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 #include <afsconfig.h>
12 #include "afs/param.h"
14 #include <afs/param.h>
20 #include "afs/sysincludes.h"
21 #include "afsincludes.h"
22 #include "rx/rx_kcommon.h"
23 #include "rx/rx_clock.h"
24 #include "rx/rx_queue.h"
25 #include "rx/rx_packet.h"
26 #else /* defined(UKERNEL) */
27 #ifdef RX_KERNEL_TRACE
28 #include "../rx/rx_kcommon.h"
31 #ifndef AFS_LINUX20_ENV
34 #if defined(AFS_SGI_ENV) || defined(AFS_HPUX110_ENV)
35 #include "afs/sysincludes.h"
37 #if defined(AFS_OBSD_ENV)
41 #if !defined(AFS_SUN5_ENV) && !defined(AFS_LINUX20_ENV) && !defined(AFS_HPUX110_ENV)
42 #if !defined(AFS_OSF_ENV) && !defined(AFS_AIX41_ENV)
43 #include "sys/mount.h" /* it gets pulled in by something later anyway */
47 #include "netinet/in.h"
48 #include "afs/afs_osi.h"
49 #include "rx_kmutex.h"
50 #include "rx/rx_clock.h"
51 #include "rx/rx_queue.h"
53 #include <sys/sysmacros.h>
55 #include "rx/rx_packet.h"
56 #endif /* defined(UKERNEL) */
57 #include "rx/rx_globals.h"
59 #include "sys/types.h"
62 #if defined(AFS_NT40_ENV)
65 #define EWOULDBLOCK WSAEWOULDBLOCK
68 #include "rx_xmit_nt.h"
71 #include <sys/socket.h>
72 #include <netinet/in.h>
78 #include <sys/sysmacros.h>
80 #include "rx_packet.h"
81 #include "rx_globals.h"
91 /* rxdb_fileID is used to identify the lock location, along with line#. */
92 static int rxdb_fileID = RXDB_FILE_RX_PACKET;
93 #endif /* RX_LOCKS_DB */
94 static struct rx_packet *rx_mallocedP = 0;
96 static afs_uint32 rx_packet_id = 0;
99 extern char cml_version_number[];
101 static int AllocPacketBufs(int class, int num_pkts, struct rx_queue *q);
103 static void rxi_SendDebugPacket(struct rx_packet *apacket, osi_socket asocket,
104 afs_int32 ahost, short aport,
107 #ifdef RX_ENABLE_TSFPQ
109 rxi_FreeDataBufsTSFPQ(struct rx_packet *p, afs_uint32 first, int flush_global);
111 static int rxi_FreeDataBufsToQueue(struct rx_packet *p,
113 struct rx_queue * q);
116 /* some rules about packets:
117 * 1. When a packet is allocated, the final iov_buf contains room for
118 * a security trailer, but iov_len masks that fact. If the security
119 * package wants to add the trailer, it may do so, and then extend
120 * iov_len appropriately. For this reason, packet's niovecs and
121 * iov_len fields should be accurate before calling PreparePacket.
125 * all packet buffers (iov_base) are integral multiples of
127 * offset is an integral multiple of the word size.
130 rx_SlowGetInt32(struct rx_packet *packet, size_t offset)
134 for (l = 0, i = 1; i < packet->niovecs; i++) {
135 if (l + packet->wirevec[i].iov_len > offset) {
137 *((afs_int32 *) ((char *)(packet->wirevec[i].iov_base) +
140 l += packet->wirevec[i].iov_len;
147 * all packet buffers (iov_base) are integral multiples of the word size.
148 * offset is an integral multiple of the word size.
151 rx_SlowPutInt32(struct rx_packet * packet, size_t offset, afs_int32 data)
155 for (l = 0, i = 1; i < packet->niovecs; i++) {
156 if (l + packet->wirevec[i].iov_len > offset) {
157 *((afs_int32 *) ((char *)(packet->wirevec[i].iov_base) +
158 (offset - l))) = data;
161 l += packet->wirevec[i].iov_len;
168 * all packet buffers (iov_base) are integral multiples of the
170 * offset is an integral multiple of the word size.
172 * all buffers are contiguously arrayed in the iovec from 0..niovecs-1
175 rx_SlowReadPacket(struct rx_packet * packet, unsigned int offset, int resid,
178 unsigned int i, j, l, r;
179 for (l = 0, i = 1; i < packet->niovecs; i++) {
180 if (l + packet->wirevec[i].iov_len > offset) {
183 l += packet->wirevec[i].iov_len;
186 /* i is the iovec which contains the first little bit of data in which we
187 * are interested. l is the total length of everything prior to this iovec.
188 * j is the number of bytes we can safely copy out of this iovec.
189 * offset only applies to the first iovec.
192 while ((r > 0) && (i < packet->niovecs)) {
193 j = MIN(r, packet->wirevec[i].iov_len - (offset - l));
194 memcpy(out, (char *)(packet->wirevec[i].iov_base) + (offset - l), j);
197 l += packet->wirevec[i].iov_len;
202 return (r ? (resid - r) : resid);
207 * all packet buffers (iov_base) are integral multiples of the
209 * offset is an integral multiple of the word size.
212 rx_SlowWritePacket(struct rx_packet * packet, int offset, int resid, char *in)
214 unsigned int i, j, l, o, r;
217 for (l = 0, i = 1, o = offset; i < packet->niovecs; i++) {
218 if (l + packet->wirevec[i].iov_len > o) {
221 l += packet->wirevec[i].iov_len;
224 /* i is the iovec which contains the first little bit of data in which we
225 * are interested. l is the total length of everything prior to this iovec.
226 * j is the number of bytes we can safely copy out of this iovec.
227 * offset only applies to the first iovec.
230 while ((r > 0) && (i <= RX_MAXWVECS)) {
231 if (i >= packet->niovecs)
232 if (rxi_AllocDataBuf(packet, r, RX_PACKET_CLASS_SEND_CBUF) > 0) /* ++niovecs as a side-effect */
235 b = (char *)(packet->wirevec[i].iov_base) + (offset - l);
236 j = MIN(r, packet->wirevec[i].iov_len - (offset - l));
240 l += packet->wirevec[i].iov_len;
245 return (r ? (resid - r) : resid);
249 rxi_AllocPackets(int class, int num_pkts, struct rx_queue * q)
251 struct rx_packet *p, *np;
253 num_pkts = AllocPacketBufs(class, num_pkts, q);
255 for (queue_Scan(q, p, np, rx_packet)) {
256 RX_PACKET_IOV_FULLINIT(p);
262 #ifdef RX_ENABLE_TSFPQ
264 AllocPacketBufs(int class, int num_pkts, struct rx_queue * q)
266 struct rx_ts_info_t * rx_ts_info;
270 RX_TS_INFO_GET(rx_ts_info);
272 transfer = num_pkts - rx_ts_info->_FPQ.len;
275 MUTEX_ENTER(&rx_freePktQ_lock);
276 transfer = MAX(transfer, rx_TSFPQGlobSize);
277 if (transfer > rx_nFreePackets) {
278 /* alloc enough for us, plus a few globs for other threads */
279 rxi_MorePacketsNoLock(transfer + 4 * rx_initSendWindow);
282 RX_TS_FPQ_GTOL2(rx_ts_info, transfer);
284 MUTEX_EXIT(&rx_freePktQ_lock);
288 RX_TS_FPQ_QCHECKOUT(rx_ts_info, num_pkts, q);
292 #else /* RX_ENABLE_TSFPQ */
294 AllocPacketBufs(int class, int num_pkts, struct rx_queue * q)
305 MUTEX_ENTER(&rx_freePktQ_lock);
308 for (; (num_pkts > 0) && (rxi_OverQuota2(class,num_pkts));
309 num_pkts--, overq++);
312 rxi_NeedMorePackets = TRUE;
313 if (rx_stats_active) {
315 case RX_PACKET_CLASS_RECEIVE:
316 rx_MutexIncrement(rx_stats.receivePktAllocFailures, rx_stats_mutex);
318 case RX_PACKET_CLASS_SEND:
319 rx_MutexIncrement(rx_stats.sendPktAllocFailures, rx_stats_mutex);
321 case RX_PACKET_CLASS_SPECIAL:
322 rx_MutexIncrement(rx_stats.specialPktAllocFailures, rx_stats_mutex);
324 case RX_PACKET_CLASS_RECV_CBUF:
325 rx_MutexIncrement(rx_stats.receiveCbufPktAllocFailures, rx_stats_mutex);
327 case RX_PACKET_CLASS_SEND_CBUF:
328 rx_MutexIncrement(rx_stats.sendCbufPktAllocFailures, rx_stats_mutex);
334 if (rx_nFreePackets < num_pkts)
335 num_pkts = rx_nFreePackets;
338 rxi_NeedMorePackets = TRUE;
342 if (rx_nFreePackets < num_pkts) {
343 rxi_MorePacketsNoLock(MAX((num_pkts-rx_nFreePackets), 4 * rx_initSendWindow));
347 for (i=0, c=queue_First(&rx_freePacketQueue, rx_packet);
349 i++, c=queue_Next(c, rx_packet)) {
353 queue_SplitBeforeAppend(&rx_freePacketQueue,q,c);
355 rx_nFreePackets -= num_pkts;
360 MUTEX_EXIT(&rx_freePktQ_lock);
365 #endif /* RX_ENABLE_TSFPQ */
368 * Free a packet currently used as a continuation buffer
370 #ifdef RX_ENABLE_TSFPQ
371 /* num_pkts=0 means queue length is unknown */
373 rxi_FreePackets(int num_pkts, struct rx_queue * q)
375 struct rx_ts_info_t * rx_ts_info;
376 struct rx_packet *c, *nc;
379 osi_Assert(num_pkts >= 0);
380 RX_TS_INFO_GET(rx_ts_info);
383 for (queue_Scan(q, c, nc, rx_packet), num_pkts++) {
384 rxi_FreeDataBufsTSFPQ(c, 2, 0);
387 for (queue_Scan(q, c, nc, rx_packet)) {
388 rxi_FreeDataBufsTSFPQ(c, 2, 0);
393 RX_TS_FPQ_QCHECKIN(rx_ts_info, num_pkts, q);
396 if (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax) {
398 MUTEX_ENTER(&rx_freePktQ_lock);
400 RX_TS_FPQ_LTOG(rx_ts_info);
402 /* Wakeup anyone waiting for packets */
405 MUTEX_EXIT(&rx_freePktQ_lock);
411 #else /* RX_ENABLE_TSFPQ */
412 /* num_pkts=0 means queue length is unknown */
414 rxi_FreePackets(int num_pkts, struct rx_queue *q)
417 struct rx_packet *p, *np;
421 osi_Assert(num_pkts >= 0);
425 for (queue_Scan(q, p, np, rx_packet), num_pkts++) {
426 if (p->niovecs > 2) {
427 qlen += rxi_FreeDataBufsToQueue(p, 2, &cbs);
434 for (queue_Scan(q, p, np, rx_packet)) {
435 if (p->niovecs > 2) {
436 qlen += rxi_FreeDataBufsToQueue(p, 2, &cbs);
443 queue_SpliceAppend(q, &cbs);
449 MUTEX_ENTER(&rx_freePktQ_lock);
451 queue_SpliceAppend(&rx_freePacketQueue, q);
452 rx_nFreePackets += qlen;
454 /* Wakeup anyone waiting for packets */
457 MUTEX_EXIT(&rx_freePktQ_lock);
462 #endif /* RX_ENABLE_TSFPQ */
464 /* this one is kind of awful.
