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>
11 #include <afs/param.h>
15 # include "afs/sysincludes.h"
16 # include "afsincludes.h"
17 # include "rx_kcommon.h"
18 # else /* defined(UKERNEL) */
19 # ifdef RX_KERNEL_TRACE
20 # include "rx_kcommon.h"
23 # ifndef AFS_LINUX20_ENV
26 # if defined(AFS_SGI_ENV) || defined(AFS_HPUX110_ENV) || defined(AFS_NBSD50_ENV)
27 # include "afs/sysincludes.h"
29 # if defined(AFS_OBSD_ENV)
32 # include "h/socket.h"
33 # if !defined(AFS_SUN5_ENV) && !defined(AFS_LINUX20_ENV) && !defined(AFS_HPUX110_ENV)
34 # if !defined(AFS_AIX41_ENV)
35 # include "sys/mount.h" /* it gets pulled in by something later anyway */
39 # include "netinet/in.h"
40 # include "afs/afs_osi.h"
41 # include "rx_kmutex.h"
42 # endif /* defined(UKERNEL) */
47 # if defined(AFS_NT40_ENV)
49 # define EWOULDBLOCK WSAEWOULDBLOCK
52 # include "rx_xmit_nt.h"
58 # include <sys/sysmacros.h>
61 #include <opr/queue.h>
65 #include "rx_packet.h"
66 #include "rx_atomic.h"
67 #include "rx_globals.h"
68 #include "rx_internal.h"
76 * \brief structure used to keep track of allocated packets
78 struct rx_mallocedPacket {
79 struct opr_queue entry; /*!< chained using opr_queue */
80 struct rx_packet *addr; /*!< address of the first element */
81 afs_uint32 size; /*!< array size in bytes */
85 /* rxdb_fileID is used to identify the lock location, along with line#. */
86 static int rxdb_fileID = RXDB_FILE_RX_PACKET;
87 #endif /* RX_LOCKS_DB */
88 static struct rx_packet *rx_mallocedP = 0;
90 static afs_uint32 rx_packet_id = 0;
93 extern char cml_version_number[];
95 static int AllocPacketBufs(int class, int num_pkts, struct opr_queue *q);
97 static void rxi_SendDebugPacket(struct rx_packet *apacket, osi_socket asocket,
98 afs_uint32 ahost, short aport,
100 static struct rx_packet *rxi_AllocPacketNoLock(int class);
103 static void rxi_MorePacketsNoLock(int apackets);
106 #ifdef RX_ENABLE_TSFPQ
107 static int rxi_FreeDataBufsTSFPQ(struct rx_packet *p, afs_uint32 first,
109 static void rxi_AdjustLocalPacketsTSFPQ(int num_keep_local,
110 int allow_overcommit);
112 static void rxi_FreePacketNoLock(struct rx_packet *p);
113 static int rxi_FreeDataBufsNoLock(struct rx_packet *p, afs_uint32 first);
114 static int rxi_FreeDataBufsToQueue(struct rx_packet *p, afs_uint32 first,
115 struct opr_queue * q);
118 extern struct opr_queue rx_idleServerQueue;
120 /* some rules about packets:
121 * 1. When a packet is allocated, the final iov_buf contains room for
122 * a security trailer, but iov_len masks that fact. If the security
123 * package wants to add the trailer, it may do so, and then extend
124 * iov_len appropriately. For this reason, packet's niovecs and
125 * iov_len fields should be accurate before calling PreparePacket.
129 * all packet buffers (iov_base) are integral multiples of
131 * offset is an integral multiple of the word size.
134 rx_SlowGetInt32(struct rx_packet *packet, size_t offset)
138 for (l = 0, i = 1; i < packet->niovecs; i++) {
139 if (l + packet->wirevec[i].iov_len > offset) {
141 *((afs_int32 *) ((char *)(packet->wirevec[i].iov_base) +
144 l += packet->wirevec[i].iov_len;
151 * all packet buffers (iov_base) are integral multiples of the word size.
152 * offset is an integral multiple of the word size.
155 rx_SlowPutInt32(struct rx_packet * packet, size_t offset, afs_int32 data)
159 for (l = 0, i = 1; i < packet->niovecs; i++) {
160 if (l + packet->wirevec[i].iov_len > offset) {
161 *((afs_int32 *) ((char *)(packet->wirevec[i].iov_base) +
162 (offset - l))) = data;
165 l += packet->wirevec[i].iov_len;
172 * all packet buffers (iov_base) are integral multiples of the
174 * offset is an integral multiple of the word size.
176 * all buffers are contiguously arrayed in the iovec from 0..niovecs-1
179 rx_SlowReadPacket(struct rx_packet * packet, unsigned int offset, int resid,
182 unsigned int i, j, l, r;
183 for (l = 0, i = 1; i < packet->niovecs; i++) {
184 if (l + packet->wirevec[i].iov_len > offset) {
187 l += packet->wirevec[i].iov_len;
190 /* i is the iovec which contains the first little bit of data in which we
191 * are interested. l is the total length of everything prior to this iovec.
192 * j is the number of bytes we can safely copy out of this iovec.
193 * offset only applies to the first iovec.
196 while ((r > 0) && (i < packet->niovecs)) {
197 j = MIN(r, packet->wirevec[i].iov_len - (offset - l));
198 memcpy(out, (char *)(packet->wirevec[i].iov_base) + (offset - l), j);
201 l += packet->wirevec[i].iov_len;
206 return (r ? (resid - r) : resid);
211 * all packet buffers (iov_base) are integral multiples of the
213 * offset is an integral multiple of the word size.
216 rx_SlowWritePacket(struct rx_packet * packet, int offset, int resid, char *in)
218 unsigned int i, j, l, o, r;
221 for (l = 0, i = 1, o = offset; i < packet->niovecs; i++) {
222 if (l + packet->wirevec[i].iov_len > o) {
225 l += packet->wirevec[i].iov_len;
228 /* i is the iovec which contains the first little bit of data in which we
229 * are interested. l is the total length of everything prior to this iovec.
230 * j is the number of bytes we can safely copy out of this iovec.
231 * offset only applies to the first iovec.
234 while ((r > 0) && (i <= RX_MAXWVECS)) {
235 if (i >= packet->niovecs)
236 if (rxi_AllocDataBuf(packet, r, RX_PACKET_CLASS_SEND_CBUF) > 0) /* ++niovecs as a side-effect */
239 b = (char *)(packet->wirevec[i].iov_base) + (offset - l);
240 j = MIN(r, packet->wirevec[i].iov_len - (offset - l));
244 l += packet->wirevec[i].iov_len;
249 return (r ? (resid - r) : resid);
253 rxi_AllocPackets(int class, int num_pkts, struct opr_queue * q)
257 num_pkts = AllocPacketBufs(class, num_pkts, q);
259 for (opr_queue_Scan(q, c)) {
260 RX_PACKET_IOV_FULLINIT(opr_queue_Entry(c, struct rx_packet, entry));
266 #ifdef RX_ENABLE_TSFPQ
268 AllocPacketBufs(int class, int num_pkts, struct opr_queue * q)
270 struct rx_ts_info_t * rx_ts_info;
274 RX_TS_INFO_GET(rx_ts_info);
276 transfer = num_pkts - rx_ts_info->_FPQ.len;
279 MUTEX_ENTER(&rx_freePktQ_lock);
280 transfer = MAX(transfer, rx_TSFPQGlobSize);
281 if (transfer > rx_nFreePackets) {
282 /* alloc enough for us, plus a few globs for other threads */
283 rxi_MorePacketsNoLock(transfer + 4 * rx_initSendWindow);
286 RX_TS_FPQ_GTOL2(rx_ts_info, transfer);
288 MUTEX_EXIT(&rx_freePktQ_lock);
292 RX_TS_FPQ_QCHECKOUT(rx_ts_info, num_pkts, q);
296 #else /* RX_ENABLE_TSFPQ */
298 AllocPacketBufs(int class, int num_pkts, struct opr_queue * q)
309 MUTEX_ENTER(&rx_freePktQ_lock);
312 for (; (num_pkts > 0) && (rxi_OverQuota2(class,num_pkts));
313 num_pkts--, overq++);
316 rxi_NeedMorePackets = TRUE;
317 if (rx_stats_active) {
319 case RX_PACKET_CLASS_RECEIVE:
320 rx_atomic_inc(&rx_stats.receivePktAllocFailures);
322 case RX_PACKET_CLASS_SEND:
323 rx_atomic_inc(&rx_stats.sendPktAllocFailures);
325 case RX_PACKET_CLASS_SPECIAL:
326 rx_atomic_inc(&rx_stats.specialPktAllocFailures);
328 case RX_PACKET_CLASS_RECV_CBUF:
329 rx_atomic_inc(&rx_stats.receiveCbufPktAllocFailures);
331 case RX_PACKET_CLASS_SEND_CBUF:
332 rx_atomic_inc(&rx_stats.sendCbufPktAllocFailures);
338 if (rx_nFreePackets < num_pkts)
339 num_pkts = rx_nFreePackets;
342 rxi_NeedMorePackets = TRUE;
346 if (rx_nFreePackets < num_pkts) {
347 rxi_MorePacketsNoLock(MAX((num_pkts-rx_nFreePackets), 4 * rx_initSendWindow));
351 for (i=0, c=opr_queue_First(&rx_freePacketQueue, struct rx_packet, entry);
353 i++, c=opr_queue_Next(&c->entry, struct rx_packet, entry)) {
357 opr_queue_SplitBeforeAppend(&rx_freePacketQueue, q, &c->entry);
359 rx_nFreePackets -= num_pkts;
364 MUTEX_EXIT(&rx_freePktQ_lock);
369 #endif /* RX_ENABLE_TSFPQ */
372 * Free a packet currently used as a continuation buffer
374 #ifdef RX_ENABLE_TSFPQ
375 /* num_pkts=0 means queue length is unknown */
377 rxi_FreePackets(int num_pkts, struct opr_queue * q)
379 struct rx_ts_info_t * rx_ts_info;
380 struct opr_queue *cursor, *store;
383 osi_Assert(num_pkts >= 0);
384 RX_TS_INFO_GET(rx_ts_info);
387 for (opr_queue_ScanSafe(q, cursor, store)) {
389 rxi_FreeDataBufsTSFPQ(opr_queue_Entry(cursor, struct rx_packet,
393 for (opr_queue_ScanSafe(q, cursor, store)) {
394 rxi_FreeDataBufsTSFPQ(opr_queue_Entry(cursor, struct rx_packet,
400 RX_TS_FPQ_QCHECKIN(rx_ts_info, num_pkts, q);
403 if (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax) {
405 MUTEX_ENTER(&rx_freePktQ_lock);
407 RX_TS_FPQ_LTOG(rx_ts_info);
409 /* Wakeup anyone waiting for packets */
412 MUTEX_EXIT(&rx_freePktQ_lock);
418 #else /* RX_ENABLE_TSFPQ */
419 /* num_pkts=0 means queue length is unknown */
421 rxi_FreePackets(int num_pkts, struct opr_queue *q)
423 struct opr_queue cbs;
424 struct opr_queue *cursor, *store;
428 osi_Assert(num_pkts >= 0);
429 opr_queue_Init(&cbs);
432 for (opr_queue_ScanSafe(q, cursor, store)) {
434 = opr_queue_Entry(cursor, struct rx_packet, entry);
435 if (p->niovecs > 2) {
436 qlen += rxi_FreeDataBufsToQueue(p, 2, &cbs);
444 for (opr_queue_ScanSafe(q, cursor, store)) {
446 = opr_queue_Entry(cursor, struct rx_packet, entry);
448 if (p->niovecs > 2) {
449 qlen += rxi_FreeDataBufsToQueue(p, 2, &cbs);
456 opr_queue_SpliceAppend(q, &cbs);
462 MUTEX_ENTER(&rx_freePktQ_lock);
464 opr_queue_SpliceAppend(&rx_freePacketQueue, q);
465 rx_nFreePackets += qlen;
467 /* Wakeup anyone waiting for packets */
470 MUTEX_EXIT(&rx_freePktQ_lock);
475 #endif /* RX_ENABLE_TSFPQ */
477 /* this one is kind of awful.
