * fit into a single packet -- it should not ever be fragmented. */
5519 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
5520 if (!rqp || !call->rq.next
5521 || (rqp->header.seq > (call->rnext + call->rwind))) {
5522 #ifndef RX_ENABLE_TSFPQ
5523 if (!optionalPacket)
5526 rxi_CallError(call, RX_CALL_DEAD);
5527 return optionalPacket;
5530 while (rqp->header.seq > call->rnext + offset)
5531 ap->acks[offset++] = RX_ACK_TYPE_NACK;
5532 ap->acks[offset++] = RX_ACK_TYPE_ACK;
5534 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
5535 #ifndef RX_ENABLE_TSFPQ
5536 if (!optionalPacket)
5539 rxi_CallError(call, RX_CALL_DEAD);
5540 return optionalPacket;
5546 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
5548 /* these are new for AFS 3.3 */
5549 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
5550 templ = htonl(templ);
5551 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
5552 templ = htonl(call->conn->peer->ifMTU);
5553 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
5554 sizeof(afs_int32), &templ);
5556 /* new for AFS 3.4 */
5557 templ = htonl(call->rwind);
5558 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
5559 sizeof(afs_int32), &templ);
5561 /* new for AFS 3.5 */
5562 templ = htonl(call->conn->peer->ifDgramPackets);
5563 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
5564 sizeof(afs_int32), &templ);
5566 p->header.serviceId = call->conn->serviceId;
5567 p->header.cid = (call->conn->cid | call->channel);
5568 p->header.callNumber = *call->callNumber;
5570 p->header.securityIndex = call->conn->securityIndex;
5571 p->header.epoch = call->conn->epoch;
5572 p->header.type = RX_PACKET_TYPE_ACK;
5573 p->header.flags = RX_SLOW_START_OK;
5574 if (reason == RX_ACK_PING) {
5575 p->header.flags |= RX_REQUEST_ACK;
5577 p->length = padbytes +
5578 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
5581 /* not fast but we can potentially use this if truncated
5582 * fragments are delivered to figure out the mtu.
5584 rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
5585 sizeof(afs_int32), sizeof(afs_int32),
5589 if (call->conn->type == RX_CLIENT_CONNECTION)
5590 p->header.flags |= RX_CLIENT_INITIATED;
5594 if (rxdebug_active) {
5598 len = _snprintf(msg, sizeof(msg),
5599 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
5600 GetCurrentThreadId(), rx_ack_reason(ap->reason),
5601 ntohl(ap->serial), ntohl(ap->previousPacket),
5602 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
5603 ap->nAcks, ntohs(ap->bufferSpace) );
5607 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
5608 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
5612 OutputDebugString(msg);
5614 #else /* AFS_NT40_ENV */
5616 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
5617 ap->reason, ntohl(ap->previousPacket),
5618 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
5620 for (offset = 0; offset < ap->nAcks; offset++)
5621 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
5626 #endif /* AFS_NT40_ENV */
5629 int i, nbytes = p->length;
5631 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
5632 if (nbytes <= p->wirevec[i].iov_len) {
5635 savelen = p->wirevec[i].iov_len;
5637 p->wirevec[i].iov_len = nbytes;
5639 rxi_Send(call, p, istack);
5640 p->wirevec[i].iov_len = savelen;
5644 nbytes -= p->wirevec[i].iov_len;
5647 if (rx_stats_active)
5648 rx_atomic_inc(&rx_stats.ackPacketsSent);
5649 #ifndef RX_ENABLE_TSFPQ
5650 if (!optionalPacket)
5653 return optionalPacket; /* Return packet for re-use by caller */
5657 struct rx_packet **list;
5662 /* Send all of the packets in the list in single datagram */
5664 rxi_SendList(struct rx_call *call, struct xmitlist *xmit,
5665 int istack, int moreFlag)
5671 struct rx_connection *conn = call->conn;
5672 struct rx_peer *peer = conn->peer;
5674 MUTEX_ENTER(&peer->peer_lock);
5675 peer->nSent += xmit->len;
5676 if (xmit->resending)
5677 peer->reSends += xmit->len;
5678 MUTEX_EXIT(&peer->peer_lock);
5680 if (rx_stats_active) {
5681 if (xmit->resending)
5682 rx_atomic_add(&rx_stats.dataPacketsReSent, xmit->len);
5684 rx_atomic_add(&rx_stats.dataPacketsSent, xmit->len);
5687 clock_GetTime(&now);
5689 if (xmit->list[xmit->len - 1]->header.flags & RX_LAST_PACKET) {
5693 /* Set the packet flags and schedule the resend events */
5694 /* Only request an ack for the last packet in the list */
5695 for (i = 0; i < xmit->len; i++) {
5696 struct rx_packet *packet = xmit->list[i];
5698 /* Record the time sent */
5699 packet->timeSent = now;
5700 packet->flags |= RX_PKTFLAG_SENT;
5702 /* Ask for an ack on retransmitted packets, on every other packet
5703 * if the peer doesn't support slow start. Ask for an ack on every
5704 * packet until the congestion window reaches the ack rate. */
5705 if (packet->header.serial) {
5708 packet->firstSent = now;
5709 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5710 || (!(call->flags & RX_CALL_SLOW_START_OK)
5711 && (packet->header.seq & 1)))) {
5716 /* Tag this packet as not being the last in this group,
5717 * for the receiver's benefit */
5718 if (i < xmit->len - 1 || moreFlag) {
5719 packet->header.flags |= RX_MORE_PACKETS;
5724 xmit->list[xmit->len - 1]->header.flags |= RX_REQUEST_ACK;
5727 /* Since we're about to send a data packet to the peer, it's
5728 * safe to nuke any scheduled end-of-packets ack */
5729 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5731 MUTEX_EXIT(&call->lock);
5732 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5733 if (xmit->len > 1) {
5734 rxi_SendPacketList(call, conn, xmit->list, xmit->len, istack);
5736 rxi_SendPacket(call, conn, xmit->list[0], istack);
5738 MUTEX_ENTER(&call->lock);
5739 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5741 /* Tell the RTO calculation engine that we have sent a packet, and
5742 * if it was the last one */
5743 rxi_rto_packet_sent(call, lastPacket, istack);
5745 /* Update last send time for this call (for keep-alive
5746 * processing), and for the connection (so that we can discover
5747 * idle connections) */
5748 conn->lastSendTime = call->lastSendTime = clock_Sec();
5749 /* Let a set of retransmits trigger an idle timeout */
5750 if (!xmit->resending)
5751 call->lastSendData = call->lastSendTime;
5754 /* When sending packets we need to follow these rules:
5755 * 1. Never send more than maxDgramPackets in a jumbogram.
5756 * 2. Never send a packet with more than two iovecs in a jumbogram.
5757 * 3. Never send a retransmitted packet in a jumbogram.
5758 * 4. Never send more than cwind/4 packets in a jumbogram
5759 * We always keep the last list we should have sent so we
5760 * can set the RX_MORE_PACKETS flags correctly.
5764 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5769 struct xmitlist working;
5770 struct xmitlist last;
5772 struct rx_peer *peer = call->conn->peer;
5773 int morePackets = 0;
5775 memset(&last, 0, sizeof(struct xmitlist));
5776 working.list = &list[0];
5778 working.resending = 0;
5780 recovery = call->flags & RX_CALL_FAST_RECOVER;
5782 for (i = 0; i < len; i++) {
5783 /* Does the current packet force us to flush the current list? */
5785 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5786 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5788 /* This sends the 'last' list and then rolls the current working
5789 * set into the 'last' one, and resets the working set */
5792 rxi_SendList(call, &last, istack, 1);
5793 /* If the call enters an error state stop sending, or if
5794 * we entered congestion recovery mode, stop sending */
5796 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5801 working.resending = 0;
5802 working.list = &list[i];
5804 /* Add the current packet to the list if it hasn't been acked.
5805 * Otherwise adjust the list pointer to skip the current packet. */
5806 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5809 if (list[i]->header.serial)
5810 working.resending = 1;
5812 /* Do we need to flush the list? */
5813 if (working.len >= (int)peer->maxDgramPackets
5814 || working.len >= (int)call->nDgramPackets
5815 || working.len >= (int)call->cwind
5816 || list[i]->header.serial
5817 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5819 rxi_SendList(call, &last, istack, 1);
5820 /* If the call enters an error state stop sending, or if
5821 * we entered congestion recovery mode, stop sending */
5823 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5828 working.resending = 0;
5829 working.list = &list[i + 1];
5832 if (working.len != 0) {
5833 osi_Panic("rxi_SendList error");
5835 working.list = &list[i + 1];
5839 /* Send the whole list when the call is in receive mode, when
5840 * the call is in eof mode, when we are in fast recovery mode,
5841 * and when we have the last packet */
5842 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5843 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5844 || (call->flags & RX_CALL_FAST_RECOVER)) {
5845 /* Check for the case where the current list contains
5846 * an acked packet. Since we always send retransmissions
5847 * in a separate packet, we only need to check the first
5848 * packet in the list */
5849 if (working.len > 0 && !(working.list[0]->flags & RX_PKTFLAG_ACKED)) {
5853 rxi_SendList(call, &last, istack, morePackets);
5854 /* If the call enters an error state stop sending, or if
5855 * we entered congestion recovery mode, stop sending */
5857 || (!recovery && (call->flags & RX_CALL_FAST_RECOVER)))
5861 rxi_SendList(call, &working, istack, 0);
5863 } else if (last.len > 0) {
5864 rxi_SendList(call, &last, istack, 0);
5865 /* Packets which are in 'working' are not sent by this call */
5870 rxi_Resend(struct rxevent *event, void *arg0, void *arg1, int istack)
5872 struct rx_call *call = arg0;
5873 struct rx_peer *peer;
5874 struct rx_packet *p, *nxp;
5875 struct clock maxTimeout = { 60, 0 };
5877 MUTEX_ENTER(&call->lock);
5879 peer = call->conn->peer;
5881 /* Make sure that the event pointer is removed from the call
5882 * structure, since there is no longer a per-call retransmission
5884 if (event == call->resendEvent) {
5885 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5886 rxevent_Put(call->resendEvent);
5887 call->resendEvent = NULL;
5890 if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
5891 rxi_CheckBusy(call);
5894 if (queue_IsEmpty(&call->tq)) {
5895 /* Nothing to do. This means that we've been raced, and that an
5896 * ACK has come in between when we were triggered, and when we
5897 * actually got to run. */
5901 /* We're in loss recovery */
5902 call->flags |= RX_CALL_FAST_RECOVER;
5904 /* Mark all of the pending packets in the queue as being lost */
5905 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5906 if (!(p->flags & RX_PKTFLAG_ACKED))
5907 p->flags &= ~RX_PKTFLAG_SENT;
5910 /* We're resending, so we double the timeout of the call. This will be
5911 * dropped back down by the first successful ACK that we receive.