465 * In rxkad, the packet has been all shortened, and everything, ready for
466 * sending. All of a sudden, we discover we need some of that space back.
467 * This isn't terribly general, because it knows that the packets are only
468 * rounded up to the EBS (userdata + security header).
471 rxi_RoundUpPacket(struct rx_packet *p, unsigned int nb)
475 if (p->wirevec[i].iov_base == (caddr_t) p->localdata) {
476 if (p->wirevec[i].iov_len <= RX_FIRSTBUFFERSIZE - nb) {
477 p->wirevec[i].iov_len += nb;
481 if (p->wirevec[i].iov_len <= RX_CBUFFERSIZE - nb) {
482 p->wirevec[i].iov_len += nb;
490 /* get sufficient space to store nb bytes of data (or more), and hook
491 * it into the supplied packet. Return nbytes<=0 if successful, otherwise
492 * returns the number of bytes >0 which it failed to come up with.
493 * Don't need to worry about locking on packet, since only
494 * one thread can manipulate one at a time. Locking on continution
495 * packets is handled by AllocPacketBufs */
496 /* MTUXXX don't need to go throught the for loop if we can trust niovecs */
498 rxi_AllocDataBuf(struct rx_packet *p, int nb, int class)
502 struct rx_packet *cb, *ncb;
504 /* compute the number of cbuf's we need */
505 nv = nb / RX_CBUFFERSIZE;
506 if ((nv * RX_CBUFFERSIZE) < nb)
508 if ((nv + p->niovecs) > RX_MAXWVECS)
509 nv = RX_MAXWVECS - p->niovecs;
513 /* allocate buffers */
515 nv = AllocPacketBufs(class, nv, &q);
517 /* setup packet iovs */
518 for (i = p->niovecs, queue_Scan(&q, cb, ncb, rx_packet), i++) {
520 p->wirevec[i].iov_base = (caddr_t) cb->localdata;
521 p->wirevec[i].iov_len = RX_CBUFFERSIZE;
524 nb -= (nv * RX_CBUFFERSIZE);
525 p->length += (nv * RX_CBUFFERSIZE);
531 /* Add more packet buffers */
532 #ifdef RX_ENABLE_TSFPQ
534 rxi_MorePackets(int apackets)
536 struct rx_packet *p, *e;
537 struct rx_ts_info_t * rx_ts_info;
541 getme = apackets * sizeof(struct rx_packet);
542 p = (struct rx_packet *)osi_Alloc(getme);
545 PIN(p, getme); /* XXXXX */
547 RX_TS_INFO_GET(rx_ts_info);
549 RX_TS_FPQ_LOCAL_ALLOC(rx_ts_info,apackets);
550 /* TSFPQ patch also needs to keep track of total packets */
552 MUTEX_ENTER(&rx_packets_mutex);
553 rx_nPackets += apackets;
554 RX_TS_FPQ_COMPUTE_LIMITS;
555 MUTEX_EXIT(&rx_packets_mutex);
557 for (e = p + apackets; p < e; p++) {
558 RX_PACKET_IOV_INIT(p);
561 RX_TS_FPQ_CHECKIN(rx_ts_info,p);
564 MUTEX_ENTER(&rx_freePktQ_lock);
565 #ifdef RXDEBUG_PACKET
566 p->packetId = rx_packet_id++;
567 p->allNextp = rx_mallocedP;
568 #endif /* RXDEBUG_PACKET */
570 MUTEX_EXIT(&rx_freePktQ_lock);
573 rx_ts_info->_FPQ.delta += apackets;
575 if (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax) {
577 MUTEX_ENTER(&rx_freePktQ_lock);
579 RX_TS_FPQ_LTOG(rx_ts_info);
580 rxi_NeedMorePackets = FALSE;
583 MUTEX_EXIT(&rx_freePktQ_lock);
587 #else /* RX_ENABLE_TSFPQ */
589 rxi_MorePackets(int apackets)
591 struct rx_packet *p, *e;
595 getme = apackets * sizeof(struct rx_packet);
596 p = (struct rx_packet *)osi_Alloc(getme);
599 PIN(p, getme); /* XXXXX */
602 MUTEX_ENTER(&rx_freePktQ_lock);
604 for (e = p + apackets; p < e; p++) {
605 RX_PACKET_IOV_INIT(p);
606 p->flags |= RX_PKTFLAG_FREE;
609 queue_Append(&rx_freePacketQueue, p);
610 #ifdef RXDEBUG_PACKET
611 p->packetId = rx_packet_id++;
612 p->allNextp = rx_mallocedP;
613 #endif /* RXDEBUG_PACKET */
617 rx_nFreePackets += apackets;
618 rxi_NeedMorePackets = FALSE;
621 MUTEX_EXIT(&rx_freePktQ_lock);
624 #endif /* RX_ENABLE_TSFPQ */
626 #ifdef RX_ENABLE_TSFPQ
628 rxi_MorePacketsTSFPQ(int apackets, int flush_global, int num_keep_local)
630 struct rx_packet *p, *e;
631 struct rx_ts_info_t * rx_ts_info;
635 getme = apackets * sizeof(struct rx_packet);
636 p = (struct rx_packet *)osi_Alloc(getme);
638 PIN(p, getme); /* XXXXX */
640 RX_TS_INFO_GET(rx_ts_info);
642 RX_TS_FPQ_LOCAL_ALLOC(rx_ts_info,apackets);
643 /* TSFPQ patch also needs to keep track of total packets */
644 MUTEX_ENTER(&rx_packets_mutex);
645 rx_nPackets += apackets;
646 RX_TS_FPQ_COMPUTE_LIMITS;
647 MUTEX_EXIT(&rx_packets_mutex);
649 for (e = p + apackets; p < e; p++) {
650 RX_PACKET_IOV_INIT(p);
652 RX_TS_FPQ_CHECKIN(rx_ts_info,p);
655 MUTEX_ENTER(&rx_freePktQ_lock);
656 #ifdef RXDEBUG_PACKET
657 p->packetId = rx_packet_id++;
658 p->allNextp = rx_mallocedP;
659 #endif /* RXDEBUG_PACKET */
661 MUTEX_EXIT(&rx_freePktQ_lock);
664 rx_ts_info->_FPQ.delta += apackets;
667 (num_keep_local < apackets)) {
669 MUTEX_ENTER(&rx_freePktQ_lock);
671 RX_TS_FPQ_LTOG2(rx_ts_info, (apackets - num_keep_local));
672 rxi_NeedMorePackets = FALSE;
675 MUTEX_EXIT(&rx_freePktQ_lock);
679 #endif /* RX_ENABLE_TSFPQ */
682 /* Add more packet buffers */
684 rxi_MorePacketsNoLock(int apackets)
686 #ifdef RX_ENABLE_TSFPQ
687 struct rx_ts_info_t * rx_ts_info;
688 #endif /* RX_ENABLE_TSFPQ */
689 struct rx_packet *p, *e;
692 /* allocate enough packets that 1/4 of the packets will be able
693 * to hold maximal amounts of data */
694 apackets += (apackets / 4)
695 * ((rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE) / RX_CBUFFERSIZE);
697 getme = apackets * sizeof(struct rx_packet);
698 p = (struct rx_packet *)osi_Alloc(getme);
700 apackets -= apackets / 4;
701 osi_Assert(apackets > 0);
706 #ifdef RX_ENABLE_TSFPQ
707 RX_TS_INFO_GET(rx_ts_info);
708 RX_TS_FPQ_GLOBAL_ALLOC(rx_ts_info,apackets);
709 #endif /* RX_ENABLE_TSFPQ */
711 for (e = p + apackets; p < e; p++) {
712 RX_PACKET_IOV_INIT(p);
713 p->flags |= RX_PKTFLAG_FREE;
716 queue_Append(&rx_freePacketQueue, p);
717 #ifdef RXDEBUG_PACKET
718 p->packetId = rx_packet_id++;
719 p->allNextp = rx_mallocedP;
720 #endif /* RXDEBUG_PACKET */
724 rx_nFreePackets += apackets;
725 #ifdef RX_ENABLE_TSFPQ
726 /* TSFPQ patch also needs to keep track of total packets */
727 MUTEX_ENTER(&rx_packets_mutex);
728 rx_nPackets += apackets;
729 RX_TS_FPQ_COMPUTE_LIMITS;
730 MUTEX_EXIT(&rx_packets_mutex);
731 #endif /* RX_ENABLE_TSFPQ */
732 rxi_NeedMorePackets = FALSE;
738 rxi_FreeAllPackets(void)
740 /* must be called at proper interrupt level, etcetera */
741 /* MTUXXX need to free all Packets */
742 osi_Free(rx_mallocedP,
743 (rx_maxReceiveWindow + 2) * sizeof(struct rx_packet));
744 UNPIN(rx_mallocedP, (rx_maxReceiveWindow + 2) * sizeof(struct rx_packet));
747 #ifdef RX_ENABLE_TSFPQ
749 rxi_AdjustLocalPacketsTSFPQ(int num_keep_local, int allow_overcommit)
751 struct rx_ts_info_t * rx_ts_info;
755 RX_TS_INFO_GET(rx_ts_info);
757 if (num_keep_local != rx_ts_info->_FPQ.len) {
759 MUTEX_ENTER(&rx_freePktQ_lock);
760 if (num_keep_local < rx_ts_info->_FPQ.len) {
761 xfer = rx_ts_info->_FPQ.len - num_keep_local;
762 RX_TS_FPQ_LTOG2(rx_ts_info, xfer);
765 xfer = num_keep_local - rx_ts_info->_FPQ.len;
766 if ((num_keep_local > rx_TSFPQLocalMax) && !allow_overcommit)
767 xfer = rx_TSFPQLocalMax - rx_ts_info->_FPQ.len;
768 if (rx_nFreePackets < xfer) {
769 rxi_MorePacketsNoLock(MAX(xfer - rx_nFreePackets, 4 * rx_initSendWindow));
771 RX_TS_FPQ_GTOL2(rx_ts_info, xfer);
773 MUTEX_EXIT(&rx_freePktQ_lock);
779 rxi_FlushLocalPacketsTSFPQ(void)
781 rxi_AdjustLocalPacketsTSFPQ(0, 0);
783 #endif /* RX_ENABLE_TSFPQ */
785 /* Allocate more packets iff we need more continuation buffers */
786 /* In kernel, can't page in memory with interrupts disabled, so we
787 * don't use the event mechanism. */
789 rx_CheckPackets(void)
791 if (rxi_NeedMorePackets) {
792 rxi_MorePackets(rx_initSendWindow);
796 /* In the packet freeing routine below, the assumption is that
797 we want all of the packets to be used equally frequently, so that we
798 don't get packet buffers paging out. It would be just as valid to
799 assume that we DO want them to page out if not many are being used.