478 * In rxkad, the packet has been all shortened, and everything, ready for
479 * sending. All of a sudden, we discover we need some of that space back.
480 * This isn't terribly general, because it knows that the packets are only
481 * rounded up to the EBS (userdata + security header).
484 rxi_RoundUpPacket(struct rx_packet *p, unsigned int nb)
488 if (p->wirevec[i].iov_base == (caddr_t) p->localdata) {
489 if (p->wirevec[i].iov_len <= RX_FIRSTBUFFERSIZE - nb) {
490 p->wirevec[i].iov_len += nb;
494 if (p->wirevec[i].iov_len <= RX_CBUFFERSIZE - nb) {
495 p->wirevec[i].iov_len += nb;
503 /* get sufficient space to store nb bytes of data (or more), and hook
504 * it into the supplied packet. Return nbytes<=0 if successful, otherwise
505 * returns the number of bytes >0 which it failed to come up with.
506 * Don't need to worry about locking on packet, since only
507 * one thread can manipulate one at a time. Locking on continution
508 * packets is handled by AllocPacketBufs */
509 /* MTUXXX don't need to go throught the for loop if we can trust niovecs */
511 rxi_AllocDataBuf(struct rx_packet *p, int nb, int class)
514 struct opr_queue q, *cursor, *store;
516 /* compute the number of cbuf's we need */
517 nv = nb / RX_CBUFFERSIZE;
518 if ((nv * RX_CBUFFERSIZE) < nb)
520 if ((nv + p->niovecs) > RX_MAXWVECS)
521 nv = RX_MAXWVECS - p->niovecs;
525 /* allocate buffers */
527 nv = AllocPacketBufs(class, nv, &q);
529 /* setup packet iovs */
531 for (opr_queue_ScanSafe(&q, cursor, store)) {
533 = opr_queue_Entry(cursor, struct rx_packet, entry);
535 opr_queue_Remove(&cb->entry);
536 p->wirevec[i].iov_base = (caddr_t) cb->localdata;
537 p->wirevec[i].iov_len = RX_CBUFFERSIZE;
541 nb -= (nv * RX_CBUFFERSIZE);
542 p->length += (nv * RX_CBUFFERSIZE);
549 * Register allocated packets.
551 * @param[in] addr array of packets
552 * @param[in] npkt number of packets
557 registerPackets(struct rx_packet *addr, afs_uint32 npkt)
559 struct rx_mallocedPacket *mp;
561 mp = osi_Alloc(sizeof(*mp));
563 osi_Assert(mp != NULL);
564 memset(mp, 0, sizeof(*mp));
567 mp->size = npkt * sizeof(struct rx_packet);
568 osi_Assert(npkt <= MAX_AFS_UINT32 / sizeof(struct rx_packet));
570 MUTEX_ENTER(&rx_mallocedPktQ_lock);
571 opr_queue_Append(&rx_mallocedPacketQueue, &mp->entry);
572 MUTEX_EXIT(&rx_mallocedPktQ_lock);
575 /* Add more packet buffers */
576 #ifdef RX_ENABLE_TSFPQ
578 rxi_MorePackets(int apackets)
580 struct rx_packet *p, *e;
581 struct rx_ts_info_t * rx_ts_info;
585 getme = apackets * sizeof(struct rx_packet);
586 p = osi_Alloc(getme);
588 registerPackets(p, apackets);
590 PIN(p, getme); /* XXXXX */
592 RX_TS_INFO_GET(rx_ts_info);
594 RX_TS_FPQ_LOCAL_ALLOC(rx_ts_info,apackets);
595 /* TSFPQ patch also needs to keep track of total packets */
597 MUTEX_ENTER(&rx_packets_mutex);
598 rx_nPackets += apackets;
599 RX_TS_FPQ_COMPUTE_LIMITS;
600 MUTEX_EXIT(&rx_packets_mutex);
602 for (e = p + apackets; p < e; p++) {
603 RX_PACKET_IOV_INIT(p);
606 RX_TS_FPQ_CHECKIN(rx_ts_info,p);
609 MUTEX_ENTER(&rx_freePktQ_lock);
610 #ifdef RXDEBUG_PACKET
611 p->packetId = rx_packet_id++;
612 p->allNextp = rx_mallocedP;
613 #endif /* RXDEBUG_PACKET */
615 MUTEX_EXIT(&rx_freePktQ_lock);
618 rx_ts_info->_FPQ.delta += apackets;
620 if (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax) {
622 MUTEX_ENTER(&rx_freePktQ_lock);
624 RX_TS_FPQ_LTOG(rx_ts_info);
625 rxi_NeedMorePackets = FALSE;
628 MUTEX_EXIT(&rx_freePktQ_lock);
632 #else /* RX_ENABLE_TSFPQ */
634 rxi_MorePackets(int apackets)
636 struct rx_packet *p, *e;
640 getme = apackets * sizeof(struct rx_packet);
641 p = osi_Alloc(getme);
643 registerPackets(p, apackets);
645 PIN(p, getme); /* XXXXX */
648 MUTEX_ENTER(&rx_freePktQ_lock);
650 for (e = p + apackets; p < e; p++) {
651 RX_PACKET_IOV_INIT(p);
652 #ifdef RX_TRACK_PACKETS
653 p->flags |= RX_PKTFLAG_FREE;
657 opr_queue_Append(&rx_freePacketQueue, &p->entry);
658 #ifdef RXDEBUG_PACKET
659 p->packetId = rx_packet_id++;
660 p->allNextp = rx_mallocedP;
661 #endif /* RXDEBUG_PACKET */
665 rx_nPackets += apackets;
666 rx_nFreePackets += apackets;
667 rxi_NeedMorePackets = FALSE;
670 MUTEX_EXIT(&rx_freePktQ_lock);
673 #endif /* RX_ENABLE_TSFPQ */
675 #ifdef RX_ENABLE_TSFPQ
677 rxi_MorePacketsTSFPQ(int apackets, int flush_global, int num_keep_local)
679 struct rx_packet *p, *e;
680 struct rx_ts_info_t * rx_ts_info;
684 getme = apackets * sizeof(struct rx_packet);
685 p = osi_Alloc(getme);
686 registerPackets(p, apackets);
688 PIN(p, getme); /* XXXXX */
690 RX_TS_INFO_GET(rx_ts_info);
692 RX_TS_FPQ_LOCAL_ALLOC(rx_ts_info,apackets);
693 /* TSFPQ patch also needs to keep track of total packets */
694 MUTEX_ENTER(&rx_packets_mutex);
695 rx_nPackets += apackets;
696 RX_TS_FPQ_COMPUTE_LIMITS;
697 MUTEX_EXIT(&rx_packets_mutex);
699 for (e = p + apackets; p < e; p++) {
700 RX_PACKET_IOV_INIT(p);
702 RX_TS_FPQ_CHECKIN(rx_ts_info,p);
705 MUTEX_ENTER(&rx_freePktQ_lock);
706 #ifdef RXDEBUG_PACKET
707 p->packetId = rx_packet_id++;
708 p->allNextp = rx_mallocedP;
709 #endif /* RXDEBUG_PACKET */
711 MUTEX_EXIT(&rx_freePktQ_lock);
714 rx_ts_info->_FPQ.delta += apackets;
717 (num_keep_local < apackets)) {
719 MUTEX_ENTER(&rx_freePktQ_lock);
721 RX_TS_FPQ_LTOG2(rx_ts_info, (apackets - num_keep_local));
722 rxi_NeedMorePackets = FALSE;
725 MUTEX_EXIT(&rx_freePktQ_lock);
729 #endif /* RX_ENABLE_TSFPQ */
732 /* Add more packet buffers */
734 rxi_MorePacketsNoLock(int apackets)
736 #ifdef RX_ENABLE_TSFPQ
737 struct rx_ts_info_t * rx_ts_info;
738 #endif /* RX_ENABLE_TSFPQ */
739 struct rx_packet *p, *e;
742 /* allocate enough packets that 1/4 of the packets will be able
743 * to hold maximal amounts of data */
744 apackets += (apackets / 4)
745 * ((rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE) / RX_CBUFFERSIZE);
747 getme = apackets * sizeof(struct rx_packet);
748 p = osi_Alloc(getme);
750 apackets -= apackets / 4;
751 osi_Assert(apackets > 0);
755 registerPackets(p, apackets);
757 #ifdef RX_ENABLE_TSFPQ
758 RX_TS_INFO_GET(rx_ts_info);
759 RX_TS_FPQ_GLOBAL_ALLOC(rx_ts_info,apackets);
760 #endif /* RX_ENABLE_TSFPQ */
762 for (e = p + apackets; p < e; p++) {
763 RX_PACKET_IOV_INIT(p);
764 #ifdef RX_TRACK_PACKETS
765 p->flags |= RX_PKTFLAG_FREE;
769 opr_queue_Append(&rx_freePacketQueue, &p->entry);
770 #ifdef RXDEBUG_PACKET
771 p->packetId = rx_packet_id++;
772 p->allNextp = rx_mallocedP;
773 #endif /* RXDEBUG_PACKET */
777 rx_nFreePackets += apackets;
778 MUTEX_ENTER(&rx_packets_mutex);
779 rx_nPackets += apackets;
780 #ifdef RX_ENABLE_TSFPQ
781 RX_TS_FPQ_COMPUTE_LIMITS;
782 #endif /* RX_ENABLE_TSFPQ */
783 MUTEX_EXIT(&rx_packets_mutex);
784 rxi_NeedMorePackets = FALSE;
790 rxi_FreeAllPackets(void)
792 struct rx_mallocedPacket *mp;
794 MUTEX_ENTER(&rx_mallocedPktQ_lock);
796 while (!opr_queue_IsEmpty(&rx_mallocedPacketQueue)) {
797 mp = opr_queue_First(&rx_mallocedPacketQueue,
798 struct rx_mallocedPacket, entry);
799 opr_queue_Remove(&mp->entry);
800 osi_Free(mp->addr, mp->size);
801 UNPIN(mp->addr, mp->size);
802 osi_Free(mp, sizeof(*mp));
804 MUTEX_EXIT(&rx_mallocedPktQ_lock);
807 #ifdef RX_ENABLE_TSFPQ
809 rxi_AdjustLocalPacketsTSFPQ(int num_keep_local, int allow_overcommit)
811 struct rx_ts_info_t * rx_ts_info;
815 RX_TS_INFO_GET(rx_ts_info);
817 if (num_keep_local != rx_ts_info->_FPQ.len) {
819 MUTEX_ENTER(&rx_freePktQ_lock);
820 if (num_keep_local < rx_ts_info->_FPQ.len) {
821 xfer = rx_ts_info->_FPQ.len - num_keep_local;
822 RX_TS_FPQ_LTOG2(rx_ts_info, xfer);
825 xfer = num_keep_local - rx_ts_info->_FPQ.len;
826 if ((num_keep_local > rx_TSFPQLocalMax) && !allow_overcommit)
827 xfer = rx_TSFPQLocalMax - rx_ts_info->_FPQ.len;
828 if (rx_nFreePackets < xfer) {
829 rxi_MorePacketsNoLock(MAX(xfer - rx_nFreePackets, 4 * rx_initSendWindow));
831 RX_TS_FPQ_GTOL2(rx_ts_info, xfer);
833 MUTEX_EXIT(&rx_freePktQ_lock);
839 rxi_FlushLocalPacketsTSFPQ(void)
841 rxi_AdjustLocalPacketsTSFPQ(0, 0);
843 #endif /* RX_ENABLE_TSFPQ */
845 /* Allocate more packets iff we need more continuation buffers */
846 /* In kernel, can't page in memory with interrupts disabled, so we
847 * don't use the event mechanism. */
849 rx_CheckPackets(void)
851 if (rxi_NeedMorePackets) {
852 rxi_MorePackets(rx_maxSendWindow);
856 /* In the packet freeing routine below, the assumption is that
857 we want all of the packets to be used equally frequently, so that we
858 don't get packet buffers paging out. It would be just as valid to
859 assume that we DO want them to page out if not many are being used.
