[U-Boot] [PATCH] Add simple IP/UDP fragmentation support
From: Frank Haverkamp haver@vnet.ibm.com
http://tools.ietf.org/html/rfc2348 describes the TFTP block size option which allows larger packtes than the 512 byte default. This reduces the number of TFTP ACKs significantly and improves performance.
To get the most benefit out of the tftp block size option the support of defragementation of IP/UDP packet is helpful. The current implemenation should work even with packets received out of order. To enable the large packet size the user should set "tftp_block_size" so a value like 16352.
We experimented with different packet sizes and found that more than those 16KiB do not contribute much to the performance anymore. Therefor I limited the defragmentation buffer to 16KiB no too waste memory.
Signed-off-by: Frank Haverkamp haver@vnet.ibm.com Signed-off-by: Josh Boyer jwboyer@linux.vnet.ibm.com
--- include/net.h | 17 ++++++ net/net.c | 156 ++++++++++++++++++++++++++++++++++++++++++++++++++-------- net/tftp.c | 22 ++++++++ net/tftp.h | 10 +++ 4 files changed, 185 insertions(+), 20 deletions(-)
--- u-boot.git.orig/include/net.h +++ u-boot.git/include/net.h @@ -200,6 +200,13 @@ typedef struct { ushort udp_xsum; /* Checksum */ } IP_t;
+#define IP_OFFS 0x1FFF /* ip offset *= 8 */ +#define IP_OFFS_SHIFT 3 /* in 8 byte steps */ +#define IP_FLAGS 0xE000 /* first 3 bits */ +#define IP_FLAGS_RES 0x8000 /* reserved */ +#define IP_FLAGS_DFRAG 0x4000 /* don't fragments */ +#define IP_FLAGS_MFRAG 0x2000 /* more fragments */ + #define IP_HDR_SIZE_NO_UDP (sizeof (IP_t) - 8) #define IP_HDR_SIZE (sizeof (IP_t))
@@ -282,6 +289,16 @@ typedef struct icmphdr { #define PKTSIZE_ALIGN 1536 /*#define PKTSIZE 608*/
+ /* + * IP/UDP Fragmentation support + * See: http://en.wikipedia.org/wiki/IPv4#Fragmentation_and_reassembly + * MAX possible UDP packet size is 64 KiB, if there is memory available. + */ +#define NET_ETH_MTU 1500 +#define NET_FRAG_BUF_SIZE (16 * 1024) /* MAX is 64 KiB */ +#define NET_UDP_FRAG_SIZE (NET_ETH_MTU - IP_HDR_SIZE_NO_UDP) /* 1480 */ +#define NET_FRAG_BUF_USED (NET_FRAG_BUF_SIZE / NET_UDP_FRAG_SIZE + 1) + /* * Maximum receive ring size; that is, the number of packets * we can buffer before overflow happens. Basically, this just --- u-boot.git.orig/net/net.c +++ u-boot.git/net/net.c @@ -192,6 +192,15 @@ volatile uchar PktBuf[(PKTBUFSRX+1) * PK
volatile uchar *NetRxPackets[PKTBUFSRX]; /* Receive packets */
+/* Packet fragmentation support */ +static uint16_t ip_id = 0; /* sequence number */ +static uint16_t udp_len = 0; +static uint16_t udp_src = 0; +static uint16_t udp_dst = 0; +static int max_idx = 0; +static uchar NetFragBuf[NET_FRAG_BUF_SIZE]; +static char NetFragBufUsed[NET_FRAG_BUF_USED] = { 0, }; + static rxhand_f *packetHandler; /* Current RX packet handler */ static thand_f *timeHandler; /* Current timeout handler */ static ulong timeStart; /* Time base value */ @@ -288,6 +297,13 @@ NetLoop(proto_t protocol) { bd_t *bd = gd->bd;