800 In any event, we assume the former, and append the packets to the end
802 /* This explanation is bogus. The free list doesn't remain in any kind of
803 useful order for afs_int32: the packets in use get pretty much randomly scattered
804 across all the pages. In order to permit unused {packets,bufs} to page out, they
805 must be stored so that packets which are adjacent in memory are adjacent in the
806 free list. An array springs rapidly to mind.
809 /* Actually free the packet p. */
810 #ifdef RX_ENABLE_TSFPQ
812 rxi_FreePacketNoLock(struct rx_packet *p)
814 struct rx_ts_info_t * rx_ts_info;
815 dpf(("Free %"AFS_PTR_FMT"\n", p));
817 RX_TS_INFO_GET(rx_ts_info);
818 RX_TS_FPQ_CHECKIN(rx_ts_info,p);
819 if (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax) {
820 RX_TS_FPQ_LTOG(rx_ts_info);
823 #else /* RX_ENABLE_TSFPQ */
825 rxi_FreePacketNoLock(struct rx_packet *p)
827 dpf(("Free %"AFS_PTR_FMT"\n", p));
831 queue_Append(&rx_freePacketQueue, p);
833 #endif /* RX_ENABLE_TSFPQ */
835 #ifdef RX_ENABLE_TSFPQ
837 rxi_FreePacketTSFPQ(struct rx_packet *p, int flush_global)
839 struct rx_ts_info_t * rx_ts_info;
840 dpf(("Free %"AFS_PTR_FMT"\n", p));
842 RX_TS_INFO_GET(rx_ts_info);
843 RX_TS_FPQ_CHECKIN(rx_ts_info,p);
845 if (flush_global && (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax)) {
847 MUTEX_ENTER(&rx_freePktQ_lock);
849 RX_TS_FPQ_LTOG(rx_ts_info);
851 /* Wakeup anyone waiting for packets */
854 MUTEX_EXIT(&rx_freePktQ_lock);
858 #endif /* RX_ENABLE_TSFPQ */
861 * free continuation buffers off a packet into a queue
863 * [IN] p -- packet from which continuation buffers will be freed
864 * [IN] first -- iovec offset of first continuation buffer to free
865 * [IN] q -- queue into which continuation buffers will be chained
868 * number of continuation buffers freed
870 #ifndef RX_ENABLE_TSFPQ
872 rxi_FreeDataBufsToQueue(struct rx_packet *p, afs_uint32 first, struct rx_queue * q)
875 struct rx_packet * cb;
878 for (first = MAX(2, first); first < p->niovecs; first++, count++) {
879 iov = &p->wirevec[first];
881 osi_Panic("rxi_FreeDataBufsToQueue: unexpected NULL iov");
882 cb = RX_CBUF_TO_PACKET(iov->iov_base, p);
883 RX_FPQ_MARK_FREE(cb);
894 * free packet continuation buffers into the global free packet pool
896 * [IN] p -- packet from which to free continuation buffers
897 * [IN] first -- iovec offset of first continuation buffer to free
903 rxi_FreeDataBufsNoLock(struct rx_packet *p, afs_uint32 first)
907 for (first = MAX(2, first); first < p->niovecs; first++) {
908 iov = &p->wirevec[first];
910 osi_Panic("rxi_FreeDataBufsNoLock: unexpected NULL iov");
911 rxi_FreePacketNoLock(RX_CBUF_TO_PACKET(iov->iov_base, p));
919 #ifdef RX_ENABLE_TSFPQ
921 * free packet continuation buffers into the thread-local free pool
923 * [IN] p -- packet from which continuation buffers will be freed
924 * [IN] first -- iovec offset of first continuation buffer to free
925 * any value less than 2, the min number of iovecs,
926 * is treated as if it is 2.
927 * [IN] flush_global -- if nonzero, we will flush overquota packets to the
928 * global free pool before returning
934 rxi_FreeDataBufsTSFPQ(struct rx_packet *p, afs_uint32 first, int flush_global)
937 struct rx_ts_info_t * rx_ts_info;
939 RX_TS_INFO_GET(rx_ts_info);
941 for (first = MAX(2, first); first < p->niovecs; first++) {
942 iov = &p->wirevec[first];
944 osi_Panic("rxi_FreeDataBufsTSFPQ: unexpected NULL iov");
945 RX_TS_FPQ_CHECKIN(rx_ts_info,RX_CBUF_TO_PACKET(iov->iov_base, p));
950 if (flush_global && (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax)) {
952 MUTEX_ENTER(&rx_freePktQ_lock);
954 RX_TS_FPQ_LTOG(rx_ts_info);
956 /* Wakeup anyone waiting for packets */
959 MUTEX_EXIT(&rx_freePktQ_lock);
964 #endif /* RX_ENABLE_TSFPQ */
966 int rxi_nBadIovecs = 0;
968 /* rxi_RestoreDataBufs
970 * Restore the correct sizes to the iovecs. Called when reusing a packet
971 * for reading off the wire.
974 rxi_RestoreDataBufs(struct rx_packet *p)
977 struct iovec *iov = &p->wirevec[2];
979 RX_PACKET_IOV_INIT(p);
981 for (i = 2, iov = &p->wirevec[2]; i < p->niovecs; i++, iov++) {
982 if (!iov->iov_base) {
987 iov->iov_len = RX_CBUFFERSIZE;
991 #ifdef RX_ENABLE_TSFPQ
993 rxi_TrimDataBufs(struct rx_packet *p, int first)
996 struct iovec *iov, *end;
997 struct rx_ts_info_t * rx_ts_info;
1001 osi_Panic("TrimDataBufs 1: first must be 1");
1003 /* Skip over continuation buffers containing message data */
1004 iov = &p->wirevec[2];
1005 end = iov + (p->niovecs - 2);
1006 length = p->length - p->wirevec[1].iov_len;
1007 for (; iov < end && length > 0; iov++) {
1009 osi_Panic("TrimDataBufs 3: vecs 1-niovecs must not be NULL");
1010 length -= iov->iov_len;
1013 /* iov now points to the first empty data buffer. */
1017 RX_TS_INFO_GET(rx_ts_info);
1018 for (; iov < end; iov++) {
1020 osi_Panic("TrimDataBufs 4: vecs 2-niovecs must not be NULL");
1021 RX_TS_FPQ_CHECKIN(rx_ts_info,RX_CBUF_TO_PACKET(iov->iov_base, p));
1024 if (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax) {
1026 MUTEX_ENTER(&rx_freePktQ_lock);
1028 RX_TS_FPQ_LTOG(rx_ts_info);
1029 rxi_PacketsUnWait();
1031 MUTEX_EXIT(&rx_freePktQ_lock);
1037 #else /* RX_ENABLE_TSFPQ */
1039 rxi_TrimDataBufs(struct rx_packet *p, int first)
1042 struct iovec *iov, *end;
1046 osi_Panic("TrimDataBufs 1: first must be 1");
1048 /* Skip over continuation buffers containing message data */
1049 iov = &p->wirevec[2];
1050 end = iov + (p->niovecs - 2);
1051 length = p->length - p->wirevec[1].iov_len;
1052 for (; iov < end && length > 0; iov++) {
1054 osi_Panic("TrimDataBufs 3: vecs 1-niovecs must not be NULL");
1055 length -= iov->iov_len;
1058 /* iov now points to the first empty data buffer. */
1063 MUTEX_ENTER(&rx_freePktQ_lock);
1065 for (; iov < end; iov++) {
1067 osi_Panic("TrimDataBufs 4: vecs 2-niovecs must not be NULL");
1068 rxi_FreePacketNoLock(RX_CBUF_TO_PACKET(iov->iov_base, p));
1071 rxi_PacketsUnWait();
1073 MUTEX_EXIT(&rx_freePktQ_lock);
1078 #endif /* RX_ENABLE_TSFPQ */
1080 /* Free the packet p. P is assumed not to be on any queue, i.e.
1081 * remove it yourself first if you call this routine. */
1082 #ifdef RX_ENABLE_TSFPQ
1084 rxi_FreePacket(struct rx_packet *p)
1086 rxi_FreeDataBufsTSFPQ(p, 2, 0);
1087 rxi_FreePacketTSFPQ(p, RX_TS_FPQ_FLUSH_GLOBAL);
1089 #else /* RX_ENABLE_TSFPQ */
1091 rxi_FreePacket(struct rx_packet *p)
1096 MUTEX_ENTER(&rx_freePktQ_lock);
1098 rxi_FreeDataBufsNoLock(p, 2);
1099 rxi_FreePacketNoLock(p);
1100 /* Wakeup anyone waiting for packets */
1101 rxi_PacketsUnWait();
1103 MUTEX_EXIT(&rx_freePktQ_lock);
1106 #endif /* RX_ENABLE_TSFPQ */
1108 /* rxi_AllocPacket sets up p->length so it reflects the number of
1109 * bytes in the packet at this point, **not including** the header.