860 In any event, we assume the former, and append the packets to the end
862 /* This explanation is bogus. The free list doesn't remain in any kind of
863 useful order for afs_int32: the packets in use get pretty much randomly scattered
864 across all the pages. In order to permit unused {packets,bufs} to page out, they
865 must be stored so that packets which are adjacent in memory are adjacent in the
866 free list. An array springs rapidly to mind.
869 /* Actually free the packet p. */
870 #ifndef RX_ENABLE_TSFPQ
872 rxi_FreePacketNoLock(struct rx_packet *p)
874 dpf(("Free %p\n", p));
878 opr_queue_Append(&rx_freePacketQueue, &p->entry);
880 #endif /* RX_ENABLE_TSFPQ */
882 #ifdef RX_ENABLE_TSFPQ
884 rxi_FreePacketTSFPQ(struct rx_packet *p, int flush_global)
886 struct rx_ts_info_t * rx_ts_info;
887 dpf(("Free %p\n", p));
889 RX_TS_INFO_GET(rx_ts_info);
890 RX_TS_FPQ_CHECKIN(rx_ts_info,p);
892 if (flush_global && (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax)) {
894 MUTEX_ENTER(&rx_freePktQ_lock);
896 RX_TS_FPQ_LTOG(rx_ts_info);
898 /* Wakeup anyone waiting for packets */
901 MUTEX_EXIT(&rx_freePktQ_lock);
905 #endif /* RX_ENABLE_TSFPQ */
908 * free continuation buffers off a packet into a queue
910 * [IN] p -- packet from which continuation buffers will be freed
911 * [IN] first -- iovec offset of first continuation buffer to free
912 * [IN] q -- queue into which continuation buffers will be chained
915 * number of continuation buffers freed
917 #ifndef RX_ENABLE_TSFPQ
919 rxi_FreeDataBufsToQueue(struct rx_packet *p, afs_uint32 first, struct opr_queue * q)
922 struct rx_packet * cb;
925 for (first = MAX(2, first); first < p->niovecs; first++, count++) {
926 iov = &p->wirevec[first];
928 osi_Panic("rxi_FreeDataBufsToQueue: unexpected NULL iov");
929 cb = RX_CBUF_TO_PACKET(iov->iov_base, p);
930 RX_FPQ_MARK_FREE(cb);
931 opr_queue_Append(q, &cb->entry);
940 * free packet continuation buffers into the global free packet pool
942 * [IN] p -- packet from which to free continuation buffers
943 * [IN] first -- iovec offset of first continuation buffer to free
949 rxi_FreeDataBufsNoLock(struct rx_packet *p, afs_uint32 first)
953 for (first = MAX(2, first); first < p->niovecs; first++) {
954 iov = &p->wirevec[first];
956 osi_Panic("rxi_FreeDataBufsNoLock: unexpected NULL iov");
957 rxi_FreePacketNoLock(RX_CBUF_TO_PACKET(iov->iov_base, p));
968 * free packet continuation buffers into the thread-local free pool
970 * [IN] p -- packet from which continuation buffers will be freed
971 * [IN] first -- iovec offset of first continuation buffer to free
972 * any value less than 2, the min number of iovecs,
973 * is treated as if it is 2.
974 * [IN] flush_global -- if nonzero, we will flush overquota packets to the
975 * global free pool before returning
981 rxi_FreeDataBufsTSFPQ(struct rx_packet *p, afs_uint32 first, int flush_global)
984 struct rx_ts_info_t * rx_ts_info;
986 RX_TS_INFO_GET(rx_ts_info);
988 for (first = MAX(2, first); first < p->niovecs; first++) {
989 iov = &p->wirevec[first];
991 osi_Panic("rxi_FreeDataBufsTSFPQ: unexpected NULL iov");
992 RX_TS_FPQ_CHECKIN(rx_ts_info,RX_CBUF_TO_PACKET(iov->iov_base, p));
997 if (flush_global && (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax)) {
999 MUTEX_ENTER(&rx_freePktQ_lock);
1001 RX_TS_FPQ_LTOG(rx_ts_info);
1003 /* Wakeup anyone waiting for packets */
1004 rxi_PacketsUnWait();
1006 MUTEX_EXIT(&rx_freePktQ_lock);
1011 #endif /* RX_ENABLE_TSFPQ */
1013 int rxi_nBadIovecs = 0;
1015 /* rxi_RestoreDataBufs
1017 * Restore the correct sizes to the iovecs. Called when reusing a packet
1018 * for reading off the wire.
1021 rxi_RestoreDataBufs(struct rx_packet *p)
1026 RX_PACKET_IOV_INIT(p);
1028 for (i = 2, iov = &p->wirevec[2]; i < p->niovecs; i++, iov++) {
1029 if (!iov->iov_base) {
1034 iov->iov_len = RX_CBUFFERSIZE;
1038 #ifdef RX_ENABLE_TSFPQ
1040 rxi_TrimDataBufs(struct rx_packet *p, int first)
1043 struct iovec *iov, *end;
1044 struct rx_ts_info_t * rx_ts_info;
1048 osi_Panic("TrimDataBufs 1: first must be 1");
1050 /* Skip over continuation buffers containing message data */
1051 iov = &p->wirevec[2];
1052 end = iov + (p->niovecs - 2);
1053 length = p->length - p->wirevec[1].iov_len;
1054 for (; iov < end && length > 0; iov++) {
1056 osi_Panic("TrimDataBufs 3: vecs 1-niovecs must not be NULL");
1057 length -= iov->iov_len;
1060 /* iov now points to the first empty data buffer. */
1064 RX_TS_INFO_GET(rx_ts_info);
1065 for (; iov < end; iov++) {
1067 osi_Panic("TrimDataBufs 4: vecs 2-niovecs must not be NULL");
1068 RX_TS_FPQ_CHECKIN(rx_ts_info,RX_CBUF_TO_PACKET(iov->iov_base, p));
1071 if (rx_ts_info->_FPQ.len > rx_TSFPQLocalMax) {
1073 MUTEX_ENTER(&rx_freePktQ_lock);
1075 RX_TS_FPQ_LTOG(rx_ts_info);
1076 rxi_PacketsUnWait();
1078 MUTEX_EXIT(&rx_freePktQ_lock);
1084 #else /* RX_ENABLE_TSFPQ */
1086 rxi_TrimDataBufs(struct rx_packet *p, int first)
1089 struct iovec *iov, *end;
1093 osi_Panic("TrimDataBufs 1: first must be 1");
1095 /* Skip over continuation buffers containing message data */
1096 iov = &p->wirevec[2];
1097 end = iov + (p->niovecs - 2);
1098 length = p->length - p->wirevec[1].iov_len;
1099 for (; iov < end && length > 0; iov++) {
1101 osi_Panic("TrimDataBufs 3: vecs 1-niovecs must not be NULL");
1102 length -= iov->iov_len;
1105 /* iov now points to the first empty data buffer. */
1110 MUTEX_ENTER(&rx_freePktQ_lock);
1112 for (; iov < end; iov++) {
1114 osi_Panic("TrimDataBufs 4: vecs 2-niovecs must not be NULL");
1115 rxi_FreePacketNoLock(RX_CBUF_TO_PACKET(iov->iov_base, p));
1118 rxi_PacketsUnWait();
1120 MUTEX_EXIT(&rx_freePktQ_lock);
1125 #endif /* RX_ENABLE_TSFPQ */
1127 /* Free the packet p. P is assumed not to be on any queue, i.e.
1128 * remove it yourself first if you call this routine. */
1129 #ifdef RX_ENABLE_TSFPQ
1131 rxi_FreePacket(struct rx_packet *p)
1133 rxi_FreeDataBufsTSFPQ(p, 2, 0);
1134 rxi_FreePacketTSFPQ(p, RX_TS_FPQ_FLUSH_GLOBAL);
1136 #else /* RX_ENABLE_TSFPQ */
1138 rxi_FreePacket(struct rx_packet *p)
1143 MUTEX_ENTER(&rx_freePktQ_lock);
1145 rxi_FreeDataBufsNoLock(p, 2);
1146 rxi_FreePacketNoLock(p);
1147 /* Wakeup anyone waiting for packets */
1148 rxi_PacketsUnWait();
1150 MUTEX_EXIT(&rx_freePktQ_lock);
1153 #endif /* RX_ENABLE_TSFPQ */
1155 /* rxi_AllocPacket sets up p->length so it reflects the number of
1156 * bytes in the packet at this point, **not including** the header.