+ /* Packet fragmentation support */ + ip_id = udp_len = udp_src = udp_dst = max_idx = 0; + memset(NetFragBuf, 0xFF, sizeof(NetFragBuf)); + memset(NetFragBufUsed, 0, sizeof(NetFragBufUsed)); + printf("NetFragBuf @ %08x max tftp_block_size=%d udp_frag_size=%d\n", + NetFragBuf, TFTP_BLOCK_SIZE_MAX, NET_UDP_FRAG_SIZE); + #ifdef CONFIG_NET_MULTI NetRestarted = 0; NetDevExists = 0; @@ -1150,6 +1166,39 @@ static void CDPStart(void) } #endif
+#ifdef CONFIG_UDP_CHECKSUM +/* + * @sumptr: Points to UDP data + * @sumlen: Size of UDP data + * @xsum: UDP checksum across IP source, destination address, protocol and size + * + * Returns 0 when checksum is correct and 1 if it is not. + */ +static int udp_checksum(ushort *sumptr, ushort sumlen, ulong xsum) +{ + while (sumlen > 1) { + ushort sumdata; + + sumdata = *sumptr++; + xsum += ntohs(sumdata); + sumlen -= 2; + } + if (sumlen > 0) { + ushort sumdata; + + sumdata = *(unsigned char *) sumptr; + sumdata = (sumdata << 8) & 0xff00; + xsum += sumdata; + } + while ((xsum >> 16) != 0) { + xsum = (xsum & 0x0000ffff) + ((xsum >> 16) & 0x0000ffff); + } + if ((xsum != 0x00000000) && (xsum != 0x0000ffff)) + return 1; + + return 0; +} +#endif /* CONFIG_UDP_CHECKSUM */
void NetReceive(volatile uchar * inpkt, int len) @@ -1164,6 +1213,7 @@ NetReceive(volatile uchar * inpkt, int l int iscdp; #endif ushort cti = 0, vlanid = VLAN_NONE, myvlanid, mynvlanid; + uint32_t off; /* ip_off for fragmentation */
#ifdef ET_DEBUG printf("packet received\n"); @@ -1404,9 +1454,11 @@ NetReceive(volatile uchar * inpkt, int l if ((ip->ip_hl_v & 0xf0) != 0x40) { return; } +#if 0 /* Obsolete after adding the fragmentation support */ if (ip->ip_off & htons(0x1fff)) { /* Can't deal w/ fragments */ return; } +#endif /* can't deal with headers > 20 bytes */ if ((ip->ip_hl_v & 0x0f) > 0x05) { return; @@ -1422,6 +1474,88 @@ NetReceive(volatile uchar * inpkt, int l #endif return; } + + /* + * Fragmentation support. We need to check the ip_id + * and if all fragments were received correctly. + */ + off = (ntohs(ip->ip_off) & IP_OFFS) << IP_OFFS_SHIFT; + if ((off != 0) || (ip->ip_off & htons(IP_FLAGS_MFRAG))) { + int size, idx, complete; + char *start; + + /* New fragmented packet arrived, clear data. */ + if (ntohs(ip->ip_id) != ip_id) { + ip_id = ntohs(ip->ip_id); + memset(NetFragBufUsed, 0, sizeof(NetFragBufUsed)); + udp_len = udp_src = udp_dst = max_idx = 0; + } + + idx = off / NET_UDP_FRAG_SIZE; + + /* Packet does not fit into IP/UDP fragmentation buf */ + if (idx >= NET_FRAG_BUF_USED) { + return; + } + + NetFragBufUsed[idx] = 1; + + /* Copy the UDP hdr with the data for 1st + fragment, else copy just payload */ + if (off == 0) { + udp_len = ntohs(ip->udp_len); + udp_src = ntohs(ip->udp_src); + udp_dst = ntohs(ip->udp_dst); + } + size = ntohs(ip->ip_len) - IP_HDR_SIZE_NO_UDP; + start = (char *)ip + IP_HDR_SIZE_NO_UDP; + memcpy(NetFragBuf + off, start, size); + + /* + * When last fragement has been received we + * know the number of fragments we expect. If + * all have arrived we process the packet. + */ + if (((off != 0) && !