1110 * The header is absolutely necessary, besides, this is the way the
1111 * length field is usually used */
1112 #ifdef RX_ENABLE_TSFPQ
1114 rxi_AllocPacketNoLock(int class)
1116 struct rx_packet *p;
1117 struct rx_ts_info_t * rx_ts_info;
1119 RX_TS_INFO_GET(rx_ts_info);
1122 if (rxi_OverQuota(class)) {
1123 rxi_NeedMorePackets = TRUE;
1124 if (rx_stats_active) {
1126 case RX_PACKET_CLASS_RECEIVE:
1127 rx_MutexIncrement(rx_stats.receivePktAllocFailures, rx_stats_mutex);
1129 case RX_PACKET_CLASS_SEND:
1130 rx_MutexIncrement(rx_stats.sendPktAllocFailures, rx_stats_mutex);
1132 case RX_PACKET_CLASS_SPECIAL:
1133 rx_MutexIncrement(rx_stats.specialPktAllocFailures, rx_stats_mutex);
1135 case RX_PACKET_CLASS_RECV_CBUF:
1136 rx_MutexIncrement(rx_stats.receiveCbufPktAllocFailures, rx_stats_mutex);
1138 case RX_PACKET_CLASS_SEND_CBUF:
1139 rx_MutexIncrement(rx_stats.sendCbufPktAllocFailures, rx_stats_mutex);
1143 return (struct rx_packet *)0;
1147 if (rx_stats_active)
1148 rx_MutexIncrement(rx_stats.packetRequests, rx_stats_mutex);
1149 if (queue_IsEmpty(&rx_ts_info->_FPQ)) {
1152 if (queue_IsEmpty(&rx_freePacketQueue))
1153 osi_Panic("rxi_AllocPacket error");
1155 if (queue_IsEmpty(&rx_freePacketQueue))
1156 rxi_MorePacketsNoLock(4 * rx_initSendWindow);
1160 RX_TS_FPQ_GTOL(rx_ts_info);
1163 RX_TS_FPQ_CHECKOUT(rx_ts_info,p);
1165 dpf(("Alloc %"AFS_PTR_FMT", class %d\n", p, class));
1168 /* have to do this here because rx_FlushWrite fiddles with the iovs in
1169 * order to truncate outbound packets. In the near future, may need
1170 * to allocate bufs from a static pool here, and/or in AllocSendPacket
1172 RX_PACKET_IOV_FULLINIT(p);
1175 #else /* RX_ENABLE_TSFPQ */
1177 rxi_AllocPacketNoLock(int class)
1179 struct rx_packet *p;
1182 if (rxi_OverQuota(class)) {
1183 rxi_NeedMorePackets = TRUE;
1184 if (rx_stats_active) {
1186 case RX_PACKET_CLASS_RECEIVE:
1187 rx_MutexIncrement(rx_stats.receivePktAllocFailures, rx_stats_mutex);
1189 case RX_PACKET_CLASS_SEND:
1190 rx_MutexIncrement(rx_stats.sendPktAllocFailures, rx_stats_mutex);
1192 case RX_PACKET_CLASS_SPECIAL:
1193 rx_MutexIncrement(rx_stats.specialPktAllocFailures, rx_stats_mutex);
1195 case RX_PACKET_CLASS_RECV_CBUF:
1196 rx_MutexIncrement(rx_stats.receiveCbufPktAllocFailures, rx_stats_mutex);
1198 case RX_PACKET_CLASS_SEND_CBUF:
1199 rx_MutexIncrement(rx_stats.sendCbufPktAllocFailures, rx_stats_mutex);
1203 return (struct rx_packet *)0;
1207 if (rx_stats_active)
1208 rx_MutexIncrement(rx_stats.packetRequests, rx_stats_mutex);
1211 if (queue_IsEmpty(&rx_freePacketQueue))
1212 osi_Panic("rxi_AllocPacket error");
1214 if (queue_IsEmpty(&rx_freePacketQueue))
1215 rxi_MorePacketsNoLock(4 * rx_initSendWindow);
1219 p = queue_First(&rx_freePacketQueue, rx_packet);
1221 RX_FPQ_MARK_USED(p);
1223 dpf(("Alloc %"AFS_PTR_FMT", class %d\n", p, class));
1226 /* have to do this here because rx_FlushWrite fiddles with the iovs in
1227 * order to truncate outbound packets. In the near future, may need
1228 * to allocate bufs from a static pool here, and/or in AllocSendPacket
1230 RX_PACKET_IOV_FULLINIT(p);
1233 #endif /* RX_ENABLE_TSFPQ */
1235 #ifdef RX_ENABLE_TSFPQ
1237 rxi_AllocPacketTSFPQ(int class, int pull_global)
1239 struct rx_packet *p;
1240 struct rx_ts_info_t * rx_ts_info;
1242 RX_TS_INFO_GET(rx_ts_info);
1244 if (rx_stats_active)
1245 rx_MutexIncrement(rx_stats.packetRequests, rx_stats_mutex);
1246 if (pull_global && queue_IsEmpty(&rx_ts_info->_FPQ)) {
1247 MUTEX_ENTER(&rx_freePktQ_lock);
1249 if (queue_IsEmpty(&rx_freePacketQueue))
1250 rxi_MorePacketsNoLock(4 * rx_initSendWindow);
1252 RX_TS_FPQ_GTOL(rx_ts_info);
1254 MUTEX_EXIT(&rx_freePktQ_lock);
1255 } else if (queue_IsEmpty(&rx_ts_info->_FPQ)) {
1259 RX_TS_FPQ_CHECKOUT(rx_ts_info,p);
1261 dpf(("Alloc %"AFS_PTR_FMT", class %d\n", p, class));
1263 /* have to do this here because rx_FlushWrite fiddles with the iovs in
1264 * order to truncate outbound packets. In the near future, may need
1265 * to allocate bufs from a static pool here, and/or in AllocSendPacket
1267 RX_PACKET_IOV_FULLINIT(p);
1270 #endif /* RX_ENABLE_TSFPQ */
1272 #ifdef RX_ENABLE_TSFPQ
1274 rxi_AllocPacket(int class)
1276 struct rx_packet *p;
1278 p = rxi_AllocPacketTSFPQ(class, RX_TS_FPQ_PULL_GLOBAL);
1281 #else /* RX_ENABLE_TSFPQ */
1283 rxi_AllocPacket(int class)
1285 struct rx_packet *p;
1287 MUTEX_ENTER(&rx_freePktQ_lock);
1288 p = rxi_AllocPacketNoLock(class);
1289 MUTEX_EXIT(&rx_freePktQ_lock);
1292 #endif /* RX_ENABLE_TSFPQ */
1294 /* This guy comes up with as many buffers as it {takes,can get} given
1295 * the MTU for this call. It also sets the packet length before
1296 * returning. caution: this is often called at NETPRI
1297 * Called with call locked.
1300 rxi_AllocSendPacket(struct rx_call *call, int want)
1302 struct rx_packet *p = (struct rx_packet *)0;
1307 mud = call->MTU - RX_HEADER_SIZE;
1309 rx_GetSecurityHeaderSize(rx_ConnectionOf(call)) +
1310 rx_GetSecurityMaxTrailerSize(rx_ConnectionOf(call));
1312 #ifdef RX_ENABLE_TSFPQ
1313 if ((p = rxi_AllocPacketTSFPQ(RX_PACKET_CLASS_SEND, 0))) {
1315 want = MIN(want, mud);
1317 if ((unsigned)want > p->length)
1318 (void)rxi_AllocDataBuf(p, (want - p->length),
1319 RX_PACKET_CLASS_SEND_CBUF);
1321 if (p->length > mud)
1324 if (delta >= p->length) {
1332 #endif /* RX_ENABLE_TSFPQ */
1334 while (!(call->error)) {
1335 MUTEX_ENTER(&rx_freePktQ_lock);
1336 /* if an error occurred, or we get the packet we want, we're done */
1337 if ((p = rxi_AllocPacketNoLock(RX_PACKET_CLASS_SEND))) {
1338 MUTEX_EXIT(&rx_freePktQ_lock);
1341 want = MIN(want, mud);
1343 if ((unsigned)want > p->length)
1344 (void)rxi_AllocDataBuf(p, (want - p->length),
1345 RX_PACKET_CLASS_SEND_CBUF);
1347 if (p->length > mud)
1350 if (delta >= p->length) {
1359 /* no error occurred, and we didn't get a packet, so we sleep.
1360 * At this point, we assume that packets will be returned
1361 * sooner or later, as packets are acknowledged, and so we
1364 call->flags |= RX_CALL_WAIT_PACKETS;
1365 CALL_HOLD(call, RX_CALL_REFCOUNT_PACKET);
1366 MUTEX_EXIT(&call->lock);
1367 rx_waitingForPackets = 1;
1369 #ifdef RX_ENABLE_LOCKS
1370 CV_WAIT(&rx_waitingForPackets_cv, &rx_freePktQ_lock);
1372 osi_rxSleep(&rx_waitingForPackets);
1374 MUTEX_EXIT(&rx_freePktQ_lock);
1375 MUTEX_ENTER(&call->lock);
1376 CALL_RELE(call, RX_CALL_REFCOUNT_PACKET);
1377 call->flags &= ~RX_CALL_WAIT_PACKETS;
1386 /* Windows does not use file descriptors. */
1387 #define CountFDs(amax) 0
1389 /* count the number of used FDs */
1398 for (i = 0; i < amax; i++) {
1399 code = fstat(i, &tstat);
1405 #endif /* AFS_NT40_ENV */
1408 #define CountFDs(amax) amax
1412 #if !defined(KERNEL) || defined(UKERNEL)
1414 /* This function reads a single packet from the interface into the
1415 * supplied packet buffer (*p). Return 0 if the packet is bogus. The
1416 * (host,port) of the sender are stored in the supplied variables, and
1417 * the data length of the packet is stored in the packet structure.
1418 * The header is decoded. */
1420 rxi_ReadPacket(osi_socket socket, struct rx_packet *p, afs_uint32 * host,
1423 struct sockaddr_in from;
1424 unsigned int nbytes;
1426 afs_uint32 tlen, savelen;
1428 rx_computelen(p, tlen);
1429 rx_SetDataSize(p, tlen); /* this is the size of the user data area */
1431 tlen += RX_HEADER_SIZE; /* now this is the size of the entire packet */
1432 rlen = rx_maxJumboRecvSize; /* this is what I am advertising. Only check
1433 * it once in order to avoid races. */
1436 tlen = rxi_AllocDataBuf(p, tlen, RX_PACKET_CLASS_SEND_CBUF);
1444 /* Extend the last iovec for padding, it's just to make sure that the
1445 * read doesn't return more data than we expect, and is done to get around
1446 * our problems caused by the lack of a length field in the rx header.
1447 * Use the extra buffer that follows the localdata in each packet
1449 savelen = p->wirevec[p->niovecs - 1].iov_len;
1450 p->wirevec[p->niovecs - 1].iov_len += RX_EXTRABUFFERSIZE;
1452 memset(&msg, 0, sizeof(msg));
1453 msg.msg_name = (char *)&from;
1454 msg.msg_namelen = sizeof(struct sockaddr_in);
1455 msg.msg_iov = p->wirevec;
1456 msg.msg_iovlen = p->niovecs;
1457 nbytes = rxi_Recvmsg(socket, &msg, 0);
1459 /* restore the vec to its correct state */
1460 p->wirevec[p->niovecs - 1].iov_len = savelen;
1462 p->length = (u_short)(nbytes - RX_HEADER_SIZE);
1463 if ((nbytes > tlen) || (p->length & 0x8000)) { /* Bogus packet */
1464 if (nbytes < 0 && errno == EWOULDBLOCK) {
1465 if (rx_stats_active)
1466 rx_MutexIncrement(rx_stats.noPacketOnRead, rx_stats_mutex);
1467 } else if (nbytes <= 0) {
1468 if (rx_stats_active) {
1469 MUTEX_ENTER(&rx_stats_mutex);
1470 rx_stats.bogusPacketOnRead++;
1471 rx_stats.bogusHost = from.sin_addr.s_addr;
1472 MUTEX_EXIT(&rx_stats_mutex);
1474 dpf(("B: bogus packet from [%x,%d] nb=%d", ntohl(from.sin_addr.s_addr),
1475 ntohs(from.sin_port), nbytes));
1480 else if ((rx_intentionallyDroppedOnReadPer100 > 0)
1481 && (random() % 100 < rx_intentionallyDroppedOnReadPer100)) {
1482 rxi_DecodePacketHeader(p);
1484 *host = from.sin_addr.s_addr;
1485 *port = from.sin_port;
1487 dpf(("Dropped %d %s: %x.%u.%u.%u.%u.%u.%u flags %d len %d",
1488 p->header.serial, rx_packetTypes[p->header.type - 1], ntohl(*host), ntohs(*port), p->header.serial,
1489 p->header.epoch, p->header.cid, p->header.callNumber, p->header.seq, p->header.flags,
1491 #ifdef RX_TRIMDATABUFS
1492 rxi_TrimDataBufs(p, 1);
1498 /* Extract packet header. */
1499 rxi_DecodePacketHeader(p);
1501 *host = from.sin_addr.s_addr;
1502 *port = from.sin_port;
1503 if (p->header.type > 0 && p->header.type < RX_N_PACKET_TYPES) {
1504 struct rx_peer *peer;
1505 if (rx_stats_active)
1506 rx_MutexIncrement(rx_stats.packetsRead[p->header.type - 1], rx_stats_mutex);
1508 * Try to look up this peer structure. If it doesn't exist,
1509 * don't create a new one -
1510 * we don't keep count of the bytes sent/received if a peer
1511 * structure doesn't already exist.