1157 * The header is absolutely necessary, besides, this is the way the
1158 * length field is usually used */
1159 #ifdef RX_ENABLE_TSFPQ
1160 static struct rx_packet *
1161 rxi_AllocPacketNoLock(int class)
1163 struct rx_packet *p;
1164 struct rx_ts_info_t * rx_ts_info;
1166 RX_TS_INFO_GET(rx_ts_info);
1168 if (rx_stats_active)
1169 rx_atomic_inc(&rx_stats.packetRequests);
1170 if (opr_queue_IsEmpty(&rx_ts_info->_FPQ.queue)) {
1173 if (opr_queue_IsEmpty(&rx_freePacketQueue))
1174 osi_Panic("rxi_AllocPacket error");
1176 if (opr_queue_IsEmpty(&rx_freePacketQueue))
1177 rxi_MorePacketsNoLock(rx_maxSendWindow);
1181 RX_TS_FPQ_GTOL(rx_ts_info);
1184 RX_TS_FPQ_CHECKOUT(rx_ts_info,p);
1186 dpf(("Alloc %p, class %d\n", p, class));
1189 /* have to do this here because rx_FlushWrite fiddles with the iovs in
1190 * order to truncate outbound packets. In the near future, may need
1191 * to allocate bufs from a static pool here, and/or in AllocSendPacket
1193 RX_PACKET_IOV_FULLINIT(p);
1196 #else /* RX_ENABLE_TSFPQ */
1197 static struct rx_packet *
1198 rxi_AllocPacketNoLock(int class)
1200 struct rx_packet *p;
1203 if (rxi_OverQuota(class)) {
1204 rxi_NeedMorePackets = TRUE;
1205 if (rx_stats_active) {
1207 case RX_PACKET_CLASS_RECEIVE:
1208 rx_atomic_inc(&rx_stats.receivePktAllocFailures);
1210 case RX_PACKET_CLASS_SEND:
1211 rx_atomic_inc(&rx_stats.sendPktAllocFailures);
1213 case RX_PACKET_CLASS_SPECIAL:
1214 rx_atomic_inc(&rx_stats.specialPktAllocFailures);
1216 case RX_PACKET_CLASS_RECV_CBUF:
1217 rx_atomic_inc(&rx_stats.receiveCbufPktAllocFailures);
1219 case RX_PACKET_CLASS_SEND_CBUF:
1220 rx_atomic_inc(&rx_stats.sendCbufPktAllocFailures);
1224 return (struct rx_packet *)0;
1228 if (rx_stats_active)
1229 rx_atomic_inc(&rx_stats.packetRequests);
1232 if (opr_queue_IsEmpty(&rx_freePacketQueue))
1233 osi_Panic("rxi_AllocPacket error");
1235 if (opr_queue_IsEmpty(&rx_freePacketQueue))
1236 rxi_MorePacketsNoLock(rx_maxSendWindow);
1240 p = opr_queue_First(&rx_freePacketQueue, struct rx_packet, entry);
1241 opr_queue_Remove(&p->entry);
1242 RX_FPQ_MARK_USED(p);
1244 dpf(("Alloc %p, class %d\n", p, class));
1247 /* have to do this here because rx_FlushWrite fiddles with the iovs in
1248 * order to truncate outbound packets. In the near future, may need
1249 * to allocate bufs from a static pool here, and/or in AllocSendPacket
1251 RX_PACKET_IOV_FULLINIT(p);
1254 #endif /* RX_ENABLE_TSFPQ */
1256 #ifdef RX_ENABLE_TSFPQ
1257 static struct rx_packet *
1258 rxi_AllocPacketTSFPQ(int class, int pull_global)
1260 struct rx_packet *p;
1261 struct rx_ts_info_t * rx_ts_info;
1263 RX_TS_INFO_GET(rx_ts_info);
1265 if (rx_stats_active)
1266 rx_atomic_inc(&rx_stats.packetRequests);
1267 if (pull_global && opr_queue_IsEmpty(&rx_ts_info->_FPQ.queue)) {
1268 MUTEX_ENTER(&rx_freePktQ_lock);
1270 if (opr_queue_IsEmpty(&rx_freePacketQueue))
1271 rxi_MorePacketsNoLock(rx_maxSendWindow);
1273 RX_TS_FPQ_GTOL(rx_ts_info);
1275 MUTEX_EXIT(&rx_freePktQ_lock);
1276 } else if (opr_queue_IsEmpty(&rx_ts_info->_FPQ.queue)) {
1280 RX_TS_FPQ_CHECKOUT(rx_ts_info,p);
1282 dpf(("Alloc %p, class %d\n", p, class));
1284 /* have to do this here because rx_FlushWrite fiddles with the iovs in
1285 * order to truncate outbound packets. In the near future, may need
1286 * to allocate bufs from a static pool here, and/or in AllocSendPacket
1288 RX_PACKET_IOV_FULLINIT(p);
1291 #endif /* RX_ENABLE_TSFPQ */
1293 #ifdef RX_ENABLE_TSFPQ
1295 rxi_AllocPacket(int class)
1297 struct rx_packet *p;
1299 p = rxi_AllocPacketTSFPQ(class, RX_TS_FPQ_PULL_GLOBAL);
1302 #else /* RX_ENABLE_TSFPQ */
1304 rxi_AllocPacket(int class)
1306 struct rx_packet *p;
1308 MUTEX_ENTER(&rx_freePktQ_lock);
1309 p = rxi_AllocPacketNoLock(class);
1310 MUTEX_EXIT(&rx_freePktQ_lock);
1313 #endif /* RX_ENABLE_TSFPQ */
1315 /* This guy comes up with as many buffers as it {takes,can get} given
1316 * the MTU for this call. It also sets the packet length before
1317 * returning. caution: this is often called at NETPRI
1318 * Called with call locked.
1321 rxi_AllocSendPacket(struct rx_call *call, int want)
1323 struct rx_packet *p = (struct rx_packet *)0;
1328 mud = call->MTU - RX_HEADER_SIZE;
1330 rx_GetSecurityHeaderSize(rx_ConnectionOf(call)) +
1331 rx_GetSecurityMaxTrailerSize(rx_ConnectionOf(call));
1333 #ifdef RX_ENABLE_TSFPQ
1334 if ((p = rxi_AllocPacketTSFPQ(RX_PACKET_CLASS_SEND, 0))) {
1336 want = MIN(want, mud);
1338 if ((unsigned)want > p->length)
1339 (void)rxi_AllocDataBuf(p, (want - p->length),
1340 RX_PACKET_CLASS_SEND_CBUF);
1342 if (p->length > mud)
1345 if (delta >= p->length) {
1353 #endif /* RX_ENABLE_TSFPQ */
1355 while (!(call->error)) {
1356 MUTEX_ENTER(&rx_freePktQ_lock);
1357 /* if an error occurred, or we get the packet we want, we're done */
1358 if ((p = rxi_AllocPacketNoLock(RX_PACKET_CLASS_SEND))) {
1359 MUTEX_EXIT(&rx_freePktQ_lock);
1362 want = MIN(want, mud);
1364 if ((unsigned)want > p->length)
1365 (void)rxi_AllocDataBuf(p, (want - p->length),
1366 RX_PACKET_CLASS_SEND_CBUF);
1368 if (p->length > mud)
1371 if (delta >= p->length) {
1380 /* no error occurred, and we didn't get a packet, so we sleep.
1381 * At this point, we assume that packets will be returned
1382 * sooner or later, as packets are acknowledged, and so we
1385 call->flags |= RX_CALL_WAIT_PACKETS;
1386 CALL_HOLD(call, RX_CALL_REFCOUNT_PACKET);
1387 MUTEX_EXIT(&call->lock);
1388 rx_waitingForPackets = 1;
1390 #ifdef RX_ENABLE_LOCKS
1391 CV_WAIT(&rx_waitingForPackets_cv, &rx_freePktQ_lock);
1393 osi_rxSleep(&rx_waitingForPackets);
1395 MUTEX_EXIT(&rx_freePktQ_lock);
1396 MUTEX_ENTER(&call->lock);
1397 CALL_RELE(call, RX_CALL_REFCOUNT_PACKET);
1398 call->flags &= ~RX_CALL_WAIT_PACKETS;
1407 /* Windows does not use file descriptors. */
1408 #define CountFDs(amax) 0
1410 /* count the number of used FDs */
1419 for (i = 0; i < amax; i++) {
1420 code = fstat(i, &tstat);
1426 #endif /* AFS_NT40_ENV */
1429 #define CountFDs(amax) amax
1433 #if !defined(KERNEL) || defined(UKERNEL)
1435 /* This function reads a single packet from the interface into the
1436 * supplied packet buffer (*p). Return 0 if the packet is bogus. The
1437 * (host,port) of the sender are stored in the supplied variables, and
1438 * the data length of the packet is stored in the packet structure.
1439 * The header is decoded. */
1441 rxi_ReadPacket(osi_socket socket, struct rx_packet *p, afs_uint32 * host,
1444 struct sockaddr_in from;
1447 afs_uint32 tlen, savelen;
1449 rx_computelen(p, tlen);
1450 rx_SetDataSize(p, tlen); /* this is the size of the user data area */
1452 tlen += RX_HEADER_SIZE; /* now this is the size of the entire packet */
1453 rlen = rx_maxJumboRecvSize; /* this is what I am advertising. Only check
1454 * it once in order to avoid races. */
1457 tlen = rxi_AllocDataBuf(p, tlen, RX_PACKET_CLASS_SEND_CBUF);
1465 /* Extend the last iovec for padding, it's just to make sure that the
1466 * read doesn't return more data than we expect, and is done to get around
1467 * our problems caused by the lack of a length field in the rx header.
1468 * Use the extra buffer that follows the localdata in each packet
1470 savelen = p->wirevec[p->niovecs - 1].iov_len;
1471 p->wirevec[p->niovecs - 1].iov_len += RX_EXTRABUFFERSIZE;
1473 memset(&msg, 0, sizeof(msg));
1474 msg.msg_name = (char *)&from;
1475 msg.msg_namelen = sizeof(struct sockaddr_in);
1476 msg.msg_iov = p->wirevec;
1477 msg.msg_iovlen = p->niovecs;
1478 nbytes = rxi_Recvmsg(socket, &msg, 0);
1480 /* restore the vec to its correct state */
1481 p->wirevec[p->niovecs - 1].iov_len = savelen;
1483 p->length = (u_short)(nbytes - RX_HEADER_SIZE);
1484 if (nbytes < 0 || (nbytes > tlen) || (p->length & 0x8000)) { /* Bogus packet */
1485 if (nbytes < 0 && errno == EWOULDBLOCK) {
1486 if (rx_stats_active)
1487 rx_atomic_inc(&rx_stats.noPacketOnRead);
1488 } else if (nbytes <= 0) {
1489 if (rx_stats_active) {
1490 rx_atomic_inc(&rx_stats.bogusPacketOnRead);
1491 rx_stats.bogusHost = from.sin_addr.s_addr;
1493 dpf(("B: bogus packet from [%x,%d] nb=%d\n", ntohl(from.sin_addr.s_addr),
1494 ntohs(from.sin_port), nbytes));
1499 else if ((rx_intentionallyDroppedOnReadPer100 > 0)
1500 && (random() % 100 < rx_intentionallyDroppedOnReadPer100)) {
1501 rxi_DecodePacketHeader(p);
1503 *host = from.sin_addr.s_addr;
1504 *port = from.sin_port;
1506 dpf(("Dropped %d %s: %x.%u.%u.%u.%u.%u.%u flags %d len %d\n",
1507 p->header.serial, rx_packetTypes[p->header.type - 1], ntohl(*host), ntohs(*port), p->header.serial,
1508 p->header.epoch, p->header.cid, p->header.callNumber, p->header.seq, p->header.flags,
1510 #ifdef RX_TRIMDATABUFS
1511 rxi_TrimDataBufs(p, 1);
1517 /* Extract packet header. */
1518 rxi_DecodePacketHeader(p);
1520 *host = from.sin_addr.s_addr;
1521 *port = from.sin_port;
1523 && p->header.type > 0 && p->header.type <= RX_N_PACKET_TYPES) {
1525 rx_atomic_inc(&rx_stats.packetsRead[p->header.type - 1]);
1528 #ifdef RX_TRIMDATABUFS
1529 /* Free any empty packet buffers at the end of this packet */
1530 rxi_TrimDataBufs(p, 1);
1536 #endif /* !KERNEL || UKERNEL */
1538 /* This function splits off the first packet in a jumbo packet.