(ip->ip_off & htons(IP_FLAGS_MFRAG)))) + max_idx = idx; + + if (max_idx == 0) + return; + + complete = 1; + for (idx = 0; idx < max_idx; idx++) { + if (NetFragBufUsed[idx] == 0) { + complete = 0; + break; + } + } + if (!complete) + return; +#ifdef CONFIG_UDP_CHECKSUM + if (ip->udp_xsum != 0) { + ulong xsum = ip->ip_p; + uint16_t *sumptr; + + xsum += udp_len; + xsum += (ntohl(ip->ip_src) >> 16) & 0xffff; + xsum += (ntohl(ip->ip_src) >> 0) & 0xffff; + xsum += (ntohl(ip->ip_dst) >> 16) & 0xffff; + xsum += (ntohl(ip->ip_dst) >> 0) & 0xffff; + sumptr = (ushort *)NetFragBuf; + + if (udp_checksum(sumptr, udp_len, xsum)) { + putc('U'); + return; + } + } +#endif /* CONFIG_UDP_CHECKSUM */ + (*packetHandler)(NetFragBuf + 8, + udp_dst, + udp_src, + udp_len - 8); + return; + } + /* * watch for ICMP host redirects * @@ -1502,26 +1636,8 @@ NetReceive(volatile uchar * inpkt, int l sumlen = ntohs(ip->udp_len); sumptr = (ushort *) &(ip->udp_src);
- while (sumlen > 1) { - ushort sumdata; - - sumdata = *sumptr++; - xsum += ntohs(sumdata); - sumlen -= 2; - } - if (sumlen > 0) { - ushort sumdata; - - sumdata = *(unsigned char *) sumptr; - sumdata = (sumdata << 8) & 0xff00; - xsum += sumdata; - } - while ((xsum >> 16) != 0) { - xsum = (xsum & 0x0000ffff) + ((xsum >> 16) & 0x0000ffff); - } - if ((xsum != 0x00000000) && (xsum != 0x0000ffff)) { - printf(" UDP wrong checksum %08lx %08x\n", - xsum, ntohs(ip->udp_xsum)); + if (udp_checksum(sumptr, sumlen, xsum)) { + putc('U'); return; } } --- u-boot.git.orig/net/tftp.c +++ u-boot.git/net/tftp.c @@ -456,6 +456,7 @@ TftpTimeout (void) void TftpStart (void) { + char *s, *err; #ifdef CONFIG_TFTP_PORT char *ep; /* Environment pointer */ #endif @@ -518,6 +519,27 @@ TftpStart (void)
puts ("Loading: *\b");
+ /* Get alternate tftp_block_size */ + if ((s = getenv("tftp_block_size")) != NULL) { + err = NULL; + + TftpBlkSizeOption = simple_strtoul(s, &err, 10); + if (*err) { + printf("ERR: "tftp_block_size" is not a number\n"); + TftpBlkSizeOption = TFTP_BLOCK_SIZE; + } + /* + * Reject values which require extensive handling. + * block size of 1428 octets (Ethernet MTU, less + * the TFTP, UDP and IP header lengths). + */ + if (TftpBlkSizeOption > TFTP_BLOCK_SIZE_MAX) { + printf("ERR: tftp_block_sizes larger than %d not " + "supported\n", TFTP_BLOCK_SIZE_MAX); + TftpBlkSizeOption = TFTP_BLOCK_SIZE; + } + } + NetSetTimeout (TIMEOUT * CFG_HZ, TftpTimeout); NetSetHandler (TftpHandler);
--- u-boot.git.orig/net/tftp.h +++ u-boot.git/net/tftp.h @@ -8,11 +8,21 @@ #ifndef __TFTP_H__ #define __TFTP_H__
+#include <net.h> + /**********************************************************************/ /* * Global functions and variables. */
+/* + * Maximum TFTP block size bound to max size of fragmented IP/UDP + * packets minus TFTP and UDP/IP overhead. TFTP overhead is 2 byte + * opcode and 2 byte block-number. + */ +#define TFTP_BLOCK_SIZE_MAX (NET_FRAG_BUF_SIZE - sizeof(IP_t) - 4) + + /* tftp.c */ extern void TftpStart (void); /* Begin TFTP get */
Josh Boyer wrote:
From: Frank Haverkamp haver@vnet.ibm.com
http://tools.ietf.org/html/rfc2348 describes the TFTP block size option which allows larger packtes than the 512 byte default. This reduces the number of TFTP ACKs significantly and improves performance.