5913 * We apply a maximum value here of 60 seconds
5915 clock_Add(&call->rto, &call->rto);
5916 if (clock_Gt(&call->rto, &maxTimeout))
5917 call->rto = maxTimeout;
5919 /* Packet loss is most likely due to congestion, so drop our window size
5920 * and start again from the beginning */
5921 if (peer->maxDgramPackets >1) {
5922 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5923 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5925 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5926 call->nDgramPackets = 1;
5928 call->nextCwind = 1;
5931 MUTEX_ENTER(&peer->peer_lock);
5932 peer->MTU = call->MTU;
5933 peer->cwind = call->cwind;
5934 peer->nDgramPackets = 1;
5936 call->congestSeq = peer->congestSeq;
5937 MUTEX_EXIT(&peer->peer_lock);
5939 rxi_Start(call, istack);
5942 MUTEX_EXIT(&call->lock);
5945 /* This routine is called when new packets are readied for
5946 * transmission and when retransmission may be necessary, or when the
5947 * transmission window or burst count are favourable. This should be
5948 * better optimized for new packets, the usual case, now that we've
5949 * got rid of queues of send packets. XXXXXXXXXXX */
5951 rxi_Start(struct rx_call *call, int istack)
5954 struct rx_packet *p;
5955 struct rx_packet *nxp; /* Next pointer for queue_Scan */
5960 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5961 if (rx_stats_active)
5962 rx_atomic_inc(&rx_tq_debug.rxi_start_in_error);
5967 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5969 /* Send (or resend) any packets that need it, subject to
5970 * window restrictions and congestion burst control
5971 * restrictions. Ask for an ack on the last packet sent in
5972 * this burst. For now, we're relying upon the window being
5973 * considerably bigger than the largest number of packets that
5974 * are typically sent at once by one initial call to
5975 * rxi_Start. This is probably bogus (perhaps we should ask
5976 * for an ack when we're half way through the current
5977 * window?). Also, for non file transfer applications, this
5978 * may end up asking for an ack for every packet. Bogus. XXXX
5981 * But check whether we're here recursively, and let the other guy
5984 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5985 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5986 call->flags |= RX_CALL_TQ_BUSY;
5988 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5990 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5991 call->flags &= ~RX_CALL_NEED_START;
5992 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5994 maxXmitPackets = MIN(call->twind, call->cwind);
5995 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5996 #ifdef RX_TRACK_PACKETS
5997 if ((p->flags & RX_PKTFLAG_FREE)
5998 || (!queue_IsEnd(&call->tq, nxp)
5999 && (nxp->flags & RX_PKTFLAG_FREE))
6000 || (p == (struct rx_packet *)&rx_freePacketQueue)
6001 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
6002 osi_Panic("rxi_Start: xmit queue clobbered");
6005 if (p->flags & RX_PKTFLAG_ACKED) {
6006 /* Since we may block, don't trust this */
6007 if (rx_stats_active)
6008 rx_atomic_inc(&rx_stats.ignoreAckedPacket);
6009 continue; /* Ignore this packet if it has been acknowledged */
6012 /* Turn off all flags except these ones, which are the same
6013 * on each transmission */
6014 p->header.flags &= RX_PRESET_FLAGS;
6016 if (p->header.seq >=
6017 call->tfirst + MIN((int)call->twind,
6018 (int)(call->nSoftAcked +
6020 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
6021 /* Note: if we're waiting for more window space, we can
6022 * still send retransmits; hence we don't return here, but
6023 * break out to schedule a retransmit event */
6024 dpf(("call %d waiting for window (seq %d, twind %d, nSoftAcked %d, cwind %d)\n",
6025 *(call->callNumber), p->header.seq, call->twind, call->nSoftAcked,
6030 /* Transmit the packet if it needs to be sent. */
6031 if (!(p->flags & RX_PKTFLAG_SENT)) {
6032 if (nXmitPackets == maxXmitPackets) {
6033 rxi_SendXmitList(call, call->xmitList,
6034 nXmitPackets, istack);
6037 dpf(("call %d xmit packet %"AFS_PTR_FMT"\n",
6038 *(call->callNumber), p));
6039 call->xmitList[nXmitPackets++] = p;
6043 /* xmitList now hold pointers to all of the packets that are
6044 * ready to send. Now we loop to send the packets */
6045 if (nXmitPackets > 0) {
6046 rxi_SendXmitList(call, call->xmitList, nXmitPackets,
6050 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6052 /* We went into the error state while sending packets. Now is
6053 * the time to reset the call. This will also inform the using
6054 * process that the call is in an error state.
6056 if (rx_stats_active)
6057 rx_atomic_inc(&rx_tq_debug.rxi_start_aborted);
6058 call->flags &= ~RX_CALL_TQ_BUSY;
6059 rxi_WakeUpTransmitQueue(call);
6060 rxi_CallError(call, call->error);
6063 #ifdef RX_ENABLE_LOCKS
6064 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
6066 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
6067 /* Some packets have received acks. If they all have, we can clear
6068 * the transmit queue.
6071 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
6072 if (p->header.seq < call->tfirst
6073 && (p->flags & RX_PKTFLAG_ACKED)) {
6075 #ifdef RX_TRACK_PACKETS
6076 p->flags &= ~RX_PKTFLAG_TQ;
6078 #ifdef RXDEBUG_PACKET
6086 call->flags |= RX_CALL_TQ_CLEARME;
6088 #endif /* RX_ENABLE_LOCKS */
6089 if (call->flags & RX_CALL_TQ_CLEARME)
6090 rxi_ClearTransmitQueue(call, 1);
6091 } while (call->flags & RX_CALL_NEED_START);
6093 * TQ references no longer protected by this flag; they must remain
6094 * protected by the global lock.
6096 call->flags &= ~RX_CALL_TQ_BUSY;
6097 rxi_WakeUpTransmitQueue(call);
6099 call->flags |= RX_CALL_NEED_START;
6101 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
6103 rxi_rto_cancel(call);
6107 /* Also adjusts the keep alive parameters for the call, to reflect
6108 * that we have just sent a packet (so keep alives aren't sent
6111 rxi_Send(struct rx_call *call, struct rx_packet *p,
6114 struct rx_connection *conn = call->conn;
6116 /* Stamp each packet with the user supplied status */
6117 p->header.userStatus = call->localStatus;
6119 /* Allow the security object controlling this call's security to
6120 * make any last-minute changes to the packet */
6121 RXS_SendPacket(conn->securityObject, call, p);
6123 /* Since we're about to send SOME sort of packet to the peer, it's
6124 * safe to nuke any scheduled end-of-packets ack */
6125 rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
6127 /* Actually send the packet, filling in more connection-specific fields */
6128 MUTEX_EXIT(&call->lock);
6129 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
6130 rxi_SendPacket(call, conn, p, istack);
6131 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
6132 MUTEX_ENTER(&call->lock);
6134 /* Update last send time for this call (for keep-alive
6135 * processing), and for the connection (so that we can discover
6136 * idle connections) */
6137 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6138 (((struct rx_ackPacket *)rx_DataOf(p))->reason == RX_ACK_PING) ||
6139 (p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
6141 conn->lastSendTime = call->lastSendTime = clock_Sec();
6142 /* Don't count keepalive ping/acks here, so idleness can be tracked. */
6143 if ((p->header.type != RX_PACKET_TYPE_ACK) ||
6144 ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
6145 (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
6146 RX_ACK_PING_RESPONSE)))
6147 call->lastSendData = call->lastSendTime;
6151 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
6152 * that things are fine. Also called periodically to guarantee that nothing
6153 * falls through the cracks (e.g. (error + dally) connections have keepalive
6154 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
6156 * haveCTLock Set if calling from rxi_ReapConnections
6158 #ifdef RX_ENABLE_LOCKS
6160 static rxi_CheckCall(struct rx_call *call, int haveCTLock)
6161 #else /* RX_ENABLE_LOCKS */
6163 static rxi_CheckCall(struct rx_call *call)
6164 #endif /* RX_ENABLE_LOCKS */
6166 struct rx_connection *conn = call->conn;
6168 afs_uint32 deadTime, idleDeadTime = 0, hardDeadTime = 0;
6169 afs_uint32 fudgeFactor;
6172 int idle_timeout = 0;
6173 afs_int32 clock_diff = 0;
6177 /* Large swings in the clock can have a significant impact on
6178 * the performance of RX call processing. Forward clock shifts
6179 * will result in premature event triggering or timeouts.
6180 * Backward shifts can result in calls not completing until
6181 * the clock catches up with the original start clock value.
6183 * If a backward clock shift of more than five minutes is noticed,
6184 * just fail the call.
6186 if (now < call->lastSendTime)
6187 clock_diff = call->lastSendTime - now;
6188 if (now < call->startWait)
6189 clock_diff = MAX(clock_diff, call->startWait - now);
6190 if (now < call->lastReceiveTime)
6191 clock_diff = MAX(clock_diff, call->lastReceiveTime - now);
6192 if (clock_diff > 5 * 60)
6194 if (call->state == RX_STATE_ACTIVE)
6195 rxi_CallError(call, RX_CALL_TIMEOUT);
6199 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
6200 if (call->flags & RX_CALL_TQ_BUSY) {
6201 /* Call is active and will be reset by rxi_Start if it's
6202 * in an error state.
6207 /* RTT + 8*MDEV, rounded up to the next second. */
6208 fudgeFactor = (((afs_uint32) call->rtt >> 3) +
6209 ((afs_uint32) call->rtt_dev << 1) + 1023) >> 10;
6211 deadTime = conn->secondsUntilDead + fudgeFactor;
6212 /* These are computed to the second (+- 1 second). But that's
6213 * good enough for these values, which should be a significant
6214 * number of seconds. */
6215 if (now > (call->lastReceiveTime + deadTime)) {
6216 if (call->state == RX_STATE_ACTIVE) {
6218 #if defined(KERNEL) && defined(AFS_SUN5_ENV)
6220 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6221 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
6222 ip_stack_t *ipst = ns->netstack_ip;
6224 ire = ire_cache_lookup(conn->peer->host
6225 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
6227 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
6229 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
6236 if (ire && ire->ire_max_frag > 0)
6237 rxi_SetPeerMtu(NULL, conn->peer->host, 0,
6239 #if defined(GLOBAL_NETSTACKID)
6243 #endif /* ADAPT_PMTU */
6244 cerror = RX_CALL_DEAD;
6247 #ifdef RX_ENABLE_LOCKS
6248 /* Cancel pending events */
6249 rxevent_Cancel(&call->delayedAckEvent, call,
6250 RX_CALL_REFCOUNT_DELAY);
6251 rxi_rto_cancel(call);
6252 rxevent_Cancel(&call->keepAliveEvent, call,
6253 RX_CALL_REFCOUNT_ALIVE);
6254 rxevent_Cancel(&call->growMTUEvent, call,
6255 RX_CALL_REFCOUNT_MTU);
6256 MUTEX_ENTER(&rx_refcnt_mutex);
6257 /* if rxi_FreeCall returns 1 it has freed the call */
6258 if (call->refCount == 0 &&
6259 rxi_FreeCall(call, haveCTLock))
6261 MUTEX_EXIT(&rx_refcnt_mutex);