1539 * As of AFS 3.5, jumbograms contain more than one fixed size
1540 * packet, and the RX_JUMBO_PACKET flag is set in all but the
1541 * last packet header. All packets (except the last) are padded to
1542 * fall on RX_CBUFFERSIZE boundaries.
1543 * HACK: We store the length of the first n-1 packets in the
1544 * last two pad bytes. */
1547 rxi_SplitJumboPacket(struct rx_packet *p)
1549 struct rx_packet *np;
1550 struct rx_jumboHeader *jp;
1556 /* All but the last packet in each jumbogram are RX_JUMBOBUFFERSIZE
1557 * bytes in length. All but the first packet are preceded by
1558 * an abbreviated four byte header. The length of the last packet
1559 * is calculated from the size of the jumbogram. */
1560 length = RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
1562 if ((int)p->length < length) {
1563 dpf(("rxi_SplitJumboPacket: bogus length %d\n", p->length));
1566 niov = p->niovecs - 2;
1568 dpf(("rxi_SplitJumboPacket: bogus niovecs %d\n", p->niovecs));
1571 iov = &p->wirevec[2];
1572 np = RX_CBUF_TO_PACKET(iov->iov_base, p);
1574 /* Get a pointer to the abbreviated packet header */
1575 jp = (struct rx_jumboHeader *)
1576 ((char *)(p->wirevec[1].iov_base) + RX_JUMBOBUFFERSIZE);
1578 /* Set up the iovecs for the next packet */
1579 np->wirevec[0].iov_base = (char *)(&np->wirehead[0]);
1580 np->wirevec[0].iov_len = sizeof(struct rx_header);
1581 np->wirevec[1].iov_base = (char *)(&np->localdata[0]);
1582 np->wirevec[1].iov_len = length - RX_JUMBOHEADERSIZE;
1583 np->niovecs = niov + 1;
1584 for (i = 2, iov++; i <= niov; i++, iov++) {
1585 np->wirevec[i] = *iov;
1587 np->length = p->length - length;
1588 p->length = RX_JUMBOBUFFERSIZE;
1591 /* Convert the jumbo packet header to host byte order */
1592 temp = ntohl(*(afs_uint32 *) jp);
1593 jp->flags = (u_char) (temp >> 24);
1594 jp->cksum = (u_short) (temp);
1596 /* Fill in the packet header */
1597 np->header = p->header;
1598 np->header.serial = p->header.serial + 1;
1599 np->header.seq = p->header.seq + 1;
1600 np->header.userStatus = 0;
1601 np->header.flags = jp->flags;
1602 np->header.spare = jp->cksum;
1608 /* Send a udp datagram */
1610 osi_NetSend(osi_socket socket, void *addr, struct iovec *dvec, int nvecs,
1611 int length, int istack)
1616 memset(&msg, 0, sizeof(msg));
1618 msg.msg_iovlen = nvecs;
1619 msg.msg_name = addr;
1620 msg.msg_namelen = sizeof(struct sockaddr_in);
1622 ret = rxi_Sendmsg(socket, &msg, 0);
1626 #elif !defined(UKERNEL)
1628 * message receipt is done in rxk_input or rx_put.
1631 #if defined(AFS_SUN5_ENV) || defined(AFS_HPUX110_ENV)
1633 * Copy an mblock to the contiguous area pointed to by cp.
1634 * MTUXXX Supposed to skip <off> bytes and copy <len> bytes,
1635 * but it doesn't really.
1636 * Returns the number of bytes not transferred.
1637 * The message is NOT changed.
1640 cpytoc(mblk_t * mp, int off, int len, char *cp)
1644 for (; mp && len > 0; mp = mp->b_cont) {
1645 if (mp->b_datap->db_type != M_DATA) {
1648 n = MIN(len, (mp->b_wptr - mp->b_rptr));
1649 memcpy(cp, (char *)mp->b_rptr, n);
1657 /* MTUXXX Supposed to skip <off> bytes and copy <len> bytes,
1658 * but it doesn't really.
1659 * This sucks, anyway, do it like m_cpy.... below
1662 cpytoiovec(mblk_t * mp, int off, int len, struct iovec *iovs,
1667 for (i = -1, t = 0; i < niovs && mp && len > 0; mp = mp->b_cont) {
1668 if (mp->b_datap->db_type != M_DATA) {
1671 n = MIN(len, (mp->b_wptr - mp->b_rptr));
1677 t = iovs[i].iov_len;
1680 memcpy(iovs[i].iov_base + o, (char *)mp->b_rptr, m);
1690 #define m_cpytoc(a, b, c, d) cpytoc(a, b, c, d)
1691 #define m_cpytoiovec(a, b, c, d, e) cpytoiovec(a, b, c, d, e)
1693 #if !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN80_ENV)
1695 m_cpytoiovec(struct mbuf *m, int off, int len, struct iovec iovs[], int niovs)
1698 unsigned int l1, l2, i, t;
1700 if (m == NULL || off < 0 || len < 0 || iovs == NULL)
1701 osi_Panic("m_cpytoiovec"); /* MTUXXX probably don't need this check */
1704 if (m->m_len <= off) {
1714 p1 = mtod(m, caddr_t) + off;
1715 l1 = m->m_len - off;
1717 p2 = iovs[0].iov_base;
1718 l2 = iovs[0].iov_len;
1721 t = MIN(l1, MIN(l2, (unsigned int)len));
1732 p1 = mtod(m, caddr_t);
1738 p2 = iovs[i].iov_base;
1739 l2 = iovs[i].iov_len;
1747 #endif /* AFS_SUN5_ENV */
1749 #if !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN80_ENV)
1750 #if defined(AFS_NBSD_ENV)
1752 rx_mb_to_packet(struct mbuf *amb, void (*free) (struct mbuf *), int hdr_len, int data_len, struct rx_packet *phandle)
1755 rx_mb_to_packet(amb, free, hdr_len, data_len, phandle)
1756 #if defined(AFS_SUN5_ENV) || defined(AFS_HPUX110_ENV)
1762 struct rx_packet *phandle;
1763 int hdr_len, data_len;
1764 #endif /* AFS_NBSD_ENV */
1769 m_cpytoiovec(amb, hdr_len, data_len, phandle->wirevec,
1776 #endif /*KERNEL && !UKERNEL */
1779 /* send a response to a debug packet */
1782 rxi_ReceiveDebugPacket(struct rx_packet *ap, osi_socket asocket,
1783 afs_uint32 ahost, short aport, int istack)
1785 struct rx_debugIn tin;
1789 * Only respond to client-initiated Rx debug packets,
1790 * and clear the client flag in the response.
1792 if (ap->header.flags & RX_CLIENT_INITIATED) {
1793 ap->header.flags = ap->header.flags & ~RX_CLIENT_INITIATED;
1794 rxi_EncodePacketHeader(ap);
1799 rx_packetread(ap, 0, sizeof(struct rx_debugIn), (char *)&tin);
1800 /* all done with packet, now set length to the truth, so we can
1801 * reuse this packet */
1802 rx_computelen(ap, ap->length);
1804 tin.type = ntohl(tin.type);
1805 tin.index = ntohl(tin.index);
1807 case RX_DEBUGI_GETSTATS:{
1808 struct rx_debugStats tstat;
1810 /* get basic stats */
1811 memset(&tstat, 0, sizeof(tstat)); /* make sure spares are zero */
1812 tstat.version = RX_DEBUGI_VERSION;
1813 #ifndef RX_ENABLE_LOCKS
1814 tstat.waitingForPackets = rx_waitingForPackets;
1816 MUTEX_ENTER(&rx_serverPool_lock);
1817 tstat.nFreePackets = htonl(rx_nFreePackets);
1818 tstat.nPackets = htonl(rx_nPackets);
1819 tstat.callsExecuted = htonl(rxi_nCalls);
1820 tstat.packetReclaims = htonl(rx_packetReclaims);
1821 tstat.usedFDs = CountFDs(64);
1822 tstat.nWaiting = htonl(rx_atomic_read(&rx_nWaiting));
1823 tstat.nWaited = htonl(rx_atomic_read(&rx_nWaited));
1824 tstat.idleThreads = opr_queue_Count(&rx_idleServerQueue);
1825 MUTEX_EXIT(&rx_serverPool_lock);
1826 tstat.idleThreads = htonl(tstat.idleThreads);
1827 tl = sizeof(struct rx_debugStats) - ap->length;
1829 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
1832 rx_packetwrite(ap, 0, sizeof(struct rx_debugStats),
1834 ap->length = sizeof(struct rx_debugStats);
1835 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
1836 rx_computelen(ap, ap->length);
1841 case RX_DEBUGI_GETALLCONN:
1842 case RX_DEBUGI_GETCONN:{
1844 struct rx_connection *tc;
1845 struct rx_call *tcall;
1846 struct rx_debugConn tconn;
1847 int all = (tin.type == RX_DEBUGI_GETALLCONN);
1850 tl = sizeof(struct rx_debugConn) - ap->length;
1852 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
1856 memset(&tconn, 0, sizeof(tconn)); /* make sure spares are zero */
1857 /* get N'th (maybe) "interesting" connection info */
1858 for (i = 0; i < rx_hashTableSize; i++) {
1859 #if !defined(KERNEL)
1860 /* the time complexity of the algorithm used here
1861 * exponentially increses with the number of connections.
1863 #ifdef AFS_PTHREAD_ENV
1869 MUTEX_ENTER(&rx_connHashTable_lock);
1870 /* We might be slightly out of step since we are not
1871 * locking each call, but this is only debugging output.
1873 for (tc = rx_connHashTable[i]; tc; tc = tc->next) {
1874 if ((all || rxi_IsConnInteresting(tc))
1875 && tin.index-- <= 0) {
1876 int do_secstats = 0;
1877 tconn.host = tc->peer->host;
1878 tconn.port = tc->peer->port;
1879 tconn.cid = htonl(tc->cid);
1880 tconn.epoch = htonl(tc->epoch);
1881 tconn.serial = htonl(tc->serial);
1882 for (j = 0; j < RX_MAXCALLS; j++) {
1883 tconn.callNumber[j] = htonl(tc->callNumber[j]);
1884 if ((tcall = tc->call[j])) {
1885 tconn.callState[j] = tcall->state;
1886 tconn.callMode[j] = tcall->app.mode;
1887 tconn.callFlags[j] = tcall->flags;
1888 if (!opr_queue_IsEmpty(&tcall->rq))
1889 tconn.callOther[j] |= RX_OTHER_IN;
1890 if (!opr_queue_IsEmpty(&tcall->tq))
1891 tconn.callOther[j] |= RX_OTHER_OUT;
1893 tconn.callState[j] = RX_STATE_NOTINIT;
1896 tconn.natMTU = htonl(tc->peer->natMTU);
1897 tconn.error = htonl(tc->error);
1898 tconn.flags = (u_char) (tc->flags & 0xff); /* compat. */
1899 tconn.type = tc->type;
1900 tconn.securityIndex = tc->securityIndex;
1901 if (tc->securityObject) {
1903 code = RXS_GetStats(tc->securityObject, tc,
1910 #define DOHTONL(a) (tconn.secStats.a = htonl(tconn.secStats.a))
1911 #define DOHTONS(a) (tconn.secStats.a = htons(tconn.secStats.a))
1914 DOHTONL(packetsReceived);
1915 DOHTONL(packetsSent);
1916 DOHTONL(bytesReceived);
1920 sizeof(tconn.secStats.spares) /
1925 sizeof(tconn.secStats.sparel) /
1926 sizeof(afs_int32); i++)
1929 memset(&tconn.secStats, 0, sizeof(tconn.secStats));
1932 MUTEX_EXIT(&rx_connHashTable_lock);
1933 rx_packetwrite(ap, 0, sizeof(struct rx_debugConn),
1936 ap->length = sizeof(struct rx_debugConn);
1937 rxi_SendDebugPacket(ap, asocket, ahost, aport,
1943 MUTEX_EXIT(&rx_connHashTable_lock);
1945 /* if we make it here, there are no interesting packets */
1946 tconn.cid = htonl(0xffffffff); /* means end */
1947 rx_packetwrite(ap, 0, sizeof(struct rx_debugConn),
1950 ap->length = sizeof(struct rx_debugConn);
1951 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
1957 * Pass back all the peer structures we have available
1960 case RX_DEBUGI_GETPEER:{
1963 struct rx_debugPeer tpeer;
1966 tl = sizeof(struct rx_debugPeer) - ap->length;
1968 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
1972 memset(&tpeer, 0, sizeof(tpeer));
1973 for (i = 0; i < rx_hashTableSize; i++) {
1974 #if !defined(KERNEL)
1975 /* the time complexity of the algorithm used here
1976 * exponentially increses with the number of peers.