1513 * The peer/connection cleanup code assumes that there is 1 peer
1514 * per connection. If we actually created a peer structure here
1515 * and this packet was an rxdebug packet, the peer structure would
1516 * never be cleaned up.
1518 peer = rxi_FindPeer(*host, *port, 0, 0);
1519 /* Since this may not be associated with a connection,
1520 * it may have no refCount, meaning we could race with
1523 if (peer && (peer->refCount > 0)) {
1524 MUTEX_ENTER(&peer->peer_lock);
1525 hadd32(peer->bytesReceived, p->length);
1526 MUTEX_EXIT(&peer->peer_lock);
1530 #ifdef RX_TRIMDATABUFS
1531 /* Free any empty packet buffers at the end of this packet */
1532 rxi_TrimDataBufs(p, 1);
1538 #endif /* !KERNEL || UKERNEL */
1540 /* This function splits off the first packet in a jumbo packet.
1541 * As of AFS 3.5, jumbograms contain more than one fixed size
1542 * packet, and the RX_JUMBO_PACKET flag is set in all but the
1543 * last packet header. All packets (except the last) are padded to
1544 * fall on RX_CBUFFERSIZE boundaries.
1545 * HACK: We store the length of the first n-1 packets in the
1546 * last two pad bytes. */
1549 rxi_SplitJumboPacket(struct rx_packet *p, afs_int32 host, short port,
1552 struct rx_packet *np;
1553 struct rx_jumboHeader *jp;
1559 /* All but the last packet in each jumbogram are RX_JUMBOBUFFERSIZE
1560 * bytes in length. All but the first packet are preceded by
1561 * an abbreviated four byte header. The length of the last packet
1562 * is calculated from the size of the jumbogram. */
1563 length = RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
1565 if ((int)p->length < length) {
1566 dpf(("rxi_SplitJumboPacket: bogus length %d\n", p->length));
1569 niov = p->niovecs - 2;
1571 dpf(("rxi_SplitJumboPacket: bogus niovecs %d\n", p->niovecs));
1574 iov = &p->wirevec[2];
1575 np = RX_CBUF_TO_PACKET(iov->iov_base, p);
1577 /* Get a pointer to the abbreviated packet header */
1578 jp = (struct rx_jumboHeader *)
1579 ((char *)(p->wirevec[1].iov_base) + RX_JUMBOBUFFERSIZE);
1581 /* Set up the iovecs for the next packet */
1582 np->wirevec[0].iov_base = (char *)(&np->wirehead[0]);
1583 np->wirevec[0].iov_len = sizeof(struct rx_header);
1584 np->wirevec[1].iov_base = (char *)(&np->localdata[0]);
1585 np->wirevec[1].iov_len = length - RX_JUMBOHEADERSIZE;
1586 np->niovecs = niov + 1;
1587 for (i = 2, iov++; i <= niov; i++, iov++) {
1588 np->wirevec[i] = *iov;
1590 np->length = p->length - length;
1591 p->length = RX_JUMBOBUFFERSIZE;
1594 /* Convert the jumbo packet header to host byte order */
1595 temp = ntohl(*(afs_uint32 *) jp);
1596 jp->flags = (u_char) (temp >> 24);
1597 jp->cksum = (u_short) (temp);
1599 /* Fill in the packet header */
1600 np->header = p->header;
1601 np->header.serial = p->header.serial + 1;
1602 np->header.seq = p->header.seq + 1;
1603 np->header.flags = jp->flags;
1604 np->header.spare = jp->cksum;
1610 /* Send a udp datagram */
1612 osi_NetSend(osi_socket socket, void *addr, struct iovec *dvec, int nvecs,
1613 int length, int istack)
1618 memset(&msg, 0, sizeof(msg));
1620 msg.msg_iovlen = nvecs;
1621 msg.msg_name = addr;
1622 msg.msg_namelen = sizeof(struct sockaddr_in);
1624 ret = rxi_Sendmsg(socket, &msg, 0);
1628 #elif !defined(UKERNEL)
1630 * message receipt is done in rxk_input or rx_put.
1633 #if defined(AFS_SUN5_ENV) || defined(AFS_HPUX110_ENV)
1635 * Copy an mblock to the contiguous area pointed to by cp.
1636 * MTUXXX Supposed to skip <off> bytes and copy <len> bytes,
1637 * but it doesn't really.
1638 * Returns the number of bytes not transferred.
1639 * The message is NOT changed.
1642 cpytoc(mblk_t * mp, int off, int len, char *cp)
1646 for (; mp && len > 0; mp = mp->b_cont) {
1647 if (mp->b_datap->db_type != M_DATA) {
1650 n = MIN(len, (mp->b_wptr - mp->b_rptr));
1651 memcpy(cp, (char *)mp->b_rptr, n);
1659 /* MTUXXX Supposed to skip <off> bytes and copy <len> bytes,
1660 * but it doesn't really.
1661 * This sucks, anyway, do it like m_cpy.... below
1664 cpytoiovec(mblk_t * mp, int off, int len, struct iovec *iovs,
1669 for (i = -1, t = 0; i < niovs && mp && len > 0; mp = mp->b_cont) {
1670 if (mp->b_datap->db_type != M_DATA) {
1673 n = MIN(len, (mp->b_wptr - mp->b_rptr));
1679 t = iovs[i].iov_len;
1682 memcpy(iovs[i].iov_base + o, (char *)mp->b_rptr, m);
1692 #define m_cpytoc(a, b, c, d) cpytoc(a, b, c, d)
1693 #define m_cpytoiovec(a, b, c, d, e) cpytoiovec(a, b, c, d, e)
1695 #if !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN80_ENV)
1697 m_cpytoiovec(struct mbuf *m, int off, int len, struct iovec iovs[], int niovs)
1700 unsigned int l1, l2, i, t;
1702 if (m == NULL || off < 0 || len < 0 || iovs == NULL)
1703 osi_Panic("m_cpytoiovec"); /* MTUXXX probably don't need this check */
1706 if (m->m_len <= off) {
1716 p1 = mtod(m, caddr_t) + off;
1717 l1 = m->m_len - off;
1719 p2 = iovs[0].iov_base;
1720 l2 = iovs[0].iov_len;
1723 t = MIN(l1, MIN(l2, (unsigned int)len));
1734 p1 = mtod(m, caddr_t);
1740 p2 = iovs[i].iov_base;
1741 l2 = iovs[i].iov_len;
1749 #endif /* AFS_SUN5_ENV */
1751 #if !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN80_ENV)
1753 rx_mb_to_packet(amb, free, hdr_len, data_len, phandle)
1754 #if defined(AFS_SUN5_ENV) || defined(AFS_HPUX110_ENV)
1760 struct rx_packet *phandle;
1761 int hdr_len, data_len;
1766 m_cpytoiovec(amb, hdr_len, data_len, phandle->wirevec,
1773 #endif /*KERNEL && !UKERNEL */
1776 /* send a response to a debug packet */
1779 rxi_ReceiveDebugPacket(struct rx_packet *ap, osi_socket asocket,
1780 afs_int32 ahost, short aport, int istack)
1782 struct rx_debugIn tin;
1784 struct rx_serverQueueEntry *np, *nqe;
1787 * Only respond to client-initiated Rx debug packets,
1788 * and clear the client flag in the response.
1790 if (ap->header.flags & RX_CLIENT_INITIATED) {
1791 ap->header.flags = ap->header.flags & ~RX_CLIENT_INITIATED;
1792 rxi_EncodePacketHeader(ap);
1797 rx_packetread(ap, 0, sizeof(struct rx_debugIn), (char *)&tin);
1798 /* all done with packet, now set length to the truth, so we can
1799 * reuse this packet */
1800 rx_computelen(ap, ap->length);
1802 tin.type = ntohl(tin.type);
1803 tin.index = ntohl(tin.index);
1805 case RX_DEBUGI_GETSTATS:{
1806 struct rx_debugStats tstat;
1808 /* get basic stats */
1809 memset(&tstat, 0, sizeof(tstat)); /* make sure spares are zero */
1810 tstat.version = RX_DEBUGI_VERSION;
1811 #ifndef RX_ENABLE_LOCKS
1812 tstat.waitingForPackets = rx_waitingForPackets;
1814 MUTEX_ENTER(&rx_serverPool_lock);
1815 tstat.nFreePackets = htonl(rx_nFreePackets);
1816 tstat.nPackets = htonl(rx_nPackets);
1817 tstat.callsExecuted = htonl(rxi_nCalls);
1818 tstat.packetReclaims = htonl(rx_packetReclaims);
1819 tstat.usedFDs = CountFDs(64);
1820 tstat.nWaiting = htonl(rx_nWaiting);
1821 tstat.nWaited = htonl(rx_nWaited);
1822 queue_Count(&rx_idleServerQueue, np, nqe, rx_serverQueueEntry,
1824 MUTEX_EXIT(&rx_serverPool_lock);
1825 tstat.idleThreads = htonl(tstat.idleThreads);
1826 tl = sizeof(struct rx_debugStats) - ap->length;
1828 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
1831 rx_packetwrite(ap, 0, sizeof(struct rx_debugStats),
1833 ap->length = sizeof(struct rx_debugStats);
1834 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
1835 rx_computelen(ap, ap->length);
1840 case RX_DEBUGI_GETALLCONN:
1841 case RX_DEBUGI_GETCONN:{
1843 struct rx_connection *tc;
1844 struct rx_call *tcall;
1845 struct rx_debugConn tconn;
1846 int all = (tin.type == RX_DEBUGI_GETALLCONN);
1849 tl = sizeof(struct rx_debugConn) - ap->length;
1851 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
1855 memset(&tconn, 0, sizeof(tconn)); /* make sure spares are zero */
1856 /* get N'th (maybe) "interesting" connection info */
1857 for (i = 0; i < rx_hashTableSize; i++) {
1858 #if !defined(KERNEL)
1859 /* the time complexity of the algorithm used here
1860 * exponentially increses with the number of connections.