6264 MUTEX_EXIT(&rx_refcnt_mutex);
6266 #else /* RX_ENABLE_LOCKS */
6267 rxi_FreeCall(call, 0);
6269 #endif /* RX_ENABLE_LOCKS */
6271 /* Non-active calls are destroyed if they are not responding
6272 * to pings; active calls are simply flagged in error, so the
6273 * attached process can die reasonably gracefully. */
6276 if (conn->idleDeadDetection) {
6277 if (conn->idleDeadTime) {
6278 idleDeadTime = conn->idleDeadTime + fudgeFactor;
6282 /* see if we have a non-activity timeout */
6283 if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
6284 (call->flags & RX_CALL_READER_WAIT)) {
6285 if (call->state == RX_STATE_ACTIVE) {
6286 cerror = RX_CALL_TIMEOUT;
6291 if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
6292 if (call->state == RX_STATE_ACTIVE) {
6293 cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
6301 if (conn->hardDeadTime) {
6302 hardDeadTime = conn->hardDeadTime + fudgeFactor;
6305 /* see if we have a hard timeout */
6307 && (now > (hardDeadTime + call->startTime.sec))) {
6308 if (call->state == RX_STATE_ACTIVE)
6309 rxi_CallError(call, RX_CALL_TIMEOUT);
6314 if (conn->msgsizeRetryErr && cerror != RX_CALL_TIMEOUT && !idle_timeout &&
6315 call->lastReceiveTime) {
6316 int oldMTU = conn->peer->ifMTU;
6318 /* if we thought we could send more, perhaps things got worse */
6319 if (conn->peer->maxPacketSize > conn->lastPacketSize)
6320 /* maxpacketsize will be cleared in rxi_SetPeerMtu */
6321 newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
6322 conn->lastPacketSize-(128+RX_IPUDP_SIZE));
6324 newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
6326 /* minimum capped in SetPeerMtu */
6327 rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
6330 conn->lastPacketSize = 0;
6332 /* needed so ResetCall doesn't clobber us. */
6333 call->MTU = conn->peer->ifMTU;
6335 /* if we never succeeded, let the error pass out as-is */
6336 if (conn->peer->maxPacketSize && oldMTU != conn->peer->ifMTU)
6337 cerror = conn->msgsizeRetryErr;
6340 rxi_CallError(call, cerror);
6345 rxi_NatKeepAliveEvent(struct rxevent *event, void *arg1,
6346 void *dummy, int dummy2)
6348 struct rx_connection *conn = arg1;
6349 struct rx_header theader;
6350 char tbuffer[1 + sizeof(struct rx_header)];
6351 struct sockaddr_in taddr;
6354 struct iovec tmpiov[2];
6357 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
6360 tp = &tbuffer[sizeof(struct rx_header)];
6361 taddr.sin_family = AF_INET;
6362 taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
6363 taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
6364 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6365 taddr.sin_len = sizeof(struct sockaddr_in);
6367 memset(&theader, 0, sizeof(theader));
6368 theader.epoch = htonl(999);
6370 theader.callNumber = 0;
6373 theader.type = RX_PACKET_TYPE_VERSION;
6374 theader.flags = RX_LAST_PACKET;
6375 theader.serviceId = 0;
6377 memcpy(tbuffer, &theader, sizeof(theader));
6378 memcpy(tp, &a, sizeof(a));
6379 tmpiov[0].iov_base = tbuffer;
6380 tmpiov[0].iov_len = 1 + sizeof(struct rx_header);
6382 osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
6384 MUTEX_ENTER(&conn->conn_data_lock);
6385 MUTEX_ENTER(&rx_refcnt_mutex);
6386 /* Only reschedule ourselves if the connection would not be destroyed */
6387 if (conn->refCount <= 1) {
6388 rxevent_Put(conn->natKeepAliveEvent);
6389 conn->natKeepAliveEvent = NULL;
6390 MUTEX_EXIT(&rx_refcnt_mutex);
6391 MUTEX_EXIT(&conn->conn_data_lock);
6392 rx_DestroyConnection(conn); /* drop the reference for this */
6394 conn->refCount--; /* drop the reference for this */
6395 MUTEX_EXIT(&rx_refcnt_mutex);
6396 rxevent_Put(conn->natKeepAliveEvent);
6397 conn->natKeepAliveEvent = NULL;
6398 rxi_ScheduleNatKeepAliveEvent(conn);
6399 MUTEX_EXIT(&conn->conn_data_lock);
6404 rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
6406 if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
6407 struct clock when, now;
6408 clock_GetTime(&now);
6410 when.sec += conn->secondsUntilNatPing;
6411 MUTEX_ENTER(&rx_refcnt_mutex);
6412 conn->refCount++; /* hold a reference for this */
6413 MUTEX_EXIT(&rx_refcnt_mutex);
6414 conn->natKeepAliveEvent =
6415 rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
6420 rx_SetConnSecondsUntilNatPing(struct rx_connection *conn, afs_int32 seconds)
6422 MUTEX_ENTER(&conn->conn_data_lock);
6423 conn->secondsUntilNatPing = seconds;
6425 if (!(conn->flags & RX_CONN_ATTACHWAIT))
6426 rxi_ScheduleNatKeepAliveEvent(conn);
6428 conn->flags |= RX_CONN_NAT_PING;
6430 MUTEX_EXIT(&conn->conn_data_lock);
6433 /* When a call is in progress, this routine is called occasionally to
6434 * make sure that some traffic has arrived (or been sent to) the peer.
6435 * If nothing has arrived in a reasonable amount of time, the call is
6436 * declared dead; if nothing has been sent for a while, we send a
6437 * keep-alive packet (if we're actually trying to keep the call alive)
6440 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy,
6443 struct rx_call *call = arg1;
6444 struct rx_connection *conn;
6447 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
6448 MUTEX_ENTER(&call->lock);
6450 if (event == call->keepAliveEvent) {
6451 rxevent_Put(call->keepAliveEvent);
6452 call->keepAliveEvent = NULL;
6457 #ifdef RX_ENABLE_LOCKS
6458 if (rxi_CheckCall(call, 0)) {
6459 MUTEX_EXIT(&call->lock);
6462 #else /* RX_ENABLE_LOCKS */
6463 if (rxi_CheckCall(call))
6465 #endif /* RX_ENABLE_LOCKS */
6467 /* Don't try to keep alive dallying calls */
6468 if (call->state == RX_STATE_DALLY) {
6469 MUTEX_EXIT(&call->lock);
6474 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
6475 /* Don't try to send keepalives if there is unacknowledged data */
6476 /* the rexmit code should be good enough, this little hack
6477 * doesn't quite work XXX */
6478 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
6480 rxi_ScheduleKeepAliveEvent(call);
6481 MUTEX_EXIT(&call->lock);
6484 /* Does what's on the nameplate. */
6486 rxi_GrowMTUEvent(struct rxevent *event, void *arg1, void *dummy, int dummy2)
6488 struct rx_call *call = arg1;
6489 struct rx_connection *conn;
6491 CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
6492 MUTEX_ENTER(&call->lock);
6494 if (event == call->growMTUEvent) {
6495 rxevent_Put(call->growMTUEvent);
6496 call->growMTUEvent = NULL;
6499 #ifdef RX_ENABLE_LOCKS
6500 if (rxi_CheckCall(call, 0)) {
6501 MUTEX_EXIT(&call->lock);
6504 #else /* RX_ENABLE_LOCKS */
6505 if (rxi_CheckCall(call))
6507 #endif /* RX_ENABLE_LOCKS */
6509 /* Don't bother with dallying calls */
6510 if (call->state == RX_STATE_DALLY) {
6511 MUTEX_EXIT(&call->lock);
6518 * keep being scheduled, just don't do anything if we're at peak,
6519 * or we're not set up to be properly handled (idle timeout required)
6521 if ((conn->peer->maxPacketSize != 0) &&
6522 (conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
6523 conn->idleDeadDetection)
6524 (void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
6525 rxi_ScheduleGrowMTUEvent(call, 0);
6526 MUTEX_EXIT(&call->lock);
6530 rxi_ScheduleKeepAliveEvent(struct rx_call *call)
6532 if (!call->keepAliveEvent) {
6533 struct clock when, now;
6534 clock_GetTime(&now);
6536 when.sec += call->conn->secondsUntilPing;
6537 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
6538 call->keepAliveEvent =
6539 rxevent_Post(&when, &now, rxi_KeepAliveEvent, call, NULL, 0);
6544 rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
6546 if (!call->growMTUEvent) {
6547 struct clock when, now;
6549 clock_GetTime(&now);
6552 if (call->conn->secondsUntilPing)
6553 secs = (6*call->conn->secondsUntilPing)-1;
6555 if (call->conn->secondsUntilDead)
6556 secs = MIN(secs, (call->conn->secondsUntilDead-1));
6560 CALL_HOLD(call, RX_CALL_REFCOUNT_MTU);
6561 call->growMTUEvent =
6562 rxevent_Post(&when, &now, rxi_GrowMTUEvent, call, NULL, 0);
6566 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
6568 rxi_KeepAliveOn(struct rx_call *call)
6570 /* Pretend last packet received was received now--i.e. if another
6571 * packet isn't received within the keep alive time, then the call
6572 * will die; Initialize last send time to the current time--even
6573 * if a packet hasn't been sent yet. This will guarantee that a
6574 * keep-alive is sent within the ping time */
6575 call->lastReceiveTime = call->lastSendTime = clock_Sec();
6576 rxi_ScheduleKeepAliveEvent(call);
6580 * Solely in order that callers not need to include rx_call.h
6583 rx_KeepAliveOff(struct rx_call *call)
6585 rxi_KeepAliveOff(call);
6588 rx_KeepAliveOn(struct rx_call *call)
6590 rxi_KeepAliveOn(call);
6594 rxi_GrowMTUOn(struct rx_call *call)
6596 struct rx_connection *conn = call->conn;
6597 MUTEX_ENTER(&conn->conn_data_lock);
6598 conn->lastPingSizeSer = conn->lastPingSize = 0;
6599 MUTEX_EXIT(&conn->conn_data_lock);
6600 rxi_ScheduleGrowMTUEvent(call, 1);
6603 /* This routine is called to send connection abort messages
6604 * that have been delayed to throttle looping clients. */
6606 rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1, void *unused,
6609 struct rx_connection *conn = arg1;
6612 struct rx_packet *packet;
6614 MUTEX_ENTER(&conn->conn_data_lock);
6615 rxevent_Put(conn->delayedAbortEvent);
6616 conn->delayedAbortEvent = NULL;
6617 error = htonl(conn->error);
6619 MUTEX_EXIT(&conn->conn_data_lock);
6620 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6623 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6624 RX_PACKET_TYPE_ABORT, (char *)&error,
6626 rxi_FreePacket(packet);
6630 /* This routine is called to send call abort messages
6631 * that have been delayed to throttle looping clients. */
6633 rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1, void *dummy,
6636 struct rx_call *call = arg1;
6639 struct rx_packet *packet;
6641 MUTEX_ENTER(&call->lock);
6642 rxevent_Put(call->delayedAbortEvent);
6643 call->delayedAbortEvent = NULL;
6644 error = htonl(call->error);
6646 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6649 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
6650 (char *)&error, sizeof(error), 0);
6651 rxi_FreePacket(packet);
6653 MUTEX_EXIT(&call->lock);
6654 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
6657 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
6658 * seconds) to ask the client to authenticate itself. The routine
6659 * issues a challenge to the client, which is obtained from the
6660 * security object associated with the connection */
6662 rxi_ChallengeEvent(struct rxevent *event,
6663 void *arg0, void *arg1, int tries)
6665 struct rx_connection *conn = arg0;
6668 rxevent_Put(conn->challengeEvent);
6669 conn->challengeEvent = NULL;
6672 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
6673 struct rx_packet *packet;
6674 struct clock when, now;
6677 /* We've failed to authenticate for too long.
6678 * Reset any calls waiting for authentication;
6679 * they are all in RX_STATE_PRECALL.