1978 * Yielding after processing each hash table entry
1979 * and dropping rx_peerHashTable_lock.
1980 * also increases the risk that we will miss a new
1981 * entry - but we are willing to live with this
1982 * limitation since this is meant for debugging only
1984 #ifdef AFS_PTHREAD_ENV
1990 MUTEX_ENTER(&rx_peerHashTable_lock);
1991 for (tp = rx_peerHashTable[i]; tp; tp = tp->next) {
1992 if (tin.index-- <= 0) {
1994 MUTEX_EXIT(&rx_peerHashTable_lock);
1996 MUTEX_ENTER(&tp->peer_lock);
1997 tpeer.host = tp->host;
1998 tpeer.port = tp->port;
1999 tpeer.ifMTU = htons(tp->ifMTU);
2000 tpeer.idleWhen = htonl(tp->idleWhen);
2001 tpeer.refCount = htons(tp->refCount);
2002 tpeer.burstSize = 0;
2004 tpeer.burstWait.sec = 0;
2005 tpeer.burstWait.usec = 0;
2006 tpeer.rtt = htonl(tp->rtt);
2007 tpeer.rtt_dev = htonl(tp->rtt_dev);
2008 tpeer.nSent = htonl(tp->nSent);
2009 tpeer.reSends = htonl(tp->reSends);
2010 tpeer.natMTU = htons(tp->natMTU);
2011 tpeer.maxMTU = htons(tp->maxMTU);
2012 tpeer.maxDgramPackets = htons(tp->maxDgramPackets);
2013 tpeer.ifDgramPackets = htons(tp->ifDgramPackets);
2014 tpeer.MTU = htons(tp->MTU);
2015 tpeer.cwind = htons(tp->cwind);
2016 tpeer.nDgramPackets = htons(tp->nDgramPackets);
2017 tpeer.congestSeq = htons(tp->congestSeq);
2018 tpeer.bytesSent.high =
2019 htonl(tp->bytesSent >> 32);
2020 tpeer.bytesSent.low =
2021 htonl(tp->bytesSent & MAX_AFS_UINT32);
2022 tpeer.bytesReceived.high =
2023 htonl(tp->bytesReceived >> 32);
2024 tpeer.bytesReceived.low =
2025 htonl(tp->bytesReceived & MAX_AFS_UINT32);
2026 MUTEX_EXIT(&tp->peer_lock);
2028 MUTEX_ENTER(&rx_peerHashTable_lock);
2030 MUTEX_EXIT(&rx_peerHashTable_lock);
2032 rx_packetwrite(ap, 0, sizeof(struct rx_debugPeer),
2035 ap->length = sizeof(struct rx_debugPeer);
2036 rxi_SendDebugPacket(ap, asocket, ahost, aport,
2042 MUTEX_EXIT(&rx_peerHashTable_lock);
2044 /* if we make it here, there are no interesting packets */
2045 tpeer.host = htonl(0xffffffff); /* means end */
2046 rx_packetwrite(ap, 0, sizeof(struct rx_debugPeer),
2049 ap->length = sizeof(struct rx_debugPeer);
2050 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2055 case RX_DEBUGI_RXSTATS:{
2059 tl = sizeof(rx_stats) - ap->length;
2061 tl = rxi_AllocDataBuf(ap, tl, RX_PACKET_CLASS_SEND_CBUF);
2065 /* Since its all int32s convert to network order with a loop. */
2066 if (rx_stats_active)
2067 MUTEX_ENTER(&rx_stats_mutex);
2068 s = (afs_int32 *) & rx_stats;
2069 for (i = 0; i < sizeof(rx_stats) / sizeof(afs_int32); i++, s++)
2070 rx_PutInt32(ap, i * sizeof(afs_int32), htonl(*s));
2073 ap->length = sizeof(rx_stats);
2074 if (rx_stats_active)
2075 MUTEX_EXIT(&rx_stats_mutex);
2076 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2082 /* error response packet */
2083 tin.type = htonl(RX_DEBUGI_BADTYPE);
2084 tin.index = tin.type;
2085 rx_packetwrite(ap, 0, sizeof(struct rx_debugIn), (char *)&tin);
2087 ap->length = sizeof(struct rx_debugIn);
2088 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2096 rxi_ReceiveVersionPacket(struct rx_packet *ap, osi_socket asocket,
2097 afs_uint32 ahost, short aport, int istack)
2102 * Only respond to client-initiated version requests, and
2103 * clear that flag in the response.
2105 if (ap->header.flags & RX_CLIENT_INITIATED) {
2108 ap->header.flags = ap->header.flags & ~RX_CLIENT_INITIATED;
2109 rxi_EncodePacketHeader(ap);
2110 memset(buf, 0, sizeof(buf));
2111 strncpy(buf, cml_version_number + 4, sizeof(buf) - 1);
2112 rx_packetwrite(ap, 0, 65, buf);
2115 rxi_SendDebugPacket(ap, asocket, ahost, aport, istack);
2123 /* send a debug packet back to the sender */
2125 rxi_SendDebugPacket(struct rx_packet *apacket, osi_socket asocket,
2126 afs_uint32 ahost, short aport, afs_int32 istack)
2128 struct sockaddr_in taddr;
2129 unsigned int i, nbytes, savelen = 0;
2132 int waslocked = ISAFS_GLOCK();
2135 taddr.sin_family = AF_INET;
2136 taddr.sin_port = aport;
2137 taddr.sin_addr.s_addr = ahost;
2138 memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
2139 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2140 taddr.sin_len = sizeof(struct sockaddr_in);
2143 /* We need to trim the niovecs. */
2144 nbytes = apacket->length;
2145 for (i = 1; i < apacket->niovecs; i++) {
2146 if (nbytes <= apacket->wirevec[i].iov_len) {
2147 savelen = apacket->wirevec[i].iov_len;
2148 saven = apacket->niovecs;
2149 apacket->wirevec[i].iov_len = nbytes;
2150 apacket->niovecs = i + 1; /* so condition fails because i == niovecs */
2152 nbytes -= apacket->wirevec[i].iov_len;
2155 #ifdef RX_KERNEL_TRACE
2156 if (ICL_SETACTIVE(afs_iclSetp)) {
2159 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2160 "before rxi_NetSend()");
2168 /* debug packets are not reliably delivered, hence the cast below. */
2169 (void)rxi_NetSend(asocket, &taddr, apacket->wirevec, apacket->niovecs,
2170 apacket->length + RX_HEADER_SIZE, istack);
2172 #ifdef RX_KERNEL_TRACE
2173 if (ICL_SETACTIVE(afs_iclSetp)) {
2175 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2176 "after rxi_NetSend()");
2185 if (saven) { /* means we truncated the packet above. */
2186 apacket->wirevec[i - 1].iov_len = savelen;
2187 apacket->niovecs = saven;
2193 rxi_NetSendError(struct rx_call *call, int code)
2197 if (code == -1 && WSAGetLastError() == WSAEHOSTUNREACH) {
2200 if (code == -WSAEHOSTUNREACH) {
2203 #elif defined(AFS_LINUX20_ENV)
2204 if (code == -ENETUNREACH) {
2207 #elif defined(AFS_DARWIN_ENV)
2208 if (code == EHOSTUNREACH) {
2213 call->lastReceiveTime = 0;
2217 /* Send the packet to appropriate destination for the specified
2218 * call. The header is first encoded and placed in the packet.
2221 rxi_SendPacket(struct rx_call *call, struct rx_connection *conn,
2222 struct rx_packet *p, int istack)
2228 struct sockaddr_in addr;
2229 struct rx_peer *peer = conn->peer;
2232 char deliveryType = 'S';
2234 /* The address we're sending the packet to */
2235 memset(&addr, 0, sizeof(addr));
2236 addr.sin_family = AF_INET;
2237 addr.sin_port = peer->port;
2238 addr.sin_addr.s_addr = peer->host;
2239 memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
2241 /* This stuff should be revamped, I think, so that most, if not
2242 * all, of the header stuff is always added here. We could
2243 * probably do away with the encode/decode routines. XXXXX */
2245 /* Stamp each packet with a unique serial number. The serial
2246 * number is maintained on a connection basis because some types
2247 * of security may be based on the serial number of the packet,
2248 * and security is handled on a per authenticated-connection
2250 /* Pre-increment, to guarantee no zero serial number; a zero
2251 * serial number means the packet was never sent. */
2252 MUTEX_ENTER(&conn->conn_data_lock);
2253 p->header.serial = ++conn->serial;
2254 if (p->length > conn->peer->maxPacketSize) {
2255 if ((p->header.type == RX_PACKET_TYPE_ACK) &&
2256 (p->header.flags & RX_REQUEST_ACK)) {
2257 conn->lastPingSize = p->length;
2258 conn->lastPingSizeSer = p->header.serial;
2259 } else if (p->header.seq != 0) {
2260 conn->lastPacketSize = p->length;
2261 conn->lastPacketSizeSeq = p->header.seq;
2264 MUTEX_EXIT(&conn->conn_data_lock);
2265 /* This is so we can adjust retransmit time-outs better in the face of
2266 * rapidly changing round-trip times. RTO estimation is not a la Karn.