1862 #ifdef AFS_PTHREAD_ENV
1868 MUTEX_ENTER(&rx_connHashTable_lock);
1869 /* We might be slightly out of step since we are not
1870 * locking each call, but this is only debugging output.
1872 for (tc = rx_connHashTable[i]; tc; tc = tc->next) {
1873 if ((all || rxi_IsConnInteresting(tc))
1874 && tin.index-- <= 0) {
1875 tconn.host = tc->peer->host;
1876 tconn.port = tc->peer->port;
1877 tconn.cid = htonl(tc->cid);
1878 tconn.epoch = htonl(tc->epoch);
1879 tconn.serial = htonl(tc->serial);
1880 for (j = 0; j < RX_MAXCALLS; j++) {
1881 tconn.callNumber[j] = htonl(tc->callNumber[j]);
1882 if ((tcall = tc->call[j])) {
1883 tconn.callState[j] = tcall->state;
1884 tconn.callMode[j] = tcall->mode;
1885 tconn.callFlags[j] = tcall->flags;
1886 if (queue_IsNotEmpty(&tcall->rq))
1887 tconn.callOther[j] |= RX_OTHER_IN;
1888 if (queue_IsNotEmpty(&tcall->tq))
1889 tconn.callOther[j] |= RX_OTHER_OUT;
1891 tconn.callState[j] = RX_STATE_NOTINIT;
1894 tconn.natMTU = htonl(tc->peer->natMTU);
1895 tconn.error = htonl(tc->error);
1896 tconn.flags = tc->flags;
1897 tconn.type = tc->type;
1898 tconn.securityIndex = tc->securityIndex;
1899 if (tc->securityObject) {
1900 RXS_GetStats(tc->securityObject, tc,
1902 #define DOHTONL(a) (tconn.secStats.a = htonl(tconn.secStats.a))
1903 #define DOHTONS(a) (tconn.secStats.a = htons(tconn.secStats.a))
1906 DOHTONL(packetsReceived);
1907 DOHTONL(packetsSent);
1908 DOHTONL(bytesReceived);
1912 sizeof(tconn.secStats.spares) /
1917 sizeof(tconn.secStats.sparel) /
1918 sizeof(afs_int32); i++)
1922 MUTEX_EXIT(&rx_connHashTable_lock);
1923 rx_packetwrite(ap, 0, sizeof(struct rx_debugConn),
1926 ap->length = sizeof(struct rx_debugConn);
1927 rxi_SendDebugPacket(ap, asocket, ahost, aport,
1933 MUTEX_EXIT(&rx_connHashTable_lock);
1935 /* if we make it here, there are no interesting packets */
1936 tconn.cid = htonl(0xffffffff); /* means end */
1937 rx_packetwrite(ap, 0, sizeof(struct rx_debugConn),
1940 ap->length = sizeof(struct rx_debugConn);
1941 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
1947 * Pass back all the peer structures we have available
1950 case RX_DEBUGI_GETPEER:{
1953 struct rx_debugPeer tpeer;
1956 tl = sizeof(struct rx_debugPeer) - ap->length;
1958 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
1962 memset(&tpeer, 0, sizeof(tpeer));
1963 for (i = 0; i < rx_hashTableSize; i++) {
1964 #if !defined(KERNEL)
1965 /* the time complexity of the algorithm used here
1966 * exponentially increses with the number of peers.
1968 * Yielding after processing each hash table entry
1969 * and dropping rx_peerHashTable_lock.
1970 * also increases the risk that we will miss a new
1971 * entry - but we are willing to live with this
1972 * limitation since this is meant for debugging only
1974 #ifdef AFS_PTHREAD_ENV
1980 MUTEX_ENTER(&rx_peerHashTable_lock);
1981 for (tp = rx_peerHashTable[i]; tp; tp = tp->next) {
1982 if (tin.index-- <= 0) {
1983 tpeer.host = tp->host;
1984 tpeer.port = tp->port;
1985 tpeer.ifMTU = htons(tp->ifMTU);
1986 tpeer.idleWhen = htonl(tp->idleWhen);
1987 tpeer.refCount = htons(tp->refCount);
1988 tpeer.burstSize = tp->burstSize;
1989 tpeer.burst = tp->burst;
1990 tpeer.burstWait.sec = htonl(tp->burstWait.sec);
1991 tpeer.burstWait.usec = htonl(tp->burstWait.usec);
1992 tpeer.rtt = htonl(tp->rtt);
1993 tpeer.rtt_dev = htonl(tp->rtt_dev);
1994 tpeer.timeout.sec = htonl(tp->timeout.sec);
1995 tpeer.timeout.usec = htonl(tp->timeout.usec);
1996 tpeer.nSent = htonl(tp->nSent);
1997 tpeer.reSends = htonl(tp->reSends);
1998 tpeer.inPacketSkew = htonl(tp->inPacketSkew);
1999 tpeer.outPacketSkew = htonl(tp->outPacketSkew);
2000 tpeer.rateFlag = htonl(tp->rateFlag);
2001 tpeer.natMTU = htons(tp->natMTU);
2002 tpeer.maxMTU = htons(tp->maxMTU);
2003 tpeer.maxDgramPackets = htons(tp->maxDgramPackets);
2004 tpeer.ifDgramPackets = htons(tp->ifDgramPackets);
2005 tpeer.MTU = htons(tp->MTU);
2006 tpeer.cwind = htons(tp->cwind);
2007 tpeer.nDgramPackets = htons(tp->nDgramPackets);
2008 tpeer.congestSeq = htons(tp->congestSeq);
2009 tpeer.bytesSent.high = htonl(tp->bytesSent.high);
2010 tpeer.bytesSent.low = htonl(tp->bytesSent.low);
2011 tpeer.bytesReceived.high =
2012 htonl(tp->bytesReceived.high);
2013 tpeer.bytesReceived.low =
2014 htonl(tp->bytesReceived.low);
2016 MUTEX_EXIT(&rx_peerHashTable_lock);
2017 rx_packetwrite(ap, 0, sizeof(struct rx_debugPeer),
2020 ap->length = sizeof(struct rx_debugPeer);
2021 rxi_SendDebugPacket(ap, asocket, ahost, aport,
2027 MUTEX_EXIT(&rx_peerHashTable_lock);
2029 /* if we make it here, there are no interesting packets */
2030 tpeer.host = htonl(0xffffffff); /* means end */
2031 rx_packetwrite(ap, 0, sizeof(struct rx_debugPeer),
2034 ap->length = sizeof(struct rx_debugPeer);
2035 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2040 case RX_DEBUGI_RXSTATS:{
2044 tl = sizeof(rx_stats) - ap->length;
2046 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
2050 /* Since its all int32s convert to network order with a loop. */
2051 if (rx_stats_active)
2052 MUTEX_ENTER(&rx_stats_mutex);
2053 s = (afs_int32 *) & rx_stats;
2054 for (i = 0; i < sizeof(rx_stats) / sizeof(afs_int32); i++, s++)
2055 rx_PutInt32(ap, i * sizeof(afs_int32), htonl(*s));
2058 ap->length = sizeof(rx_stats);
2059 if (rx_stats_active)
2060 MUTEX_EXIT(&rx_stats_mutex);
2061 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2067 /* error response packet */
2068 tin.type = htonl(RX_DEBUGI_BADTYPE);
2069 tin.index = tin.type;
2070 rx_packetwrite(ap, 0, sizeof(struct rx_debugIn), (char *)&tin);
2072 ap->length = sizeof(struct rx_debugIn);
2073 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2081 rxi_ReceiveVersionPacket(struct rx_packet *ap, osi_socket asocket,
2082 afs_int32 ahost, short aport, int istack)
2087 * Only respond to client-initiated version requests, and
2088 * clear that flag in the response.
2090 if (ap->header.flags & RX_CLIENT_INITIATED) {
2093 ap->header.flags = ap->header.flags & ~RX_CLIENT_INITIATED;
2094 rxi_EncodePacketHeader(ap);
2095 memset(buf, 0, sizeof(buf));
2096 strncpy(buf, cml_version_number + 4, sizeof(buf) - 1);
2097 rx_packetwrite(ap, 0, 65, buf);
2100 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2108 /* send a debug packet back to the sender */
2110 rxi_SendDebugPacket(struct rx_packet *apacket, osi_socket asocket,
2111 afs_int32 ahost, short aport, afs_int32 istack)
2113 struct sockaddr_in taddr;
2114 unsigned int i, nbytes, savelen = 0;
2117 int waslocked = ISAFS_GLOCK();
2120 taddr.sin_family = AF_INET;
2121 taddr.sin_port = aport;
2122 taddr.sin_addr.s_addr = ahost;
2123 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2124 taddr.sin_len = sizeof(struct sockaddr_in);
2127 /* We need to trim the niovecs. */
2128 nbytes = apacket->length;
2129 for (i = 1; i < apacket->niovecs; i++) {
2130 if (nbytes <= apacket->wirevec[i].iov_len) {
2131 savelen = apacket->wirevec[i].iov_len;
2132 saven = apacket->niovecs;
2133 apacket->wirevec[i].iov_len = nbytes;
2134 apacket->niovecs = i + 1; /* so condition fails because i == niovecs */
2136 nbytes -= apacket->wirevec[i].iov_len;
2139 #ifdef RX_KERNEL_TRACE
2140 if (ICL_SETACTIVE(afs_iclSetp)) {
2143 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2144 "before osi_NetSend()");
2152 /* debug packets are not reliably delivered, hence the cast below. */
2153 (void)osi_NetSend(asocket, &taddr, apacket->wirevec, apacket->niovecs,
2154 apacket->length + RX_HEADER_SIZE, istack);
2156 #ifdef RX_KERNEL_TRACE
2157 if (ICL_SETACTIVE(afs_iclSetp)) {
2159 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2160 "after osi_NetSend()");
2169 if (saven) { /* means we truncated the packet above. */
2170 apacket->wirevec[i - 1].iov_len = savelen;
2171 apacket->niovecs = saven;
2176 /* Send the packet to appropriate destination for the specified
2177 * call. The header is first encoded and placed in the packet.