6683 MUTEX_ENTER(&conn->conn_call_lock);
6684 for (i = 0; i < RX_MAXCALLS; i++) {
6685 struct rx_call *call = conn->call[i];
6687 MUTEX_ENTER(&call->lock);
6688 if (call->state == RX_STATE_PRECALL) {
6689 rxi_CallError(call, RX_CALL_DEAD);
6690 rxi_SendCallAbort(call, NULL, 0, 0);
6692 MUTEX_EXIT(&call->lock);
6695 MUTEX_EXIT(&conn->conn_call_lock);
6699 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
6701 /* If there's no packet available, do this later. */
6702 RXS_GetChallenge(conn->securityObject, conn, packet);
6703 rxi_SendSpecial((struct rx_call *)0, conn, packet,
6704 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
6705 rxi_FreePacket(packet);
6707 clock_GetTime(&now);
6709 when.sec += RX_CHALLENGE_TIMEOUT;
6710 conn->challengeEvent =
6711 rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
6716 /* Call this routine to start requesting the client to authenticate
6717 * itself. This will continue until authentication is established,
6718 * the call times out, or an invalid response is returned. The
6719 * security object associated with the connection is asked to create
6720 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
6721 * defined earlier. */
6723 rxi_ChallengeOn(struct rx_connection *conn)
6725 if (!conn->challengeEvent) {
6726 RXS_CreateChallenge(conn->securityObject, conn);
6727 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
6732 /* rxi_ComputeRoundTripTime is called with peer locked. */
6733 /* peer may be null */
6735 rxi_ComputeRoundTripTime(struct rx_packet *p,
6736 struct rx_ackPacket *ack,
6737 struct rx_call *call,
6738 struct rx_peer *peer,
6741 struct clock thisRtt, *sentp;
6745 /* If the ACK is delayed, then do nothing */
6746 if (ack->reason == RX_ACK_DELAY)
6749 /* On the wire, jumbograms are a single UDP packet. We shouldn't count
6750 * their RTT multiple times, so only include the RTT of the last packet
6752 if (p->flags & RX_JUMBO_PACKET)
6755 /* Use the serial number to determine which transmission the ACK is for,
6756 * and set the sent time to match this. If we have no serial number, then
6757 * only use the ACK for RTT calculations if the packet has not been
6761 serial = ntohl(ack->serial);
6763 if (serial == p->header.serial) {
6764 sentp = &p->timeSent;
6765 } else if (serial == p->firstSerial) {
6766 sentp = &p->firstSent;
6767 } else if (clock_Eq(&p->timeSent, &p->firstSent)) {
6768 sentp = &p->firstSent;
6772 if (clock_Eq(&p->timeSent, &p->firstSent)) {
6773 sentp = &p->firstSent;
6780 if (clock_Lt(&thisRtt, sentp))
6781 return; /* somebody set the clock back, don't count this time. */
6783 clock_Sub(&thisRtt, sentp);
6784 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rttp=%d.%06d sec)\n",
6785 p->header.callNumber, p, thisRtt.sec, thisRtt.usec));
6787 if (clock_IsZero(&thisRtt)) {
6789 * The actual round trip time is shorter than the
6790 * clock_GetTime resolution. It is most likely 1ms or 100ns.
6791 * Since we can't tell which at the moment we will assume 1ms.
6793 thisRtt.usec = 1000;
6796 if (rx_stats_active) {
6797 MUTEX_ENTER(&rx_stats_mutex);
6798 if (clock_Lt(&thisRtt, &rx_stats.minRtt))
6799 rx_stats.minRtt = thisRtt;
6800 if (clock_Gt(&thisRtt, &rx_stats.maxRtt)) {
6801 if (thisRtt.sec > 60) {
6802 MUTEX_EXIT(&rx_stats_mutex);
6803 return; /* somebody set the clock ahead */
6805 rx_stats.maxRtt = thisRtt;
6807 clock_Add(&rx_stats.totalRtt, &thisRtt);
6808 rx_atomic_inc(&rx_stats.nRttSamples);
6809 MUTEX_EXIT(&rx_stats_mutex);
6812 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
6814 /* Apply VanJacobson round-trip estimations */
6819 * srtt (call->rtt) is in units of one-eighth-milliseconds.
6820 * srtt is stored as fixed point with 3 bits after the binary
6821 * point (i.e., scaled by 8). The following magic is
6822 * equivalent to the smoothing algorithm in rfc793 with an
6823 * alpha of .875 (srtt' = rtt/8 + srtt*7/8 in fixed point).
6824 * srtt'*8 = rtt + srtt*7
6825 * srtt'*8 = srtt*8 + rtt - srtt
6826 * srtt' = srtt + rtt/8 - srtt/8
6827 * srtt' = srtt + (rtt - srtt)/8
6830 delta = _8THMSEC(&thisRtt) - call->rtt;
6831 call->rtt += (delta >> 3);
6834 * We accumulate a smoothed rtt variance (actually, a smoothed
6835 * mean difference), then set the retransmit timer to smoothed
6836 * rtt + 4 times the smoothed variance (was 2x in van's original
6837 * paper, but 4x works better for me, and apparently for him as
6839 * rttvar is stored as
6840 * fixed point with 2 bits after the binary point (scaled by
6841 * 4). The following is equivalent to rfc793 smoothing with
6842 * an alpha of .75 (rttvar' = rttvar*3/4 + |delta| / 4).
6843 * rttvar'*4 = rttvar*3 + |delta|
6844 * rttvar'*4 = rttvar*4 + |delta| - rttvar
6845 * rttvar' = rttvar + |delta|/4 - rttvar/4
6846 * rttvar' = rttvar + (|delta| - rttvar)/4
6847 * This replaces rfc793's wired-in beta.
6848 * dev*4 = dev*4 + (|actual - expected| - dev)
6854 delta -= (call->rtt_dev << 1);
6855 call->rtt_dev += (delta >> 3);
6857 /* I don't have a stored RTT so I start with this value. Since I'm
6858 * probably just starting a call, and will be pushing more data down
6859 * this, I expect congestion to increase rapidly. So I fudge a
6860 * little, and I set deviance to half the rtt. In practice,
6861 * deviance tends to approach something a little less than
6862 * half the smoothed rtt. */
6863 call->rtt = _8THMSEC(&thisRtt) + 8;
6864 call->rtt_dev = call->rtt >> 2; /* rtt/2: they're scaled differently */
6866 /* the smoothed RTT time is RTT + 4*MDEV
6868 * We allow a user specified minimum to be set for this, to allow clamping
6869 * at a minimum value in the same way as TCP. In addition, we have to allow
6870 * for the possibility that this packet is answered by a delayed ACK, so we
6871 * add on a fixed 200ms to account for that timer expiring.
6874 rtt_timeout = MAX(((call->rtt >> 3) + call->rtt_dev),
6875 rx_minPeerTimeout) + 200;
6876 clock_Zero(&call->rto);
6877 clock_Addmsec(&call->rto, rtt_timeout);
6879 /* Update the peer, so any new calls start with our values */
6880 peer->rtt_dev = call->rtt_dev;
6881 peer->rtt = call->rtt;
6883 dpf(("rxi_ComputeRoundTripTime(call=%d packet=%"AFS_PTR_FMT" rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%06d sec)\n",
6884 p->header.callNumber, p, MSEC(&thisRtt), call->rtt >> 3, call->rtt_dev >> 2, (call->rto.sec), (call->rto.usec)));
6888 /* Find all server connections that have not been active for a long time, and
6891 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2,
6894 struct clock now, when;
6895 clock_GetTime(&now);
6897 /* Find server connection structures that haven't been used for
6898 * greater than rx_idleConnectionTime */
6900 struct rx_connection **conn_ptr, **conn_end;
6901 int i, havecalls = 0;
6902 MUTEX_ENTER(&rx_connHashTable_lock);
6903 for (conn_ptr = &rx_connHashTable[0], conn_end =
6904 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6906 struct rx_connection *conn, *next;
6907 struct rx_call *call;
6911 for (conn = *conn_ptr; conn; conn = next) {
6912 /* XXX -- Shouldn't the connection be locked? */
6915 for (i = 0; i < RX_MAXCALLS; i++) {
6916 call = conn->call[i];
6920 code = MUTEX_TRYENTER(&call->lock);
6923 #ifdef RX_ENABLE_LOCKS
6924 result = rxi_CheckCall(call, 1);
6925 #else /* RX_ENABLE_LOCKS */
6926 result = rxi_CheckCall(call);
6927 #endif /* RX_ENABLE_LOCKS */
6928 MUTEX_EXIT(&call->lock);
6930 /* If CheckCall freed the call, it might
6931 * have destroyed the connection as well,
6932 * which screws up the linked lists.
6938 if (conn->type == RX_SERVER_CONNECTION) {
6939 /* This only actually destroys the connection if
6940 * there are no outstanding calls */
6941 MUTEX_ENTER(&conn->conn_data_lock);
6942 MUTEX_ENTER(&rx_refcnt_mutex);
6943 if (!havecalls && !conn->refCount
6944 && ((conn->lastSendTime + rx_idleConnectionTime) <
6946 conn->refCount++; /* it will be decr in rx_DestroyConn */
6947 MUTEX_EXIT(&rx_refcnt_mutex);
6948 MUTEX_EXIT(&conn->conn_data_lock);
6949 #ifdef RX_ENABLE_LOCKS
6950 rxi_DestroyConnectionNoLock(conn);
6951 #else /* RX_ENABLE_LOCKS */
6952 rxi_DestroyConnection(conn);
6953 #endif /* RX_ENABLE_LOCKS */
6955 #ifdef RX_ENABLE_LOCKS
6957 MUTEX_EXIT(&rx_refcnt_mutex);
6958 MUTEX_EXIT(&conn->conn_data_lock);
6960 #endif /* RX_ENABLE_LOCKS */
6964 #ifdef RX_ENABLE_LOCKS
6965 while (rx_connCleanup_list) {
6966 struct rx_connection *conn;
6967 conn = rx_connCleanup_list;
6968 rx_connCleanup_list = rx_connCleanup_list->next;
6969 MUTEX_EXIT(&rx_connHashTable_lock);
6970 rxi_CleanupConnection(conn);
6971 MUTEX_ENTER(&rx_connHashTable_lock);
6973 MUTEX_EXIT(&rx_connHashTable_lock);
6974 #endif /* RX_ENABLE_LOCKS */
6977 /* Find any peer structures that haven't been used (haven't had an
6978 * associated connection) for greater than rx_idlePeerTime */
6980 struct rx_peer **peer_ptr, **peer_end;
6984 * Why do we need to hold the rx_peerHashTable_lock across
6985 * the incrementing of peer_ptr since the rx_peerHashTable
6986 * array is not changing? We don't.
6988 * By dropping the lock periodically we can permit other
6989 * activities to be performed while a rxi_ReapConnections
6990 * call is in progress. The goal of reap connections
6991 * is to clean up quickly without causing large amounts
6992 * of contention. Therefore, it is important that global
6993 * mutexes not be held for extended periods of time.
6995 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6996 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6998 struct rx_peer *peer, *next, *prev;
7000 MUTEX_ENTER(&rx_peerHashTable_lock);
7001 for (prev = peer = *peer_ptr; peer; peer = next) {
7003 code = MUTEX_TRYENTER(&peer->peer_lock);
7004 if ((code) && (peer->refCount == 0)
7005 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
7006 rx_interface_stat_p rpc_stat, nrpc_stat;
7010 * now know that this peer object is one to be
7011 * removed from the hash table. Once it is removed
7012 * it can't be referenced by other threads.
7013 * Lets remove it first and decrement the struct
7014 * nPeerStructs count.
7016 if (peer == *peer_ptr) {
7022 if (rx_stats_active)
7023 rx_atomic_dec(&rx_stats.nPeerStructs);
7026 * Now if we hold references on 'prev' and 'next'
7027 * we can safely drop the rx_peerHashTable_lock
7028 * while we destroy this 'peer' object.
7034 MUTEX_EXIT(&rx_peerHashTable_lock);
7036 MUTEX_EXIT(&peer->peer_lock);
7037 MUTEX_DESTROY(&peer->peer_lock);
7039 (&peer->rpcStats, rpc_stat, nrpc_stat,
7040 rx_interface_stat)) {
7041 unsigned int num_funcs;
7044 queue_Remove(&rpc_stat->queue_header);
7045 queue_Remove(&rpc_stat->all_peers);
7046 num_funcs = rpc_stat->stats[0].func_total;
7048 sizeof(rx_interface_stat_t) +
7049 rpc_stat->stats[0].func_total *
7050 sizeof(rx_function_entry_v1_t);
7052 rxi_Free(rpc_stat, space);
7054 MUTEX_ENTER(&rx_rpc_stats);
7055 rxi_rpc_peer_stat_cnt -= num_funcs;
7056 MUTEX_EXIT(&rx_rpc_stats);
7061 * Regain the rx_peerHashTable_lock and
7062 * decrement the reference count on 'prev'
7065 MUTEX_ENTER(&rx_peerHashTable_lock);
7072 MUTEX_EXIT(&peer->peer_lock);
7077 MUTEX_EXIT(&rx_peerHashTable_lock);
7081 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
7082 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
7083 * GC, just below. Really, we shouldn't have to keep moving packets from
7084 * one place to another, but instead ought to always know if we can
7085 * afford to hold onto a packet in its particular use. */
7086 MUTEX_ENTER(&rx_freePktQ_lock);
7087 if (rx_waitingForPackets) {
7088 rx_waitingForPackets = 0;
7089 #ifdef RX_ENABLE_LOCKS
7090 CV_BROADCAST(&rx_waitingForPackets_cv);
7092 osi_rxWakeup(&rx_waitingForPackets);
7095 MUTEX_EXIT(&rx_freePktQ_lock);
7098 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
7099 rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
7103 /* rxs_Release - This isn't strictly necessary but, since the macro name from
7104 * rx.h is sort of strange this is better. This is called with a security
7105 * object before it is discarded. Each connection using a security object has
7106 * its own refcount to the object so it won't actually be freed until the last
7107 * connection is destroyed.