To get the most benefit out of the tftp block size option the support of defragementation of IP/UDP packet is helpful. The current implemenation should work even with packets received out of order. To enable the large packet size the user should set "tftp_block_size" so a value like 16352.
We experimented with different packet sizes and found that more than those 16KiB do not contribute much to the performance anymore. Therefor I limited the defragmentation buffer to 16KiB no too waste memory.
Signed-off-by: Frank Haverkamp haver@vnet.ibm.com Signed-off-by: Josh Boyer jwboyer@linux.vnet.ibm.com
Thanks. This looks like a good feature. I'll try to digest it tonight.
regards, Ben
Hi Josh & Frank,
A few nits. Overall a nice patch.
One question: Will multicast TFTP still work when this is applied?
Josh Boyer wrote:
From: Frank Haverkamp haver@vnet.ibm.com
http://tools.ietf.org/html/rfc2348 describes the TFTP block size option which allows larger packtes than the 512 byte default. This reduces the
s/packtes/packets/
number of TFTP ACKs significantly and improves performance.
To get the most benefit out of the tftp block size option the support of defragementation of IP/UDP packet is helpful. The current implemenation
s/defragementation/defragmentation/ s/implemenation/implementation/
should work even with packets received out of order. To enable the large packet size the user should set "tftp_block_size" so a value like 16352.
s/so/to/
We experimented with different packet sizes and found that more than those 16KiB do not contribute much to the performance anymore. Therefor I limited
s/Therefor/Therefore/
the defragmentation buffer to 16KiB no too waste memory.
so as to not waste memory
Signed-off-by: Frank Haverkamp haver@vnet.ibm.com Signed-off-by: Josh Boyer jwboyer@linux.vnet.ibm.com
include/net.h | 17 ++++++ net/net.c | 156 ++++++++++++++++++++++++++++++++++++++++++++++++++-------- net/tftp.c | 22 ++++++++ net/tftp.h | 10 +++ 4 files changed, 185 insertions(+), 20 deletions(-)
--- u-boot.git.orig/include/net.h +++ u-boot.git/include/net.h @@ -200,6 +200,13 @@ typedef struct { ushort udp_xsum; /* Checksum */ } IP_t;
+#define IP_OFFS 0x1FFF /* ip offset *= 8 */ +#define IP_OFFS_SHIFT 3 /* in 8 byte steps */ +#define IP_FLAGS 0xE000 /* first 3 bits */ +#define IP_FLAGS_RES 0x8000 /* reserved */ +#define IP_FLAGS_DFRAG 0x4000 /* don't fragments */ +#define IP_FLAGS_MFRAG 0x2000 /* more fragments */
Please align these better. After applying, it's:
#define IP_OFFS 0x1FFF /* ip offset *= 8 */ #define IP_OFFS_SHIFT 3 /* in 8 byte steps */ #define IP_FLAGS 0xE000 /* first 3 bits */ #define IP_FLAGS_RES 0x8000 /* reserved */ #define IP_FLAGS_DFRAG 0x4000 /* don't fragments */ #define IP_FLAGS_MFRAG 0x2000 /* more fragments */
#define IP_HDR_SIZE_NO_UDP (sizeof (IP_t) - 8) #define IP_HDR_SIZE (sizeof (IP_t))
@@ -282,6 +289,16 @@ typedef struct icmphdr { #define PKTSIZE_ALIGN 1536 /*#define PKTSIZE 608*/
- /*
- IP/UDP Fragmentation support
- MAX possible UDP packet size is 64 KiB, if there is memory available.