2180 rxi_SendPacket(struct rx_call *call, struct rx_connection *conn,
2181 struct rx_packet *p, int istack)
2187 struct sockaddr_in addr;
2188 struct rx_peer *peer = conn->peer;
2191 char deliveryType = 'S';
2193 /* The address we're sending the packet to */
2194 memset(&addr, 0, sizeof(addr));
2195 addr.sin_family = AF_INET;
2196 addr.sin_port = peer->port;
2197 addr.sin_addr.s_addr = peer->host;
2199 /* This stuff should be revamped, I think, so that most, if not
2200 * all, of the header stuff is always added here. We could
2201 * probably do away with the encode/decode routines. XXXXX */
2203 /* Stamp each packet with a unique serial number. The serial
2204 * number is maintained on a connection basis because some types
2205 * of security may be based on the serial number of the packet,
2206 * and security is handled on a per authenticated-connection
2208 /* Pre-increment, to guarantee no zero serial number; a zero
2209 * serial number means the packet was never sent. */
2210 MUTEX_ENTER(&conn->conn_data_lock);
2211 p->header.serial = ++conn->serial;
2212 MUTEX_EXIT(&conn->conn_data_lock);
2213 /* This is so we can adjust retransmit time-outs better in the face of
2214 * rapidly changing round-trip times. RTO estimation is not a la Karn.
2216 if (p->firstSerial == 0) {
2217 p->firstSerial = p->header.serial;
2220 /* If an output tracer function is defined, call it with the packet and
2221 * network address. Note this function may modify its arguments. */
2222 if (rx_almostSent) {
2223 int drop = (*rx_almostSent) (p, &addr);
2224 /* drop packet if return value is non-zero? */
2226 deliveryType = 'D'; /* Drop the packet */
2230 /* Get network byte order header */
2231 rxi_EncodePacketHeader(p); /* XXX in the event of rexmit, etc, don't need to
2232 * touch ALL the fields */
2234 /* Send the packet out on the same socket that related packets are being
2238 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
2241 /* Possibly drop this packet, for testing purposes */
2242 if ((deliveryType == 'D')
2243 || ((rx_intentionallyDroppedPacketsPer100 > 0)
2244 && (random() % 100 < rx_intentionallyDroppedPacketsPer100))) {
2245 deliveryType = 'D'; /* Drop the packet */
2247 deliveryType = 'S'; /* Send the packet */
2248 #endif /* RXDEBUG */
2250 /* Loop until the packet is sent. We'd prefer just to use a
2251 * blocking socket, but unfortunately the interface doesn't
2252 * allow us to have the socket block in send mode, and not
2253 * block in receive mode */
2255 waslocked = ISAFS_GLOCK();
2256 #ifdef RX_KERNEL_TRACE
2257 if (ICL_SETACTIVE(afs_iclSetp)) {
2260 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2261 "before osi_NetSend()");
2270 osi_NetSend(socket, &addr, p->wirevec, p->niovecs,
2271 p->length + RX_HEADER_SIZE, istack)) != 0) {
2272 /* send failed, so let's hurry up the resend, eh? */
2273 if (rx_stats_active)
2274 rx_MutexIncrement(rx_stats.netSendFailures, rx_stats_mutex);
2275 p->retryTime = p->timeSent; /* resend it very soon */
2276 clock_Addmsec(&(p->retryTime),
2277 10 + (((afs_uint32) p->backoff) << 8));
2278 /* Some systems are nice and tell us right away that we cannot
2279 * reach this recipient by returning an error code.
2280 * So, when this happens let's "down" the host NOW so
2281 * we don't sit around waiting for this host to timeout later.
2285 (code == -1 && WSAGetLastError() == WSAEHOSTUNREACH) || (code == -WSAEHOSTUNREACH)
2286 #elif defined(AFS_LINUX20_ENV)
2287 code == -ENETUNREACH
2288 #elif defined(AFS_DARWIN_ENV)
2289 code == EHOSTUNREACH
2294 call->lastReceiveTime = 0;
2297 #ifdef RX_KERNEL_TRACE
2298 if (ICL_SETACTIVE(afs_iclSetp)) {
2300 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2301 "after osi_NetSend()");
2312 dpf(("%c %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%0.3d len %d",
2313 deliveryType, p->header.serial, rx_packetTypes[p->header.type - 1], ntohl(peer->host),
2314 ntohs(peer->port), p->header.serial, p->header.epoch, p->header.cid, p->header.callNumber,
2315 p->header.seq, p->header.flags, p, p->retryTime.sec, p->retryTime.usec / 1000, p->length));
2317 if (rx_stats_active)
2318 rx_MutexIncrement(rx_stats.packetsSent[p->header.type - 1], rx_stats_mutex);
2319 MUTEX_ENTER(&peer->peer_lock);
2320 hadd32(peer->bytesSent, p->length);
2321 MUTEX_EXIT(&peer->peer_lock);
2324 /* Send a list of packets to appropriate destination for the specified
2325 * connection. The headers are first encoded and placed in the packets.
2328 rxi_SendPacketList(struct rx_call *call, struct rx_connection *conn,
2329 struct rx_packet **list, int len, int istack)
2331 #if defined(AFS_SUN5_ENV) && defined(KERNEL)
2334 struct sockaddr_in addr;
2335 struct rx_peer *peer = conn->peer;
2337 struct rx_packet *p = NULL;
2338 struct iovec wirevec[RX_MAXIOVECS];
2339 int i, length, code;
2342 struct rx_jumboHeader *jp;
2344 char deliveryType = 'S';
2346 /* The address we're sending the packet to */
2347 addr.sin_family = AF_INET;
2348 addr.sin_port = peer->port;
2349 addr.sin_addr.s_addr = peer->host;
2351 if (len + 1 > RX_MAXIOVECS) {
2352 osi_Panic("rxi_SendPacketList, len > RX_MAXIOVECS\n");
2356 * Stamp the packets in this jumbogram with consecutive serial numbers
2358 MUTEX_ENTER(&conn->conn_data_lock);
2359 serial = conn->serial;
2360 conn->serial += len;
2361 MUTEX_EXIT(&conn->conn_data_lock);
2364 /* This stuff should be revamped, I think, so that most, if not
2365 * all, of the header stuff is always added here. We could
2366 * probably do away with the encode/decode routines. XXXXX */
2369 length = RX_HEADER_SIZE;
2370 wirevec[0].iov_base = (char *)(&list[0]->wirehead[0]);
2371 wirevec[0].iov_len = RX_HEADER_SIZE;
2372 for (i = 0; i < len; i++) {
2375 /* The whole 3.5 jumbogram scheme relies on packets fitting
2376 * in a single packet buffer. */
2377 if (p->niovecs > 2) {
2378 osi_Panic("rxi_SendPacketList, niovecs > 2\n");
2381 /* Set the RX_JUMBO_PACKET flags in all but the last packets
2384 if (p->length != RX_JUMBOBUFFERSIZE) {
2385 osi_Panic("rxi_SendPacketList, length != jumbo size\n");
2387 p->header.flags |= RX_JUMBO_PACKET;
2388 length += RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
2389 wirevec[i + 1].iov_len = RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
2391 wirevec[i + 1].iov_len = p->length;
2392 length += p->length;
2394 wirevec[i + 1].iov_base = (char *)(&p->localdata[0]);
2396 /* Convert jumbo packet header to network byte order */
2397 temp = (afs_uint32) (p->header.flags) << 24;
2398 temp |= (afs_uint32) (p->header.spare);
2399 *(afs_uint32 *) jp = htonl(temp);
2401 jp = (struct rx_jumboHeader *)
2402 ((char *)(&p->localdata[0]) + RX_JUMBOBUFFERSIZE);
2404 /* Stamp each packet with a unique serial number. The serial
2405 * number is maintained on a connection basis because some types
2406 * of security may be based on the serial number of the packet,
2407 * and security is handled on a per authenticated-connection
2409 /* Pre-increment, to guarantee no zero serial number; a zero
2410 * serial number means the packet was never sent. */
2411 p->header.serial = ++serial;
2412 /* This is so we can adjust retransmit time-outs better in the face of
2413 * rapidly changing round-trip times. RTO estimation is not a la Karn.
2415 if (p->firstSerial == 0) {
2416 p->firstSerial = p->header.serial;
2419 /* If an output tracer function is defined, call it with the packet and
2420 * network address. Note this function may modify its arguments. */
2421 if (rx_almostSent) {
2422 int drop = (*rx_almostSent) (p, &addr);
2423 /* drop packet if return value is non-zero? */
2425 deliveryType = 'D'; /* Drop the packet */
2429 /* Get network byte order header */
2430 rxi_EncodePacketHeader(p); /* XXX in the event of rexmit, etc, don't need to
2431 * touch ALL the fields */
2434 /* Send the packet out on the same socket that related packets are being
2438 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
2441 /* Possibly drop this packet, for testing purposes */
2442 if ((deliveryType == 'D')
2443 || ((rx_intentionallyDroppedPacketsPer100 > 0)
2444 && (random() % 100 < rx_intentionallyDroppedPacketsPer100))) {
2445 deliveryType = 'D'; /* Drop the packet */
2447 deliveryType = 'S'; /* Send the packet */
2448 #endif /* RXDEBUG */
2450 /* Loop until the packet is sent. We'd prefer just to use a
2451 * blocking socket, but unfortunately the interface doesn't
2452 * allow us to have the socket block in send mode, and not
2453 * block in receive mode */
2454 #if defined(AFS_SUN5_ENV) && defined(KERNEL)
2455 waslocked = ISAFS_GLOCK();
2456 if (!istack && waslocked)
2460 osi_NetSend(socket, &addr, &wirevec[0], len + 1, length,
2462 /* send failed, so let's hurry up the resend, eh? */
2463 if (rx_stats_active)
2464 rx_MutexIncrement(rx_stats.netSendFailures, rx_stats_mutex);
2465 for (i = 0; i < len; i++) {
2467 p->retryTime = p->timeSent; /* resend it very soon */
2468 clock_Addmsec(&(p->retryTime),
2469 10 + (((afs_uint32) p->backoff) << 8));
2471 /* Some systems are nice and tell us right away that we cannot
2472 * reach this recipient by returning an error code.
2473 * So, when this happens let's "down" the host NOW so
2474 * we don't sit around waiting for this host to timeout later.