7109 * This is the only rxs module call. A hold could also be written but no one
7113 rxs_Release(struct rx_securityClass *aobj)
7115 return RXS_Close(aobj);
7123 #define TRACE_OPTION_RX_DEBUG 16
7131 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
7132 0, KEY_QUERY_VALUE, &parmKey);
7133 if (code != ERROR_SUCCESS)
7136 dummyLen = sizeof(TraceOption);
7137 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
7138 (BYTE *) &TraceOption, &dummyLen);
7139 if (code == ERROR_SUCCESS) {
7140 rxdebug_active = (TraceOption & TRACE_OPTION_RX_DEBUG) ? 1 : 0;
7142 RegCloseKey (parmKey);
7143 #endif /* AFS_NT40_ENV */
7148 rx_DebugOnOff(int on)
7152 rxdebug_active = on;
7158 rx_StatsOnOff(int on)
7160 rx_stats_active = on;
7164 /* Don't call this debugging routine directly; use dpf */
7166 rxi_DebugPrint(char *format, ...)
7175 va_start(ap, format);
7177 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
7180 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
7182 OutputDebugString(msg);
7188 va_start(ap, format);
7190 clock_GetTime(&now);
7191 fprintf(rx_Log, " %d.%06d:", (unsigned int)now.sec,
7192 (unsigned int)now.usec);
7193 vfprintf(rx_Log, format, ap);
7201 * This function is used to process the rx_stats structure that is local
7202 * to a process as well as an rx_stats structure received from a remote
7203 * process (via rxdebug). Therefore, it needs to do minimal version
7207 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
7208 afs_int32 freePackets, char version)
7212 if (size != sizeof(struct rx_statistics)) {
7214 "Unexpected size of stats structure: was %d, expected %" AFS_SIZET_FMT "\n",
7215 size, sizeof(struct rx_statistics));
7218 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
7221 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7222 fprintf(file, "alloc-failures(rcv %u/%u,send %u/%u,ack %u)\n",
7223 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
7224 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
7225 s->specialPktAllocFailures);
7227 fprintf(file, "alloc-failures(rcv %u,send %u,ack %u)\n",
7228 s->receivePktAllocFailures, s->sendPktAllocFailures,
7229 s->specialPktAllocFailures);
7233 " greedy %u, " "bogusReads %u (last from host %x), "
7234 "noPackets %u, " "noBuffers %u, " "selects %u, "
7235 "sendSelects %u\n", s->socketGreedy, s->bogusPacketOnRead,
7236 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
7237 s->selects, s->sendSelects);
7239 fprintf(file, " packets read: ");
7240 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7241 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsRead[i]);
7243 fprintf(file, "\n");
7246 " other read counters: data %u, " "ack %u, " "dup %u "
7247 "spurious %u " "dally %u\n", s->dataPacketsRead,
7248 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
7249 s->ignorePacketDally);
7251 fprintf(file, " packets sent: ");
7252 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
7253 fprintf(file, "%s %u ", rx_packetTypes[i], s->packetsSent[i]);
7255 fprintf(file, "\n");
7258 " other send counters: ack %u, " "data %u (not resends), "
7259 "resends %u, " "pushed %u, " "acked&ignored %u\n",
7260 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
7261 s->dataPacketsPushed, s->ignoreAckedPacket);
7264 " \t(these should be small) sendFailed %u, " "fatalErrors %u\n",
7265 s->netSendFailures, (int)s->fatalErrors);
7267 if (s->nRttSamples) {
7268 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
7269 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
7271 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
7272 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
7276 " %d server connections, " "%d client connections, "
7277 "%d peer structs, " "%d call structs, " "%d free call structs\n",
7278 s->nServerConns, s->nClientConns, s->nPeerStructs,
7279 s->nCallStructs, s->nFreeCallStructs);
7281 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
7282 fprintf(file, " %d clock updates\n", clock_nUpdates);
7286 /* for backward compatibility */
7288 rx_PrintStats(FILE * file)
7290 MUTEX_ENTER(&rx_stats_mutex);
7291 rx_PrintTheseStats(file, (struct rx_statistics *) &rx_stats,
7292 sizeof(rx_stats), rx_nFreePackets,
7294 MUTEX_EXIT(&rx_stats_mutex);
7298 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
7300 fprintf(file, "Peer %x.%d.\n",
7301 ntohl(peer->host), (int)ntohs(peer->port));
7304 " Rtt %d, " "total sent %d, " "resent %d\n",
7305 peer->rtt, peer->nSent, peer->reSends);
7307 fprintf(file, " Packet size %d\n", peer->ifMTU);
7311 #if defined(AFS_PTHREAD_ENV) && defined(RXDEBUG)
7313 * This mutex protects the following static variables:
7317 #define LOCK_RX_DEBUG MUTEX_ENTER(&rx_debug_mutex)
7318 #define UNLOCK_RX_DEBUG MUTEX_EXIT(&rx_debug_mutex)
7320 #define LOCK_RX_DEBUG
7321 #define UNLOCK_RX_DEBUG
7322 #endif /* AFS_PTHREAD_ENV */
7324 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7326 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
7327 u_char type, void *inputData, size_t inputLength,
7328 void *outputData, size_t outputLength)
7330 static afs_int32 counter = 100;
7331 time_t waitTime, waitCount;
7332 struct rx_header theader;
7335 struct timeval tv_now, tv_wake, tv_delta;
7336 struct sockaddr_in taddr, faddr;
7350 tp = &tbuffer[sizeof(struct rx_header)];
7351 taddr.sin_family = AF_INET;
7352 taddr.sin_port = remotePort;
7353 taddr.sin_addr.s_addr = remoteAddr;
7354 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
7355 taddr.sin_len = sizeof(struct sockaddr_in);
7358 memset(&theader, 0, sizeof(theader));
7359 theader.epoch = htonl(999);
7361 theader.callNumber = htonl(counter);
7364 theader.type = type;
7365 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
7366 theader.serviceId = 0;
7368 memcpy(tbuffer, &theader, sizeof(theader));
7369 memcpy(tp, inputData, inputLength);
7371 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
7372 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
7374 /* see if there's a packet available */
7375 gettimeofday(&tv_wake, NULL);
7376 tv_wake.tv_sec += waitTime;
7379 FD_SET(socket, &imask);
7380 tv_delta.tv_sec = tv_wake.tv_sec;
7381 tv_delta.tv_usec = tv_wake.tv_usec;
7382 gettimeofday(&tv_now, NULL);
7384 if (tv_delta.tv_usec < tv_now.tv_usec) {
7386 tv_delta.tv_usec += 1000000;
7389 tv_delta.tv_usec -= tv_now.tv_usec;
7391 if (tv_delta.tv_sec < tv_now.tv_sec) {
7395 tv_delta.tv_sec -= tv_now.tv_sec;
7398 code = select(0, &imask, 0, 0, &tv_delta);
7399 #else /* AFS_NT40_ENV */
7400 code = select(socket + 1, &imask, 0, 0, &tv_delta);
7401 #endif /* AFS_NT40_ENV */
7402 if (code == 1 && FD_ISSET(socket, &imask)) {
7403 /* now receive a packet */
7404 faddrLen = sizeof(struct sockaddr_in);
7406 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
7407 (struct sockaddr *)&faddr, &faddrLen);
7410 memcpy(&theader, tbuffer, sizeof(struct rx_header));
7411 if (counter == ntohl(theader.callNumber))
7419 /* see if we've timed out */
7427 code -= sizeof(struct rx_header);
7428 if (code > outputLength)
7429 code = outputLength;
7430 memcpy(outputData, tp, code);
7433 #endif /* RXDEBUG */
7436 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
7437 afs_uint16 remotePort, struct rx_debugStats * stat,
7438 afs_uint32 * supportedValues)
7440 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7442 struct rx_debugIn in;
7444 *supportedValues = 0;
7445 in.type = htonl(RX_DEBUGI_GETSTATS);
7448 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7449 &in, sizeof(in), stat, sizeof(*stat));
7452 * If the call was successful, fixup the version and indicate
7453 * what contents of the stat structure are valid.
7454 * Also do net to host conversion of fields here.
7458 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
7459 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
7461 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
7462 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
7464 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
7465 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
7467 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
7468 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
7470 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
7471 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
7473 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
7474 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
7476 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
7477 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
7479 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
7480 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
7482 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
7483 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
7485 stat->nFreePackets = ntohl(stat->nFreePackets);
7486 stat->packetReclaims = ntohl(stat->packetReclaims);
7487 stat->callsExecuted = ntohl(stat->callsExecuted);
7488 stat->nWaiting = ntohl(stat->nWaiting);
7489 stat->idleThreads = ntohl(stat->idleThreads);
7490 stat->nWaited = ntohl(stat->nWaited);
7491 stat->nPackets = ntohl(stat->nPackets);
7500 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
7501 afs_uint16 remotePort, struct rx_statistics * stat,
7502 afs_uint32 * supportedValues)
7504 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7506 struct rx_debugIn in;
7507 afs_int32 *lp = (afs_int32 *) stat;
7511 * supportedValues is currently unused, but added to allow future
7512 * versioning of this function.
7515 *supportedValues = 0;
7516 in.type = htonl(RX_DEBUGI_RXSTATS);
7518 memset(stat, 0, sizeof(*stat));
7520 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7521 &in, sizeof(in), stat, sizeof(*stat));
7526 * Do net to host conversion here
7529 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
7540 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
7541 afs_uint16 remotePort, size_t version_length,
7544 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7546 return MakeDebugCall(socket, remoteAddr, remotePort,
7547 RX_PACKET_TYPE_VERSION, a, 1, version,
7555 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
7556 afs_uint16 remotePort, afs_int32 * nextConnection,
7557 int allConnections, afs_uint32 debugSupportedValues,
7558 struct rx_debugConn * conn,
7559 afs_uint32 * supportedValues)
7561 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7563 struct rx_debugIn in;
7567 * supportedValues is currently unused, but added to allow future
7568 * versioning of this function.
7571 *supportedValues = 0;
7572 if (allConnections) {
7573 in.type = htonl(RX_DEBUGI_GETALLCONN);
7575 in.type = htonl(RX_DEBUGI_GETCONN);
7577 in.index = htonl(*nextConnection);
7578 memset(conn, 0, sizeof(*conn));
7580 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7581 &in, sizeof(in), conn, sizeof(*conn));
7584 *nextConnection += 1;
7587 * Convert old connection format to new structure.
7590 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
7591 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
7592 #define MOVEvL(a) (conn->a = vL->a)
7594 /* any old or unrecognized version... */
7595 for (i = 0; i < RX_MAXCALLS; i++) {
7596 MOVEvL(callState[i]);
7597 MOVEvL(callMode[i]);
7598 MOVEvL(callFlags[i]);
7599 MOVEvL(callOther[i]);
7601 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
7602 MOVEvL(secStats.type);
7603 MOVEvL(secStats.level);
7604 MOVEvL(secStats.flags);
7605 MOVEvL(secStats.expires);
7606 MOVEvL(secStats.packetsReceived);
7607 MOVEvL(secStats.packetsSent);
7608 MOVEvL(secStats.bytesReceived);
7609 MOVEvL(secStats.bytesSent);
7614 * Do net to host conversion here
7616 * I don't convert host or port since we are most likely
7617 * going to want these in NBO.