2268 if (p->firstSerial == 0) {
2269 p->firstSerial = p->header.serial;
2272 /* If an output tracer function is defined, call it with the packet and
2273 * network address. Note this function may modify its arguments. */
2274 if (rx_almostSent) {
2275 int drop = (*rx_almostSent) (p, &addr);
2276 /* drop packet if return value is non-zero? */
2278 deliveryType = 'D'; /* Drop the packet */
2282 /* Get network byte order header */
2283 rxi_EncodePacketHeader(p); /* XXX in the event of rexmit, etc, don't need to
2284 * touch ALL the fields */
2286 /* Send the packet out on the same socket that related packets are being
2290 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
2293 /* Possibly drop this packet, for testing purposes */
2294 if ((deliveryType == 'D')
2295 || ((rx_intentionallyDroppedPacketsPer100 > 0)
2296 && (random() % 100 < rx_intentionallyDroppedPacketsPer100))) {
2297 deliveryType = 'D'; /* Drop the packet */
2299 deliveryType = 'S'; /* Send the packet */
2300 #endif /* RXDEBUG */
2302 /* Loop until the packet is sent. We'd prefer just to use a
2303 * blocking socket, but unfortunately the interface doesn't
2304 * allow us to have the socket block in send mode, and not
2305 * block in receive mode */
2307 waslocked = ISAFS_GLOCK();
2308 #ifdef RX_KERNEL_TRACE
2309 if (ICL_SETACTIVE(afs_iclSetp)) {
2312 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2313 "before rxi_NetSend()");
2322 rxi_NetSend(socket, &addr, p->wirevec, p->niovecs,
2323 p->length + RX_HEADER_SIZE, istack)) != 0) {
2324 /* send failed, so let's hurry up the resend, eh? */
2325 if (rx_stats_active)
2326 rx_atomic_inc(&rx_stats.netSendFailures);
2327 p->flags &= ~RX_PKTFLAG_SENT; /* resend it very soon */
2329 /* Some systems are nice and tell us right away that we cannot
2330 * reach this recipient by returning an error code.
2331 * So, when this happens let's "down" the host NOW so
2332 * we don't sit around waiting for this host to timeout later.
2335 rxi_NetSendError(call, code);
2339 #ifdef RX_KERNEL_TRACE
2340 if (ICL_SETACTIVE(afs_iclSetp)) {
2342 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
2343 "after rxi_NetSend()");
2354 dpf(("%c %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %p len %d\n",
2355 deliveryType, p->header.serial, rx_packetTypes[p->header.type - 1], ntohl(peer->host),
2356 ntohs(peer->port), p->header.serial, p->header.epoch, p->header.cid, p->header.callNumber,
2357 p->header.seq, p->header.flags, p, p->length));
2359 if (rx_stats_active) {
2360 rx_atomic_inc(&rx_stats.packetsSent[p->header.type - 1]);
2361 MUTEX_ENTER(&peer->peer_lock);
2362 peer->bytesSent += p->length;
2363 MUTEX_EXIT(&peer->peer_lock);
2367 /* Send a list of packets to appropriate destination for the specified
2368 * connection. The headers are first encoded and placed in the packets.
2371 rxi_SendPacketList(struct rx_call *call, struct rx_connection *conn,
2372 struct rx_packet **list, int len, int istack)
2374 #if defined(AFS_SUN5_ENV) && defined(KERNEL)
2377 struct sockaddr_in addr;
2378 struct rx_peer *peer = conn->peer;
2380 struct rx_packet *p = NULL;
2381 struct iovec wirevec[RX_MAXIOVECS];
2382 int i, length, code;
2385 struct rx_jumboHeader *jp;
2387 char deliveryType = 'S';
2389 /* The address we're sending the packet to */
2390 addr.sin_family = AF_INET;
2391 addr.sin_port = peer->port;
2392 addr.sin_addr.s_addr = peer->host;
2393 memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
2395 if (len + 1 > RX_MAXIOVECS) {
2396 osi_Panic("rxi_SendPacketList, len > RX_MAXIOVECS\n");
2400 * Stamp the packets in this jumbogram with consecutive serial numbers
2402 MUTEX_ENTER(&conn->conn_data_lock);
2403 serial = conn->serial;
2404 conn->serial += len;
2405 for (i = 0; i < len; i++) {
2407 /* a ping *or* a sequenced packet can count */
2408 if (p->length > conn->peer->maxPacketSize) {
2409 if (((p->header.type == RX_PACKET_TYPE_ACK) &&
2410 (p->header.flags & RX_REQUEST_ACK)) &&
2411 ((i == 0) || (p->length >= conn->lastPingSize))) {
2412 conn->lastPingSize = p->length;
2413 conn->lastPingSizeSer = serial + i;
2414 } else if ((p->header.seq != 0) &&
2415 ((i == 0) || (p->length >= conn->lastPacketSize))) {
2416 conn->lastPacketSize = p->length;
2417 conn->lastPacketSizeSeq = p->header.seq;
2421 MUTEX_EXIT(&conn->conn_data_lock);
2424 /* This stuff should be revamped, I think, so that most, if not
2425 * all, of the header stuff is always added here. We could
2426 * probably do away with the encode/decode routines. XXXXX */
2429 length = RX_HEADER_SIZE;
2430 wirevec[0].iov_base = (char *)(&list[0]->wirehead[0]);
2431 wirevec[0].iov_len = RX_HEADER_SIZE;
2432 for (i = 0; i < len; i++) {
2435 /* The whole 3.5 jumbogram scheme relies on packets fitting
2436 * in a single packet buffer. */
2437 if (p->niovecs > 2) {
2438 osi_Panic("rxi_SendPacketList, niovecs > 2\n");
2441 /* Set the RX_JUMBO_PACKET flags in all but the last packets
2444 if (p->length != RX_JUMBOBUFFERSIZE) {
2445 osi_Panic("rxi_SendPacketList, length != jumbo size\n");
2447 p->header.flags |= RX_JUMBO_PACKET;
2448 length += RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
2449 wirevec[i + 1].iov_len = RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
2451 wirevec[i + 1].iov_len = p->length;
2452 length += p->length;
2454 wirevec[i + 1].iov_base = (char *)(&p->localdata[0]);
2456 /* Convert jumbo packet header to network byte order */
2457 temp = (afs_uint32) (p->header.flags) << 24;
2458 temp |= (afs_uint32) (p->header.spare);
2459 *(afs_uint32 *) jp = htonl(temp);
2461 jp = (struct rx_jumboHeader *)
2462 ((char *)(&p->localdata[0]) + RX_JUMBOBUFFERSIZE);
2464 /* Stamp each packet with a unique serial number. The serial
2465 * number is maintained on a connection basis because some types
2466 * of security may be based on the serial number of the packet,
2467 * and security is handled on a per authenticated-connection
2469 /* Pre-increment, to guarantee no zero serial number; a zero
2470 * serial number means the packet was never sent. */
2471 p->header.serial = ++serial;
2472 /* This is so we can adjust retransmit time-outs better in the face of
2473 * rapidly changing round-trip times. RTO estimation is not a la Karn.
2475 if (p->firstSerial == 0) {
2476 p->firstSerial = p->header.serial;
2479 /* If an output tracer function is defined, call it with the packet and
2480 * network address. Note this function may modify its arguments. */
2481 if (rx_almostSent) {
2482 int drop = (*rx_almostSent) (p, &addr);
2483 /* drop packet if return value is non-zero? */
2485 deliveryType = 'D'; /* Drop the packet */
2489 /* Get network byte order header */
2490 rxi_EncodePacketHeader(p); /* XXX in the event of rexmit, etc, don't need to
2491 * touch ALL the fields */
2494 /* Send the packet out on the same socket that related packets are being
2498 RX_CLIENT_CONNECTION ? rx_socket : conn->service->socket);
2501 /* Possibly drop this packet, for testing purposes */
2502 if ((deliveryType == 'D')
2503 || ((rx_intentionallyDroppedPacketsPer100 > 0)
2504 && (random() % 100 < rx_intentionallyDroppedPacketsPer100))) {
2505 deliveryType = 'D'; /* Drop the packet */
2507 deliveryType = 'S'; /* Send the packet */
2508 #endif /* RXDEBUG */
2510 /* Loop until the packet is sent. We'd prefer just to use a
2511 * blocking socket, but unfortunately the interface doesn't
2512 * allow us to have the socket block in send mode, and not
2513 * block in receive mode */
2514 #if defined(AFS_SUN5_ENV) && defined(KERNEL)
2515 waslocked = ISAFS_GLOCK();
2516 if (!istack && waslocked)
2520 rxi_NetSend(socket, &addr, &wirevec[0], len + 1, length,
2522 /* send failed, so let's hurry up the resend, eh? */
2523 if (rx_stats_active)
2524 rx_atomic_inc(&rx_stats.netSendFailures);
2525 for (i = 0; i < len; i++) {
2527 p->flags &= ~RX_PKTFLAG_SENT; /* resend it very soon */
2529 /* Some systems are nice and tell us right away that we cannot
2530 * reach this recipient by returning an error code.
2531 * So, when this happens let's "down" the host NOW so
2532 * we don't sit around waiting for this host to timeout later.
2535 rxi_NetSendError(call, code);
2538 #if defined(AFS_SUN5_ENV) && defined(KERNEL)
2539 if (!istack && waslocked)
2545 osi_Assert(p != NULL);
2547 dpf(("%c %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %p len %d\n",
2548 deliveryType, p->header.serial, rx_packetTypes[p->header.type - 1], ntohl(peer->host),
2549 ntohs(peer->port), p->header.serial, p->header.epoch, p->header.cid, p->header.callNumber,
2550 p->header.seq, p->header.flags, p, p->length));
2553 if (rx_stats_active) {
2554 rx_atomic_inc(&rx_stats.packetsSent[p->header.type - 1]);
2555 MUTEX_ENTER(&peer->peer_lock);
2556 peer->bytesSent += p->length;
2557 MUTEX_EXIT(&peer->peer_lock);
2561 /* Send a raw abort packet, without any call or connection structures */
2563 rxi_SendRawAbort(osi_socket socket, afs_uint32 host, u_short port,
2564 afs_uint32 serial, afs_int32 error,
2565 struct rx_packet *source, int istack)
2567 struct rx_header theader;
2568 struct sockaddr_in addr;
2569 struct iovec iov[2];
2571 memset(&theader, 0, sizeof(theader));
2572 theader.epoch = htonl(source->header.epoch);
2573 theader.callNumber = htonl(source->header.callNumber);
2574 theader.serial = htonl(serial);
2575 theader.type = RX_PACKET_TYPE_ABORT;
2576 theader.serviceId = htons(source->header.serviceId);
2577 theader.securityIndex = source->header.securityIndex;
2578 theader.cid = htonl(source->header.cid);
2581 * If the abort is being sent in response to a server initiated packet,
2582 * set client_initiated in the abort to ensure it is not associated by
2583 * the receiver with a connection in the opposite direction.