2478 (code == -1 && WSAGetLastError() == WSAEHOSTUNREACH) || (code == -WSAEHOSTUNREACH)
2479 #elif defined(AFS_LINUX20_ENV)
2480 code == -ENETUNREACH
2481 #elif defined(AFS_DARWIN_ENV)
2482 code == EHOSTUNREACH
2487 call->lastReceiveTime = 0;
2489 #if defined(AFS_SUN5_ENV) && defined(KERNEL)
2490 if (!istack && waslocked)
2498 dpf(("%c %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%0.3d len %d",
2499 deliveryType, p->header.serial, rx_packetTypes[p->header.type - 1], ntohl(peer->host),
2500 ntohs(peer->port), p->header.serial, p->header.epoch, p->header.cid, p->header.callNumber,
2501 p->header.seq, p->header.flags, p, p->retryTime.sec, p->retryTime.usec / 1000, p->length));
2504 if (rx_stats_active)
2505 rx_MutexIncrement(rx_stats.packetsSent[p->header.type - 1], rx_stats_mutex);
2506 MUTEX_ENTER(&peer->peer_lock);
2507 hadd32(peer->bytesSent, p->length);
2508 MUTEX_EXIT(&peer->peer_lock);
2512 /* Send a "special" packet to the peer connection. If call is
2513 * specified, then the packet is directed to a specific call channel
2514 * associated with the connection, otherwise it is directed to the
2515 * connection only. Uses optionalPacket if it is supplied, rather than
2516 * allocating a new packet buffer. Nbytes is the length of the data
2517 * portion of the packet. If data is non-null, nbytes of data are
2518 * copied into the packet. Type is the type of the packet, as defined
2519 * in rx.h. Bug: there's a lot of duplication between this and other
2520 * routines. This needs to be cleaned up. */
2522 rxi_SendSpecial(struct rx_call *call,
2523 struct rx_connection *conn,
2524 struct rx_packet *optionalPacket, int type, char *data,
2525 int nbytes, int istack)
2527 /* Some of the following stuff should be common code for all
2528 * packet sends (it's repeated elsewhere) */
2529 struct rx_packet *p;
2531 int savelen = 0, saven = 0;
2532 int channel, callNumber;
2534 channel = call->channel;
2535 callNumber = *call->callNumber;
2536 /* BUSY packets refer to the next call on this connection */
2537 if (type == RX_PACKET_TYPE_BUSY) {
2546 p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
2548 osi_Panic("rxi_SendSpecial failure");
2555 p->header.serviceId = conn->serviceId;
2556 p->header.securityIndex = conn->securityIndex;
2557 p->header.cid = (conn->cid | channel);
2558 p->header.callNumber = callNumber;
2560 p->header.epoch = conn->epoch;
2561 p->header.type = type;
2562 p->header.flags = 0;
2563 if (conn->type == RX_CLIENT_CONNECTION)
2564 p->header.flags |= RX_CLIENT_INITIATED;
2566 rx_packetwrite(p, 0, nbytes, data);
2568 for (i = 1; i < p->niovecs; i++) {
2569 if (nbytes <= p->wirevec[i].iov_len) {
2570 savelen = p->wirevec[i].iov_len;
2572 p->wirevec[i].iov_len = nbytes;
2573 p->niovecs = i + 1; /* so condition fails because i == niovecs */
2575 nbytes -= p->wirevec[i].iov_len;
2579 rxi_Send(call, p, istack);
2581 rxi_SendPacket((struct rx_call *)0, conn, p, istack);
2582 if (saven) { /* means we truncated the packet above. We probably don't */
2583 /* really need to do this, but it seems safer this way, given that */
2584 /* sneaky optionalPacket... */
2585 p->wirevec[i - 1].iov_len = savelen;
2588 if (!optionalPacket)
2590 return optionalPacket;
2594 /* Encode the packet's header (from the struct header in the packet to
2595 * the net byte order representation in the wire representation of the
2596 * packet, which is what is actually sent out on the wire) */
2598 rxi_EncodePacketHeader(struct rx_packet *p)
2600 afs_uint32 *buf = (afs_uint32 *) (p->wirevec[0].iov_base); /* MTUXXX */
2602 memset(buf, 0, RX_HEADER_SIZE);
2603 *buf++ = htonl(p->header.epoch);
2604 *buf++ = htonl(p->header.cid);
2605 *buf++ = htonl(p->header.callNumber);
2606 *buf++ = htonl(p->header.seq);
2607 *buf++ = htonl(p->header.serial);
2608 *buf++ = htonl((((afs_uint32) p->header.type) << 24)
2609 | (((afs_uint32) p->header.flags) << 16)
2610 | (p->header.userStatus << 8) | p->header.securityIndex);
2611 /* Note: top 16 bits of this next word were reserved */
2612 *buf++ = htonl((p->header.spare << 16) | (p->header.serviceId & 0xffff));
2615 /* Decode the packet's header (from net byte order to a struct header) */
2617 rxi_DecodePacketHeader(struct rx_packet *p)
2619 afs_uint32 *buf = (afs_uint32 *) (p->wirevec[0].iov_base); /* MTUXXX */
2622 p->header.epoch = ntohl(*buf);
2624 p->header.cid = ntohl(*buf);
2626 p->header.callNumber = ntohl(*buf);
2628 p->header.seq = ntohl(*buf);
2630 p->header.serial = ntohl(*buf);
2636 /* C will truncate byte fields to bytes for me */
2637 p->header.type = temp >> 24;
2638 p->header.flags = temp >> 16;
2639 p->header.userStatus = temp >> 8;
2640 p->header.securityIndex = temp >> 0;
2645 p->header.serviceId = (temp & 0xffff);
2646 p->header.spare = temp >> 16;
2647 /* Note: top 16 bits of this last word are the security checksum */
2651 rxi_PrepareSendPacket(struct rx_call *call,
2652 struct rx_packet *p, int last)
2654 struct rx_connection *conn = call->conn;
2656 afs_int32 len; /* len must be a signed type; it can go negative */
2658 p->flags &= ~RX_PKTFLAG_ACKED;
2659 p->header.cid = (conn->cid | call->channel);
2660 p->header.serviceId = conn->serviceId;
2661 p->header.securityIndex = conn->securityIndex;
2663 /* No data packets on call 0. Where do these come from? */
2664 if (*call->callNumber == 0)
2665 *call->callNumber = 1;
2667 p->header.callNumber = *call->callNumber;
2668 p->header.seq = call->tnext++;
2669 p->header.epoch = conn->epoch;
2670 p->header.type = RX_PACKET_TYPE_DATA;
2671 p->header.flags = 0;
2672 p->header.spare = 0;
2673 if (conn->type == RX_CLIENT_CONNECTION)
2674 p->header.flags |= RX_CLIENT_INITIATED;
2677 p->header.flags |= RX_LAST_PACKET;
2679 clock_Zero(&p->retryTime); /* Never yet transmitted */
2680 clock_Zero(&p->firstSent); /* Never yet transmitted */
2681 p->header.serial = 0; /* Another way of saying never transmitted... */
2684 /* Now that we're sure this is the last data on the call, make sure
2685 * that the "length" and the sum of the iov_lens matches. */
2686 len = p->length + call->conn->securityHeaderSize;
2688 for (i = 1; i < p->niovecs && len > 0; i++) {
2689 len -= p->wirevec[i].iov_len;
2692 osi_Panic("PrepareSendPacket 1\n"); /* MTUXXX */
2693 } else if (i < p->niovecs) {
2694 /* Free any extra elements in the wirevec */
2695 #if defined(RX_ENABLE_TSFPQ)
2696 rxi_FreeDataBufsTSFPQ(p, i, 1 /* allow global pool flush if overquota */);
2697 #else /* !RX_ENABLE_TSFPQ */
2698 MUTEX_ENTER(&rx_freePktQ_lock);
2699 rxi_FreeDataBufsNoLock(p, i);
2700 MUTEX_EXIT(&rx_freePktQ_lock);
2701 #endif /* !RX_ENABLE_TSFPQ */
2706 p->wirevec[i - 1].iov_len += len;
2707 RXS_PreparePacket(conn->securityObject, call, p);
2710 /* Given an interface MTU size, calculate an adjusted MTU size that
2711 * will make efficient use of the RX buffers when the peer is sending
2712 * either AFS 3.4a jumbograms or AFS 3.5 jumbograms. */
2714 rxi_AdjustIfMTU(int mtu)
2719 if (rxi_nRecvFrags == 1 && rxi_nSendFrags == 1)
2721 adjMTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
2722 if (mtu <= adjMTU) {
2729 frags = mtu / (RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE);
2730 return (adjMTU + (frags * (RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE)));
2733 /* Given an interface MTU size, and the peer's advertised max receive
2734 * size, calculate an adjisted maxMTU size that makes efficient use
2735 * of our packet buffers when we are sending AFS 3.4a jumbograms. */
2737 rxi_AdjustMaxMTU(int mtu, int peerMaxMTU)
2739 int maxMTU = mtu * rxi_nSendFrags;
2740 maxMTU = MIN(maxMTU, peerMaxMTU);
2741 return rxi_AdjustIfMTU(maxMTU);
2744 /* Given a packet size, figure out how many datagram packet will fit.
2745 * The first buffer always contains RX_HEADER_SIZE+RX_JUMBOBUFFERSIZE+
2746 * RX_JUMBOHEADERSIZE, the middle buffers contain RX_JUMBOBUFFERSIZE+
2747 * RX_JUMBOHEADERSIZE, and the last buffer contains RX_JUMBOBUFFERSIZE */
2749 rxi_AdjustDgramPackets(int frags, int mtu)
2752 if (mtu + IPv6_FRAG_HDR_SIZE < RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE) {
2755 maxMTU = (frags * (mtu + UDP_HDR_SIZE)) - UDP_HDR_SIZE;
2756 maxMTU = MIN(maxMTU, RX_MAX_PACKET_SIZE);
2757 /* subtract the size of the first and last packets */
2758 maxMTU -= RX_HEADER_SIZE + (2 * RX_JUMBOBUFFERSIZE) + RX_JUMBOHEADERSIZE;
2762 return (2 + (maxMTU / (RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE)));
2767 * This function can be used by the Windows Cache Manager
2768 * to dump the list of all rx packets so that we can determine
2769 * where the packet leakage is.
2771 int rx_DumpPackets(FILE *outputFile, char *cookie)
2773 #ifdef RXDEBUG_PACKET
2774 struct rx_packet *p;
2778 #define RXDPRINTF sprintf
2779 #define RXDPRINTOUT output
2781 #define RXDPRINTF fprintf
2782 #define RXDPRINTOUT outputFile
2786 MUTEX_ENTER(&rx_freePktQ_lock);
2787 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Packets - count=%u\r\n", cookie, rx_packet_id);
2789 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2792 for (p = rx_mallocedP; p; p = p->allNextp) {
2793 RXDPRINTF(RXDPRINTOUT, "%s - packet=0x%p, id=%u, firstSent=%u.%08u, timeSent=%u.%08u, retryTime=%u.%08u, firstSerial=%u, niovecs=%u, flags=0x%x, backoff=%u, length=%u header: epoch=%u, cid=%u, callNum=%u, seq=%u, serial=%u, type=%u, flags=0x%x, userStatus=%u, securityIndex=%u, serviceId=%u\r\n",
2794 cookie, p, p->packetId, p->firstSent.sec, p->firstSent.usec, p->timeSent.sec, p->timeSent.usec, p->retryTime.sec, p->retryTime.usec,
2795 p->firstSerial, p->niovecs, (afs_uint32)p->flags, (afs_uint32)p->backoff, (afs_uint32)p->length,
2796 p->header.epoch, p->header.cid, p->header.callNumber, p->header.seq, p->header.serial,
2797 (afs_uint32)p->header.type, (afs_uint32)p->header.flags, (afs_uint32)p->header.userStatus,
2798 (afs_uint32)p->header.securityIndex, (afs_uint32)p->header.serviceId);
2800 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2804 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Packets\r\n", cookie);
2806 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2809 MUTEX_EXIT(&rx_freePktQ_lock);
2811 #endif /* RXDEBUG_PACKET */