7619 conn->cid = ntohl(conn->cid);
7620 conn->serial = ntohl(conn->serial);
7621 for (i = 0; i < RX_MAXCALLS; i++) {
7622 conn->callNumber[i] = ntohl(conn->callNumber[i]);
7624 conn->error = ntohl(conn->error);
7625 conn->secStats.flags = ntohl(conn->secStats.flags);
7626 conn->secStats.expires = ntohl(conn->secStats.expires);
7627 conn->secStats.packetsReceived =
7628 ntohl(conn->secStats.packetsReceived);
7629 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
7630 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
7631 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
7632 conn->epoch = ntohl(conn->epoch);
7633 conn->natMTU = ntohl(conn->natMTU);
7642 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
7643 afs_uint16 remotePort, afs_int32 * nextPeer,
7644 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
7645 afs_uint32 * supportedValues)
7647 #if defined(RXDEBUG) || defined(MAKEDEBUGCALL)
7649 struct rx_debugIn in;
7652 * supportedValues is currently unused, but added to allow future
7653 * versioning of this function.
7656 *supportedValues = 0;
7657 in.type = htonl(RX_DEBUGI_GETPEER);
7658 in.index = htonl(*nextPeer);
7659 memset(peer, 0, sizeof(*peer));
7661 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
7662 &in, sizeof(in), peer, sizeof(*peer));
7668 * Do net to host conversion here
7670 * I don't convert host or port since we are most likely
7671 * going to want these in NBO.
7673 peer->ifMTU = ntohs(peer->ifMTU);
7674 peer->idleWhen = ntohl(peer->idleWhen);
7675 peer->refCount = ntohs(peer->refCount);
7676 peer->rtt = ntohl(peer->rtt);
7677 peer->rtt_dev = ntohl(peer->rtt_dev);
7678 peer->timeout.sec = 0;
7679 peer->timeout.usec = 0;
7680 peer->nSent = ntohl(peer->nSent);
7681 peer->reSends = ntohl(peer->reSends);
7682 peer->natMTU = ntohs(peer->natMTU);
7683 peer->maxMTU = ntohs(peer->maxMTU);
7684 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
7685 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
7686 peer->MTU = ntohs(peer->MTU);
7687 peer->cwind = ntohs(peer->cwind);
7688 peer->nDgramPackets = ntohs(peer->nDgramPackets);
7689 peer->congestSeq = ntohs(peer->congestSeq);
7690 peer->bytesSent.high = ntohl(peer->bytesSent.high);
7691 peer->bytesSent.low = ntohl(peer->bytesSent.low);
7692 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
7693 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
7702 rx_GetLocalPeers(afs_uint32 peerHost, afs_uint16 peerPort,
7703 struct rx_debugPeer * peerStats)
7706 afs_int32 error = 1; /* default to "did not succeed" */
7707 afs_uint32 hashValue = PEER_HASH(peerHost, peerPort);
7709 MUTEX_ENTER(&rx_peerHashTable_lock);
7710 for(tp = rx_peerHashTable[hashValue];
7711 tp != NULL; tp = tp->next) {
7712 if (tp->host == peerHost)
7718 MUTEX_EXIT(&rx_peerHashTable_lock);
7722 MUTEX_ENTER(&tp->peer_lock);
7723 peerStats->host = tp->host;
7724 peerStats->port = tp->port;
7725 peerStats->ifMTU = tp->ifMTU;
7726 peerStats->idleWhen = tp->idleWhen;
7727 peerStats->refCount = tp->refCount;
7728 peerStats->burstSize = 0;
7729 peerStats->burst = 0;
7730 peerStats->burstWait.sec = 0;
7731 peerStats->burstWait.usec = 0;
7732 peerStats->rtt = tp->rtt;
7733 peerStats->rtt_dev = tp->rtt_dev;
7734 peerStats->timeout.sec = 0;
7735 peerStats->timeout.usec = 0;
7736 peerStats->nSent = tp->nSent;
7737 peerStats->reSends = tp->reSends;
7738 peerStats->natMTU = tp->natMTU;
7739 peerStats->maxMTU = tp->maxMTU;
7740 peerStats->maxDgramPackets = tp->maxDgramPackets;
7741 peerStats->ifDgramPackets = tp->ifDgramPackets;
7742 peerStats->MTU = tp->MTU;
7743 peerStats->cwind = tp->cwind;
7744 peerStats->nDgramPackets = tp->nDgramPackets;
7745 peerStats->congestSeq = tp->congestSeq;
7746 peerStats->bytesSent.high = tp->bytesSent >> 32;
7747 peerStats->bytesSent.low = tp->bytesSent & MAX_AFS_UINT32;
7748 peerStats->bytesReceived.high = tp->bytesReceived >> 32;
7749 peerStats->bytesReceived.low
7750 = tp->bytesReceived & MAX_AFS_UINT32;
7751 MUTEX_EXIT(&tp->peer_lock);
7753 MUTEX_ENTER(&rx_peerHashTable_lock);
7756 MUTEX_EXIT(&rx_peerHashTable_lock);
7764 struct rx_serverQueueEntry *np;
7767 struct rx_call *call;
7768 struct rx_serverQueueEntry *sq;
7772 if (rxinit_status == 1) {
7774 return; /* Already shutdown. */
7778 #ifndef AFS_PTHREAD_ENV
7779 FD_ZERO(&rx_selectMask);
7780 #endif /* AFS_PTHREAD_ENV */
7781 rxi_dataQuota = RX_MAX_QUOTA;
7782 #ifndef AFS_PTHREAD_ENV
7784 #endif /* AFS_PTHREAD_ENV */
7787 #ifndef AFS_PTHREAD_ENV
7788 #ifndef AFS_USE_GETTIMEOFDAY
7790 #endif /* AFS_USE_GETTIMEOFDAY */
7791 #endif /* AFS_PTHREAD_ENV */
7793 while (!queue_IsEmpty(&rx_freeCallQueue)) {
7794 call = queue_First(&rx_freeCallQueue, rx_call);
7796 rxi_Free(call, sizeof(struct rx_call));
7799 while (!queue_IsEmpty(&rx_idleServerQueue)) {
7800 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
7806 struct rx_peer **peer_ptr, **peer_end;
7807 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7808 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7810 struct rx_peer *peer, *next;
7812 MUTEX_ENTER(&rx_peerHashTable_lock);
7813 for (peer = *peer_ptr; peer; peer = next) {
7814 rx_interface_stat_p rpc_stat, nrpc_stat;
7817 MUTEX_ENTER(&rx_rpc_stats);
7818 MUTEX_ENTER(&peer->peer_lock);
7820 (&peer->rpcStats, rpc_stat, nrpc_stat,
7821 rx_interface_stat)) {
7822 unsigned int num_funcs;
7825 queue_Remove(&rpc_stat->queue_header);
7826 queue_Remove(&rpc_stat->all_peers);
7827 num_funcs = rpc_stat->stats[0].func_total;
7829 sizeof(rx_interface_stat_t) +
7830 rpc_stat->stats[0].func_total *
7831 sizeof(rx_function_entry_v1_t);
7833 rxi_Free(rpc_stat, space);
7835 /* rx_rpc_stats must be held */
7836 rxi_rpc_peer_stat_cnt -= num_funcs;
7838 MUTEX_EXIT(&peer->peer_lock);
7839 MUTEX_EXIT(&rx_rpc_stats);
7843 if (rx_stats_active)
7844 rx_atomic_dec(&rx_stats.nPeerStructs);
7846 MUTEX_EXIT(&rx_peerHashTable_lock);
7849 for (i = 0; i < RX_MAX_SERVICES; i++) {
7851 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
7853 for (i = 0; i < rx_hashTableSize; i++) {
7854 struct rx_connection *tc, *ntc;
7855 MUTEX_ENTER(&rx_connHashTable_lock);
7856 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
7858 for (j = 0; j < RX_MAXCALLS; j++) {
7860 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
7863 rxi_Free(tc, sizeof(*tc));
7865 MUTEX_EXIT(&rx_connHashTable_lock);
7868 MUTEX_ENTER(&freeSQEList_lock);
7870 while ((np = rx_FreeSQEList)) {
7871 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
7872 MUTEX_DESTROY(&np->lock);
7873 rxi_Free(np, sizeof(*np));
7876 MUTEX_EXIT(&freeSQEList_lock);
7877 MUTEX_DESTROY(&freeSQEList_lock);
7878 MUTEX_DESTROY(&rx_freeCallQueue_lock);
7879 MUTEX_DESTROY(&rx_connHashTable_lock);
7880 MUTEX_DESTROY(&rx_peerHashTable_lock);
7881 MUTEX_DESTROY(&rx_serverPool_lock);
7883 osi_Free(rx_connHashTable,
7884 rx_hashTableSize * sizeof(struct rx_connection *));
7885 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7887 UNPIN(rx_connHashTable,
7888 rx_hashTableSize * sizeof(struct rx_connection *));
7889 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7891 rxi_FreeAllPackets();
7893 MUTEX_ENTER(&rx_quota_mutex);
7894 rxi_dataQuota = RX_MAX_QUOTA;
7895 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7896 MUTEX_EXIT(&rx_quota_mutex);
7901 #ifdef RX_ENABLE_LOCKS
7903 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7905 if (!MUTEX_ISMINE(lockaddr))
7906 osi_Panic("Lock not held: %s", msg);
7908 #endif /* RX_ENABLE_LOCKS */
7913 * Routines to implement connection specific data.
7917 rx_KeyCreate(rx_destructor_t rtn)
7920 MUTEX_ENTER(&rxi_keyCreate_lock);
7921 key = rxi_keyCreate_counter++;
7922 rxi_keyCreate_destructor = (rx_destructor_t *)
7923 realloc((void *)rxi_keyCreate_destructor,
7924 (key + 1) * sizeof(rx_destructor_t));
7925 rxi_keyCreate_destructor[key] = rtn;
7926 MUTEX_EXIT(&rxi_keyCreate_lock);
7931 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7934 MUTEX_ENTER(&conn->conn_data_lock);
7935 if (!conn->specific) {
7936 conn->specific = malloc((key + 1) * sizeof(void *));
7937 for (i = 0; i < key; i++)
7938 conn->specific[i] = NULL;
7939 conn->nSpecific = key + 1;
7940 conn->specific[key] = ptr;
7941 } else if (key >= conn->nSpecific) {
7942 conn->specific = (void **)
7943 realloc(conn->specific, (key + 1) * sizeof(void *));
7944 for (i = conn->nSpecific; i < key; i++)
7945 conn->specific[i] = NULL;
7946 conn->nSpecific = key + 1;
7947 conn->specific[key] = ptr;
7949 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7950 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7951 conn->specific[key] = ptr;
7953 MUTEX_EXIT(&conn->conn_data_lock);
7957 rx_SetServiceSpecific(struct rx_service *svc, int key, void *ptr)
7960 MUTEX_ENTER(&svc->svc_data_lock);
7961 if (!svc->specific) {
7962 svc->specific = malloc((key + 1) * sizeof(void *));
7963 for (i = 0; i < key; i++)
7964 svc->specific[i] = NULL;
7965 svc->nSpecific = key + 1;
7966 svc->specific[key] = ptr;
7967 } else if (key >= svc->nSpecific) {
7968 svc->specific = (void **)
7969 realloc(svc->specific, (key + 1) * sizeof(void *));
7970 for (i = svc->nSpecific; i < key; i++)
7971 svc->specific[i] = NULL;
7972 svc->nSpecific = key + 1;
7973 svc->specific[key] = ptr;
7975 if (svc->specific[key] && rxi_keyCreate_destructor[key])
7976 (*rxi_keyCreate_destructor[key]) (svc->specific[key]);
7977 svc->specific[key] = ptr;
7979 MUTEX_EXIT(&svc->svc_data_lock);
7983 rx_GetSpecific(struct rx_connection *conn, int key)
7986 MUTEX_ENTER(&conn->conn_data_lock);
7987 if (key >= conn->nSpecific)
7990 ptr = conn->specific[key];
7991 MUTEX_EXIT(&conn->conn_data_lock);
7996 rx_GetServiceSpecific(struct rx_service *svc, int key)
7999 MUTEX_ENTER(&svc->svc_data_lock);
8000 if (key >= svc->nSpecific)
8003 ptr = svc->specific[key];
8004 MUTEX_EXIT(&svc->svc_data_lock);
8009 #endif /* !KERNEL */
8012 * processStats is a queue used to store the statistics for the local
8013 * process. Its contents are similar to the contents of the rpcStats
8014 * queue on a rx_peer structure, but the actual data stored within
8015 * this queue contains totals across the lifetime of the process (assuming
8016 * the stats have not been reset) - unlike the per peer structures
8017 * which can come and go based upon the peer lifetime.