2585 if ((source->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2586 theader.flags |= RX_CLIENT_INITIATED;
2588 error = htonl(error);
2590 iov[0].iov_base = &theader;
2591 iov[0].iov_len = sizeof(struct rx_header);
2592 iov[1].iov_base = &error;
2593 iov[1].iov_len = sizeof(error);
2595 addr.sin_family = AF_INET;
2596 addr.sin_addr.s_addr = host;
2597 addr.sin_port = port;
2598 memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
2599 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2600 addr.sin_len = sizeof(struct sockaddr_in);
2603 rxi_NetSend(socket, &addr, iov, 2,
2604 sizeof(struct rx_header) + sizeof(error), istack);
2607 /* Send a "special" packet to the peer connection. If call is
2608 * specified, then the packet is directed to a specific call channel
2609 * associated with the connection, otherwise it is directed to the
2610 * connection only. Uses optionalPacket if it is supplied, rather than
2611 * allocating a new packet buffer. Nbytes is the length of the data
2612 * portion of the packet. If data is non-null, nbytes of data are
2613 * copied into the packet. Type is the type of the packet, as defined
2614 * in rx.h. Bug: there's a lot of duplication between this and other
2615 * routines. This needs to be cleaned up. */
2617 rxi_SendSpecial(struct rx_call *call,
2618 struct rx_connection *conn,
2619 struct rx_packet *optionalPacket, int type, char *data,
2620 int nbytes, int istack)
2622 /* Some of the following stuff should be common code for all
2623 * packet sends (it's repeated elsewhere) */
2624 struct rx_packet *p;
2626 int savelen = 0, saven = 0;
2627 int channel, callNumber;
2629 channel = call->channel;
2630 callNumber = *call->callNumber;
2631 /* BUSY packets refer to the next call on this connection */
2632 if (type == RX_PACKET_TYPE_BUSY) {
2641 p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
2643 osi_Panic("rxi_SendSpecial failure");
2650 p->header.serviceId = conn->serviceId;
2651 p->header.securityIndex = conn->securityIndex;
2652 p->header.cid = (conn->cid | channel);
2653 p->header.callNumber = callNumber;
2655 p->header.epoch = conn->epoch;
2656 p->header.type = type;
2657 p->header.userStatus = 0;
2658 p->header.flags = 0;
2659 if (conn->type == RX_CLIENT_CONNECTION)
2660 p->header.flags |= RX_CLIENT_INITIATED;
2662 rx_packetwrite(p, 0, nbytes, data);
2664 for (i = 1; i < p->niovecs; i++) {
2665 if (nbytes <= p->wirevec[i].iov_len) {
2666 savelen = p->wirevec[i].iov_len;
2668 p->wirevec[i].iov_len = nbytes;
2669 p->niovecs = i + 1; /* so condition fails because i == niovecs */
2671 nbytes -= p->wirevec[i].iov_len;
2675 rxi_Send(call, p, istack);
2677 rxi_SendPacket((struct rx_call *)0, conn, p, istack);
2678 if (saven) { /* means we truncated the packet above. We probably don't */
2679 /* really need to do this, but it seems safer this way, given that */
2680 /* sneaky optionalPacket... */
2681 p->wirevec[i - 1].iov_len = savelen;
2684 if (!optionalPacket)
2686 return optionalPacket;
2690 /* Encode the packet's header (from the struct header in the packet to
2691 * the net byte order representation in the wire representation of the
2692 * packet, which is what is actually sent out on the wire) */
2694 rxi_EncodePacketHeader(struct rx_packet *p)
2696 afs_uint32 *buf = (afs_uint32 *) (p->wirevec[0].iov_base); /* MTUXXX */
2698 memset(buf, 0, RX_HEADER_SIZE);
2699 *buf++ = htonl(p->header.epoch);
2700 *buf++ = htonl(p->header.cid);
2701 *buf++ = htonl(p->header.callNumber);
2702 *buf++ = htonl(p->header.seq);
2703 *buf++ = htonl(p->header.serial);
2704 *buf++ = htonl((((afs_uint32) p->header.type) << 24)
2705 | (((afs_uint32) p->header.flags) << 16)
2706 | (p->header.userStatus << 8) | p->header.securityIndex);
2707 /* Note: top 16 bits of this next word were reserved */
2708 *buf++ = htonl((p->header.spare << 16) | (p->header.serviceId & 0xffff));
2711 /* Decode the packet's header (from net byte order to a struct header) */
2713 rxi_DecodePacketHeader(struct rx_packet *p)
2715 afs_uint32 *buf = (afs_uint32 *) (p->wirevec[0].iov_base); /* MTUXXX */
2718 p->header.epoch = ntohl(*buf);
2720 p->header.cid = ntohl(*buf);
2722 p->header.callNumber = ntohl(*buf);
2724 p->header.seq = ntohl(*buf);
2726 p->header.serial = ntohl(*buf);
2732 /* C will truncate byte fields to bytes for me */
2733 p->header.type = temp >> 24;
2734 p->header.flags = temp >> 16;
2735 p->header.userStatus = temp >> 8;
2736 p->header.securityIndex = temp >> 0;
2741 p->header.serviceId = (temp & 0xffff);
2742 p->header.spare = temp >> 16;
2743 /* Note: top 16 bits of this last word are the security checksum */
2747 * LOCKS HELD: called with call->lock held.
2749 * PrepareSendPacket is the only place in the code that
2750 * can increment call->tnext. This could become an atomic
2751 * in the future. Beyond that there is nothing in this
2752 * function that requires the call being locked. This
2753 * function can only be called by the application thread.
2756 rxi_PrepareSendPacket(struct rx_call *call,
2757 struct rx_packet *p, int last)
2759 struct rx_connection *conn = call->conn;
2760 afs_uint32 seq = call->tnext++;
2762 afs_int32 len; /* len must be a signed type; it can go negative */
2765 /* No data packets on call 0. Where do these come from? */
2766 if (*call->callNumber == 0)
2767 *call->callNumber = 1;
2769 MUTEX_EXIT(&call->lock);
2770 p->flags &= ~(RX_PKTFLAG_ACKED | RX_PKTFLAG_SENT);
2772 p->header.cid = (conn->cid | call->channel);
2773 p->header.serviceId = conn->serviceId;
2774 p->header.securityIndex = conn->securityIndex;
2776 p->header.callNumber = *call->callNumber;
2777 p->header.seq = seq;
2778 p->header.epoch = conn->epoch;
2779 p->header.type = RX_PACKET_TYPE_DATA;
2780 p->header.userStatus = 0;
2781 p->header.flags = 0;
2782 p->header.spare = 0;
2783 if (conn->type == RX_CLIENT_CONNECTION)
2784 p->header.flags |= RX_CLIENT_INITIATED;
2787 p->header.flags |= RX_LAST_PACKET;
2789 clock_Zero(&p->firstSent); /* Never yet transmitted */
2790 p->header.serial = 0; /* Another way of saying never transmitted... */
2792 /* Now that we're sure this is the last data on the call, make sure
2793 * that the "length" and the sum of the iov_lens matches. */
2794 len = p->length + call->conn->securityHeaderSize;
2796 for (i = 1; i < p->niovecs && len > 0; i++) {
2797 len -= p->wirevec[i].iov_len;
2800 osi_Panic("PrepareSendPacket 1\n"); /* MTUXXX */
2801 } else if (i < p->niovecs) {
2802 /* Free any extra elements in the wirevec */
2803 #if defined(RX_ENABLE_TSFPQ)
2804 rxi_FreeDataBufsTSFPQ(p, i, 1 /* allow global pool flush if overquota */);
2805 #else /* !RX_ENABLE_TSFPQ */
2806 MUTEX_ENTER(&rx_freePktQ_lock);
2807 rxi_FreeDataBufsNoLock(p, i);
2808 MUTEX_EXIT(&rx_freePktQ_lock);
2809 #endif /* !RX_ENABLE_TSFPQ */
2814 p->wirevec[i - 1].iov_len += len;
2815 MUTEX_ENTER(&call->lock);
2816 code = RXS_PreparePacket(conn->securityObject, call, p);
2818 MUTEX_EXIT(&call->lock);
2819 rxi_ConnectionError(conn, code);
2820 MUTEX_ENTER(&conn->conn_data_lock);
2821 p = rxi_SendConnectionAbort(conn, p, 0, 0);
2822 MUTEX_EXIT(&conn->conn_data_lock);
2823 MUTEX_ENTER(&call->lock);
2824 /* setting a connection error means all calls for that conn are also
2825 * error'd. if this call does not have an error by now, something is
2826 * very wrong, and we risk sending data in the clear that is supposed
2827 * to be encrypted. */
2828 osi_Assert(call->error);
2832 /* Given an interface MTU size, calculate an adjusted MTU size that
2833 * will make efficient use of the RX buffers when the peer is sending
2834 * either AFS 3.4a jumbograms or AFS 3.5 jumbograms. */
2836 rxi_AdjustIfMTU(int mtu)
2841 if (rxi_nRecvFrags == 1 && rxi_nSendFrags == 1)
2843 adjMTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE;
2844 if (mtu <= adjMTU) {
2851 frags = mtu / (RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE);
2852 return (adjMTU + (frags * (RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE)));
2855 /* Given an interface MTU size, and the peer's advertised max receive
2856 * size, calculate an adjisted maxMTU size that makes efficient use
2857 * of our packet buffers when we are sending AFS 3.4a jumbograms. */
2859 rxi_AdjustMaxMTU(int mtu, int peerMaxMTU)
2861 int maxMTU = mtu * rxi_nSendFrags;
2862 maxMTU = MIN(maxMTU, peerMaxMTU);
2863 return rxi_AdjustIfMTU(maxMTU);
2866 /* Given a packet size, figure out how many datagram packet will fit.
2867 * The first buffer always contains RX_HEADER_SIZE+RX_JUMBOBUFFERSIZE+
2868 * RX_JUMBOHEADERSIZE, the middle buffers contain RX_JUMBOBUFFERSIZE+
2869 * RX_JUMBOHEADERSIZE, and the last buffer contains RX_JUMBOBUFFERSIZE */
2871 rxi_AdjustDgramPackets(int frags, int mtu)
2874 if (mtu + IPv6_FRAG_HDR_SIZE < RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE) {
2877 maxMTU = (frags * (mtu + UDP_HDR_SIZE)) - UDP_HDR_SIZE;
2878 maxMTU = MIN(maxMTU, RX_MAX_PACKET_SIZE);
2879 /* subtract the size of the first and last packets */
2880 maxMTU -= RX_HEADER_SIZE + (2 * RX_JUMBOBUFFERSIZE) + RX_JUMBOHEADERSIZE;
2884 return (2 + (maxMTU / (RX_JUMBOBUFFERSIZE + RX_JUMBOHEADERSIZE)));
2889 * This function can be used by the Windows Cache Manager
2890 * to dump the list of all rx packets so that we can determine
2891 * where the packet leakage is.
2893 int rx_DumpPackets(FILE *outputFile, char *cookie)
2895 #ifdef RXDEBUG_PACKET
2896 struct rx_packet *p;
2900 #define RXDPRINTF sprintf
2901 #define RXDPRINTOUT output
2903 #define RXDPRINTF fprintf
2904 #define RXDPRINTOUT outputFile
2908 MUTEX_ENTER(&rx_freePktQ_lock);
2909 RXDPRINTF(RXDPRINTOUT, "%s - Start dumping all Rx Packets - count=%u\r\n", cookie, rx_packet_id);
2911 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2914 for (p = rx_mallocedP; p; p = p->allNextp) {
2915 RXDPRINTF(RXDPRINTOUT, "%s - packet=0x%p, id=%u, firstSent=%u.%08u, timeSent=%u.%08u, firstSerial=%u, niovecs=%u, flags=0x%x, 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",
2916 cookie, p, p->packetId, p->firstSent.sec, p->firstSent.usec, p->timeSent.sec, p->timeSent.usec,
2917 p->firstSerial, p->niovecs, (afs_uint32)p->flags, (afs_uint32)p->length,
2918 p->header.epoch, p->header.cid, p->header.callNumber, p->header.seq, p->header.serial,
2919 (afs_uint32)p->header.type, (afs_uint32)p->header.flags, (afs_uint32)p->header.userStatus,
2920 (afs_uint32)p->header.securityIndex, (afs_uint32)p->header.serviceId);
2922 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2926 RXDPRINTF(RXDPRINTOUT, "%s - End dumping all Rx Packets\r\n", cookie);
2928 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2931 MUTEX_EXIT(&rx_freePktQ_lock);
2933 #endif /* RXDEBUG_PACKET */