8020 static struct rx_queue processStats = { &processStats, &processStats };
8023 * peerStats is a queue used to store the statistics for all peer structs.
8024 * Its contents are the union of all the peer rpcStats queues.
8027 static struct rx_queue peerStats = { &peerStats, &peerStats };
8030 * rxi_monitor_processStats is used to turn process wide stat collection
8034 static int rxi_monitor_processStats = 0;
8037 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
8040 static int rxi_monitor_peerStats = 0;
8043 * Given all of the information for a particular rpc
8044 * call, create (if needed) and update the stat totals for the rpc.
8047 * the queue of stats that will be updated with the new value
8049 * @param rxInterface
8050 * a unique number that identifies the rpc interface
8052 * @param currentFunc
8053 * the index of the function being invoked
8056 * the total number of functions in this interface
8059 * the amount of time this function waited for a thread
8062 * the amount of time this function invocation took to execute
8065 * the number bytes sent by this invocation
8068 * the number bytes received by this invocation
8071 * if true, this invocation was made to a server
8074 * the ip address of the remote host
8077 * the port of the remote host
8079 * @param addToPeerList
8080 * if != 0, add newly created stat to the global peer list
8083 * if a new stats structure is allocated, the counter will
8084 * be updated with the new number of allocated stat structures
8089 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
8090 afs_uint32 currentFunc, afs_uint32 totalFunc,
8091 struct clock *queueTime, struct clock *execTime,
8092 afs_uint64 bytesSent, afs_uint64 bytesRcvd, int isServer,
8093 afs_uint32 remoteHost, afs_uint32 remotePort,
8094 int addToPeerList, unsigned int *counter)
8097 rx_interface_stat_p rpc_stat, nrpc_stat;
8100 * See if there's already a structure for this interface
8103 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8104 if ((rpc_stat->stats[0].interfaceId == rxInterface)
8105 && (rpc_stat->stats[0].remote_is_server == isServer))
8110 * Didn't find a match so allocate a new structure and add it to the
8114 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
8115 || (rpc_stat->stats[0].interfaceId != rxInterface)
8116 || (rpc_stat->stats[0].remote_is_server != isServer)) {
8121 sizeof(rx_interface_stat_t) +
8122 totalFunc * sizeof(rx_function_entry_v1_t);
8124 rpc_stat = rxi_Alloc(space);
8125 if (rpc_stat == NULL) {
8129 *counter += totalFunc;
8130 for (i = 0; i < totalFunc; i++) {
8131 rpc_stat->stats[i].remote_peer = remoteHost;
8132 rpc_stat->stats[i].remote_port = remotePort;
8133 rpc_stat->stats[i].remote_is_server = isServer;
8134 rpc_stat->stats[i].interfaceId = rxInterface;
8135 rpc_stat->stats[i].func_total = totalFunc;
8136 rpc_stat->stats[i].func_index = i;
8137 rpc_stat->stats[i].invocations = 0;
8138 rpc_stat->stats[i].bytes_sent = 0;
8139 rpc_stat->stats[i].bytes_rcvd = 0;
8140 rpc_stat->stats[i].queue_time_sum.sec = 0;
8141 rpc_stat->stats[i].queue_time_sum.usec = 0;
8142 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8143 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8144 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8145 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8146 rpc_stat->stats[i].queue_time_max.sec = 0;
8147 rpc_stat->stats[i].queue_time_max.usec = 0;
8148 rpc_stat->stats[i].execution_time_sum.sec = 0;
8149 rpc_stat->stats[i].execution_time_sum.usec = 0;
8150 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8151 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8152 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8153 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8154 rpc_stat->stats[i].execution_time_max.sec = 0;
8155 rpc_stat->stats[i].execution_time_max.usec = 0;
8157 queue_Prepend(stats, rpc_stat);
8158 if (addToPeerList) {
8159 queue_Prepend(&peerStats, &rpc_stat->all_peers);
8164 * Increment the stats for this function
8167 rpc_stat->stats[currentFunc].invocations++;
8168 rpc_stat->stats[currentFunc].bytes_sent += bytesSent;
8169 rpc_stat->stats[currentFunc].bytes_rcvd += bytesRcvd;
8170 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
8171 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
8172 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
8173 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
8175 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
8176 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
8178 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
8179 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
8181 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
8182 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
8184 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
8185 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
8193 rxi_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8194 afs_uint32 currentFunc, afs_uint32 totalFunc,
8195 struct clock *queueTime, struct clock *execTime,
8196 afs_uint64 bytesSent, afs_uint64 bytesRcvd,
8200 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
8203 MUTEX_ENTER(&rx_rpc_stats);
8205 if (rxi_monitor_peerStats) {
8206 MUTEX_ENTER(&peer->peer_lock);
8207 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
8208 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8209 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
8210 MUTEX_EXIT(&peer->peer_lock);
8213 if (rxi_monitor_processStats) {
8214 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
8215 queueTime, execTime, bytesSent, bytesRcvd, isServer,
8216 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
8219 MUTEX_EXIT(&rx_rpc_stats);
8223 * Increment the times and count for a particular rpc function.
8225 * Traditionally this call was invoked from rxgen stubs. Modern stubs
8226 * call rx_RecordCallStatistics instead, so the public version of this
8227 * function is left purely for legacy callers.
8230 * The peer who invoked the rpc
8232 * @param rxInterface
8233 * A unique number that identifies the rpc interface
8235 * @param currentFunc
8236 * The index of the function being invoked
8239 * The total number of functions in this interface
8242 * The amount of time this function waited for a thread
8245 * The amount of time this function invocation took to execute
8248 * The number bytes sent by this invocation
8251 * The number bytes received by this invocation
8254 * If true, this invocation was made to a server
8258 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
8259 afs_uint32 currentFunc, afs_uint32 totalFunc,
8260 struct clock *queueTime, struct clock *execTime,
8261 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
8267 sent64 = ((afs_uint64)bytesSent->high << 32) + bytesSent->low;
8268 rcvd64 = ((afs_uint64)bytesRcvd->high << 32) + bytesRcvd->low;
8270 rxi_IncrementTimeAndCount(peer, rxInterface, currentFunc, totalFunc,
8271 queueTime, execTime, sent64, rcvd64,
8278 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
8282 * IN callerVersion - the rpc stat version of the caller.
8284 * IN count - the number of entries to marshall.
8286 * IN stats - pointer to stats to be marshalled.
8288 * OUT ptr - Where to store the marshalled data.
8295 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
8296 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
8302 * We only support the first version
8304 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
8305 *(ptr++) = stats->remote_peer;
8306 *(ptr++) = stats->remote_port;
8307 *(ptr++) = stats->remote_is_server;
8308 *(ptr++) = stats->interfaceId;
8309 *(ptr++) = stats->func_total;
8310 *(ptr++) = stats->func_index;
8311 *(ptr++) = stats->invocations >> 32;
8312 *(ptr++) = stats->invocations & MAX_AFS_UINT32;
8313 *(ptr++) = stats->bytes_sent >> 32;
8314 *(ptr++) = stats->bytes_sent & MAX_AFS_UINT32;
8315 *(ptr++) = stats->bytes_rcvd >> 32;
8316 *(ptr++) = stats->bytes_rcvd & MAX_AFS_UINT32;
8317 *(ptr++) = stats->queue_time_sum.sec;
8318 *(ptr++) = stats->queue_time_sum.usec;
8319 *(ptr++) = stats->queue_time_sum_sqr.sec;
8320 *(ptr++) = stats->queue_time_sum_sqr.usec;
8321 *(ptr++) = stats->queue_time_min.sec;
8322 *(ptr++) = stats->queue_time_min.usec;
8323 *(ptr++) = stats->queue_time_max.sec;
8324 *(ptr++) = stats->queue_time_max.usec;
8325 *(ptr++) = stats->execution_time_sum.sec;
8326 *(ptr++) = stats->execution_time_sum.usec;
8327 *(ptr++) = stats->execution_time_sum_sqr.sec;
8328 *(ptr++) = stats->execution_time_sum_sqr.usec;
8329 *(ptr++) = stats->execution_time_min.sec;
8330 *(ptr++) = stats->execution_time_min.usec;
8331 *(ptr++) = stats->execution_time_max.sec;
8332 *(ptr++) = stats->execution_time_max.usec;
8338 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
8343 * IN callerVersion - the rpc stat version of the caller
8345 * OUT myVersion - the rpc stat version of this function
8347 * OUT clock_sec - local time seconds
8349 * OUT clock_usec - local time microseconds
8351 * OUT allocSize - the number of bytes allocated to contain stats
8353 * OUT statCount - the number stats retrieved from this process.
8355 * OUT stats - the actual stats retrieved from this process.
8359 * Returns void. If successful, stats will != NULL.
8363 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8364 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8365 size_t * allocSize, afs_uint32 * statCount,
8366 afs_uint32 ** stats)
8376 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8379 * Check to see if stats are enabled
8382 MUTEX_ENTER(&rx_rpc_stats);
8383 if (!rxi_monitor_processStats) {
8384 MUTEX_EXIT(&rx_rpc_stats);
8388 clock_GetTime(&now);
8389 *clock_sec = now.sec;
8390 *clock_usec = now.usec;
8393 * Allocate the space based upon the caller version
8395 * If the client is at an older version than we are,
8396 * we return the statistic data in the older data format, but
8397 * we still return our version number so the client knows we
8398 * are maintaining more data than it can retrieve.
8401 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8402 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
8403 *statCount = rxi_rpc_process_stat_cnt;
8406 * This can't happen yet, but in the future version changes
8407 * can be handled by adding additional code here
8411 if (space > (size_t) 0) {
8413 ptr = *stats = rxi_Alloc(space);
8416 rx_interface_stat_p rpc_stat, nrpc_stat;
8420 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8422 * Copy the data based upon the caller version
8424 rx_MarshallProcessRPCStats(callerVersion,
8425 rpc_stat->stats[0].func_total,
8426 rpc_stat->stats, &ptr);
8432 MUTEX_EXIT(&rx_rpc_stats);
8437 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
8441 * IN callerVersion - the rpc stat version of the caller
8443 * OUT myVersion - the rpc stat version of this function
8445 * OUT clock_sec - local time seconds
8447 * OUT clock_usec - local time microseconds
8449 * OUT allocSize - the number of bytes allocated to contain stats
8451 * OUT statCount - the number of stats retrieved from the individual
8454 * OUT stats - the actual stats retrieved from the individual peer structures.
8458 * Returns void. If successful, stats will != NULL.
8462 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
8463 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
8464 size_t * allocSize, afs_uint32 * statCount,
8465 afs_uint32 ** stats)
8475 *myVersion = RX_STATS_RETRIEVAL_VERSION;
8478 * Check to see if stats are enabled
8481 MUTEX_ENTER(&rx_rpc_stats);
8482 if (!rxi_monitor_peerStats) {
8483 MUTEX_EXIT(&rx_rpc_stats);
8487 clock_GetTime(&now);
8488 *clock_sec = now.sec;
8489 *clock_usec = now.usec;
8492 * Allocate the space based upon the caller version
8494 * If the client is at an older version than we are,
8495 * we return the statistic data in the older data format, but
8496 * we still return our version number so the client knows we
8497 * are maintaining more data than it can retrieve.
8500 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
8501 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
8502 *statCount = rxi_rpc_peer_stat_cnt;
8505 * This can't happen yet, but in the future version changes
8506 * can be handled by adding additional code here
8510 if (space > (size_t) 0) {
8512 ptr = *stats = rxi_Alloc(space);
8515 rx_interface_stat_p rpc_stat, nrpc_stat;
8519 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8521 * We have to fix the offset of rpc_stat since we are
8522 * keeping this structure on two rx_queues. The rx_queue
8523 * package assumes that the rx_queue member is the first
8524 * member of the structure. That is, rx_queue assumes that
8525 * any one item is only on one queue at a time. We are
8526 * breaking that assumption and so we have to do a little
8527 * math to fix our pointers.
8530 fix_offset = (char *)rpc_stat;
8531 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8532 rpc_stat = (rx_interface_stat_p) fix_offset;
8535 * Copy the data based upon the caller version
8537 rx_MarshallProcessRPCStats(callerVersion,
8538 rpc_stat->stats[0].func_total,
8539 rpc_stat->stats, &ptr);
8545 MUTEX_EXIT(&rx_rpc_stats);
8550 * rx_FreeRPCStats - free memory allocated by
8551 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
8555 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
8556 * rx_RetrievePeerRPCStats
8558 * IN allocSize - the number of bytes in stats.
8566 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
8568 rxi_Free(stats, allocSize);
8572 * rx_queryProcessRPCStats - see if process rpc stat collection is
8573 * currently enabled.
8579 * Returns 0 if stats are not enabled != 0 otherwise
8583 rx_queryProcessRPCStats(void)
8586 MUTEX_ENTER(&rx_rpc_stats);
8587 rc = rxi_monitor_processStats;
8588 MUTEX_EXIT(&rx_rpc_stats);
8593 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
8599 * Returns 0 if stats are not enabled != 0 otherwise
8603 rx_queryPeerRPCStats(void)
8606 MUTEX_ENTER(&rx_rpc_stats);
8607 rc = rxi_monitor_peerStats;
8608 MUTEX_EXIT(&rx_rpc_stats);
8613 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
8623 rx_enableProcessRPCStats(void)
8625 MUTEX_ENTER(&rx_rpc_stats);
8626 rx_enable_stats = 1;
8627 rxi_monitor_processStats = 1;
8628 MUTEX_EXIT(&rx_rpc_stats);
8632 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
8642 rx_enablePeerRPCStats(void)
8644 MUTEX_ENTER(&rx_rpc_stats);
8645 rx_enable_stats = 1;
8646 rxi_monitor_peerStats = 1;
8647 MUTEX_EXIT(&rx_rpc_stats);
8651 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
8661 rx_disableProcessRPCStats(void)
8663 rx_interface_stat_p rpc_stat, nrpc_stat;
8666 MUTEX_ENTER(&rx_rpc_stats);
8669 * Turn off process statistics and if peer stats is also off, turn
8673 rxi_monitor_processStats = 0;
8674 if (rxi_monitor_peerStats == 0) {
8675 rx_enable_stats = 0;
8678 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8679 unsigned int num_funcs = 0;
8682 queue_Remove(rpc_stat);
8683 num_funcs = rpc_stat->stats[0].func_total;
8685 sizeof(rx_interface_stat_t) +
8686 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
8688 rxi_Free(rpc_stat, space);
8689 rxi_rpc_process_stat_cnt -= num_funcs;
8691 MUTEX_EXIT(&rx_rpc_stats);
8695 * rx_disablePeerRPCStats - stop rpc stat collection for peers
8705 rx_disablePeerRPCStats(void)
8707 struct rx_peer **peer_ptr, **peer_end;
8711 * Turn off peer statistics and if process stats is also off, turn
8715 rxi_monitor_peerStats = 0;
8716 if (rxi_monitor_processStats == 0) {
8717 rx_enable_stats = 0;
8720 for (peer_ptr = &rx_peerHashTable[0], peer_end =
8721 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
8723 struct rx_peer *peer, *next, *prev;
8725 MUTEX_ENTER(&rx_peerHashTable_lock);
8726 MUTEX_ENTER(&rx_rpc_stats);
8727 for (prev = peer = *peer_ptr; peer; peer = next) {
8729 code = MUTEX_TRYENTER(&peer->peer_lock);
8731 rx_interface_stat_p rpc_stat, nrpc_stat;
8734 if (prev == *peer_ptr) {
8745 MUTEX_EXIT(&rx_peerHashTable_lock);
8748 (&peer->rpcStats, rpc_stat, nrpc_stat,
8749 rx_interface_stat)) {
8750 unsigned int num_funcs = 0;
8753 queue_Remove(&rpc_stat->queue_header);
8754 queue_Remove(&rpc_stat->all_peers);
8755 num_funcs = rpc_stat->stats[0].func_total;
8757 sizeof(rx_interface_stat_t) +
8758 rpc_stat->stats[0].func_total *
8759 sizeof(rx_function_entry_v1_t);
8761 rxi_Free(rpc_stat, space);
8762 rxi_rpc_peer_stat_cnt -= num_funcs;
8764 MUTEX_EXIT(&peer->peer_lock);
8766 MUTEX_ENTER(&rx_peerHashTable_lock);
8776 MUTEX_EXIT(&rx_rpc_stats);
8777 MUTEX_EXIT(&rx_peerHashTable_lock);
8782 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
8787 * IN clearFlag - flag indicating which stats to clear
8795 rx_clearProcessRPCStats(afs_uint32 clearFlag)
8797 rx_interface_stat_p rpc_stat, nrpc_stat;
8799 MUTEX_ENTER(&rx_rpc_stats);
8801 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8802 unsigned int num_funcs = 0, i;
8803 num_funcs = rpc_stat->stats[0].func_total;
8804 for (i = 0; i < num_funcs; i++) {
8805 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8806 rpc_stat->stats[i].invocations = 0;
8808 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8809 rpc_stat->stats[i].bytes_sent = 0;
8811 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8812 rpc_stat->stats[i].bytes_rcvd = 0;
8814 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8815 rpc_stat->stats[i].queue_time_sum.sec = 0;
8816 rpc_stat->stats[i].queue_time_sum.usec = 0;
8818 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8819 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8820 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8822 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8823 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8824 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8826 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8827 rpc_stat->stats[i].queue_time_max.sec = 0;
8828 rpc_stat->stats[i].queue_time_max.usec = 0;
8830 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8831 rpc_stat->stats[i].execution_time_sum.sec = 0;
8832 rpc_stat->stats[i].execution_time_sum.usec = 0;
8834 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8835 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8836 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8838 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8839 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8840 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8842 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8843 rpc_stat->stats[i].execution_time_max.sec = 0;
8844 rpc_stat->stats[i].execution_time_max.usec = 0;
8849 MUTEX_EXIT(&rx_rpc_stats);
8853 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
8858 * IN clearFlag - flag indicating which stats to clear
8866 rx_clearPeerRPCStats(afs_uint32 clearFlag)
8868 rx_interface_stat_p rpc_stat, nrpc_stat;
8870 MUTEX_ENTER(&rx_rpc_stats);
8872 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
8873 unsigned int num_funcs = 0, i;
8876 * We have to fix the offset of rpc_stat since we are
8877 * keeping this structure on two rx_queues. The rx_queue
8878 * package assumes that the rx_queue member is the first
8879 * member of the structure. That is, rx_queue assumes that
8880 * any one item is only on one queue at a time. We are
8881 * breaking that assumption and so we have to do a little
8882 * math to fix our pointers.
8885 fix_offset = (char *)rpc_stat;
8886 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
8887 rpc_stat = (rx_interface_stat_p) fix_offset;
8889 num_funcs = rpc_stat->stats[0].func_total;
8890 for (i = 0; i < num_funcs; i++) {
8891 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
8892 rpc_stat->stats[i].invocations = 0;
8894 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
8895 rpc_stat->stats[i].bytes_sent = 0;
8897 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
8898 rpc_stat->stats[i].bytes_rcvd = 0;
8900 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
8901 rpc_stat->stats[i].queue_time_sum.sec = 0;
8902 rpc_stat->stats[i].queue_time_sum.usec = 0;
8904 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
8905 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
8906 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
8908 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
8909 rpc_stat->stats[i].queue_time_min.sec = 9999999;
8910 rpc_stat->stats[i].queue_time_min.usec = 9999999;
8912 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
8913 rpc_stat->stats[i].queue_time_max.sec = 0;
8914 rpc_stat->stats[i].queue_time_max.usec = 0;
8916 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
8917 rpc_stat->stats[i].execution_time_sum.sec = 0;
8918 rpc_stat->stats[i].execution_time_sum.usec = 0;
8920 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
8921 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
8922 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
8924 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
8925 rpc_stat->stats[i].execution_time_min.sec = 9999999;
8926 rpc_stat->stats[i].execution_time_min.usec = 9999999;
8928 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
8929 rpc_stat->stats[i].execution_time_max.sec = 0;
8930 rpc_stat->stats[i].execution_time_max.usec = 0;
8935 MUTEX_EXIT(&rx_rpc_stats);
8939 * rxi_rxstat_userok points to a routine that returns 1 if the caller
8940 * is authorized to enable/disable/clear RX statistics.
8942 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
8945 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
8947 rxi_rxstat_userok = proc;
8951 rx_RxStatUserOk(struct rx_call *call)
8953 if (!rxi_rxstat_userok)
8955 return rxi_rxstat_userok(call);
8960 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
8961 * function in the MSVC runtime DLL (msvcrt.dll).
8963 * Note: the system serializes calls to this function.
8966 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
8967 DWORD reason, /* reason function is being called */
8968 LPVOID reserved) /* reserved for future use */
8971 case DLL_PROCESS_ATTACH:
8972 /* library is being attached to a process */
8976 case DLL_PROCESS_DETACH:
8983 #endif /* AFS_NT40_ENV */
8986 int rx_DumpCalls(FILE *outputFile, char *cookie)
8988 #ifdef RXDEBUG_PACKET
8989 #ifdef KDUMP_RX_LOCK
8990 struct rx_call_rx_lock *c;
8997 #define RXDPRINTF sprintf
8998 #define RXDPRINTOUT output
9000 #define RXDPRINTF fprintf
9001 #define RXDPRINTOUT outputFile
9004 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
9006 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9009 for (c = rx_allCallsp; c; c = c->allNextp) {
9010 u_short rqc, tqc, iovqc;
9011 struct rx_packet *p, *np;
9013 MUTEX_ENTER(&c->lock);
9014 queue_Count(&c->rq, p, np, rx_packet, rqc);
9015 queue_Count(&c->tq, p, np, rx_packet, tqc);
9016 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
9018 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, "
9019 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
9020 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
9021 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
9022 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
9023 #ifdef RX_ENABLE_LOCKS
9026 #ifdef RX_REFCOUNT_CHECK
9027 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
9028 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
9031 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
9032 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
9033 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
9034 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
9035 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
9036 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
9037 #ifdef RX_ENABLE_LOCKS
9038 , (afs_uint32)c->refCount
9040 #ifdef RX_REFCOUNT_CHECK
9041 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
9044 MUTEX_EXIT(&c->lock);
9047 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9050 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Calls\r\n", cookie);
9052 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
9054 #endif /* RXDEBUG_PACKET */