- */
+#define NET_ETH_MTU 1500 +#define NET_FRAG_BUF_SIZE (16 * 1024) /* MAX is 64 KiB */ +#define NET_UDP_FRAG_SIZE (NET_ETH_MTU - IP_HDR_SIZE_NO_UDP) /* 1480 */ +#define NET_FRAG_BUF_USED (NET_FRAG_BUF_SIZE / NET_UDP_FRAG_SIZE + 1)
/*
- Maximum receive ring size; that is, the number of packets
- we can buffer before overflow happens. Basically, this just
--- u-boot.git.orig/net/net.c +++ u-boot.git/net/net.c @@ -192,6 +192,15 @@ volatile uchar PktBuf[(PKTBUFSRX+1) * PK
volatile uchar *NetRxPackets[PKTBUFSRX]; /* Receive packets */
+/* Packet fragmentation support */ +static uint16_t ip_id = 0; /* sequence number */ +static uint16_t udp_len = 0; +static uint16_t udp_src = 0; +static uint16_t udp_dst = 0; +static int max_idx = 0; +static uchar NetFragBuf[NET_FRAG_BUF_SIZE]; +static char NetFragBufUsed[NET_FRAG_BUF_USED] = { 0, };
static rxhand_f *packetHandler; /* Current RX packet handler */ static thand_f *timeHandler; /* Current timeout handler */ static ulong timeStart; /* Time base value */ @@ -288,6 +297,13 @@ NetLoop(proto_t protocol) { bd_t *bd = gd->bd;
- /* Packet fragmentation support */
- ip_id = udp_len = udp_src = udp_dst = max_idx = 0;
- memset(NetFragBuf, 0xFF, sizeof(NetFragBuf));
- memset(NetFragBufUsed, 0, sizeof(NetFragBufUsed));
- printf("NetFragBuf @ %08x max tftp_block_size=%d udp_frag_size=%d\n",
NetFragBuf, TFTP_BLOCK_SIZE_MAX, NET_UDP_FRAG_SIZE);#ifdef CONFIG_NET_MULTI NetRestarted = 0; NetDevExists = 0; @@ -1150,6 +1166,39 @@ static void CDPStart(void) } #endif
+#ifdef CONFIG_UDP_CHECKSUM +/*
- @sumptr: Points to UDP data
- @sumlen: Size of UDP data
- @xsum: UDP checksum across IP source, destination address, protocol and size
- Returns 0 when checksum is correct and 1 if it is not
Can you return -1 on failure? Not a big deal, just more conventional.
- */
+static int udp_checksum(ushort *sumptr, ushort sumlen, ulong xsum) +{
while (sumlen > 1) {ushort sumdata;sumdata = *sumptr++;xsum += ntohs(sumdata);sumlen -= 2;}if (sumlen > 0) {ushort sumdata;sumdata = *(unsigned char *) sumptr;sumdata = (sumdata << 8) & 0xff00;xsum += sumdata;}while ((xsum >> 16) != 0) {xsum = (xsum & 0x0000ffff) + ((xsum >> 16) & 0x0000ffff);}if ((xsum != 0x00000000) && (xsum != 0x0000ffff))return 1;return 0;+} +#endif /* CONFIG_UDP_CHECKSUM */
void NetReceive(volatile uchar * inpkt, int len) @@ -1164,6 +1213,7 @@ NetReceive(volatile uchar * inpkt, int l int iscdp; #endif ushort cti = 0, vlanid = VLAN_NONE, myvlanid, mynvlanid;
- uint32_t off; /* ip_off for fragmentation */
Can you pick a better variable name than 'off'? 'offset' maybe?
#ifdef ET_DEBUG printf("packet received\n"); @@ -1404,9 +1454,11 @@ NetReceive(volatile uchar * inpkt, int l if ((ip->ip_hl_v & 0xf0) != 0x40) { return; } +#if 0 /* Obsolete after adding the fragmentation support */ if (ip->ip_off & htons(0x1fff)) { /* Can't deal w/ fragments */ return; } +#endif
Please delete this block if it's now dead code.
/* can't deal with headers > 20 bytes */ if ((ip->ip_hl_v & 0x0f) > 0x05) { return;@@ -1422,6 +1474,88 @@ NetReceive(volatile uchar * inpkt, int l #endif return; }
/** Fragmentation support. We need to check the ip_id* and if all fragments were received correctly.*/off = (ntohs(ip->ip_off) & IP_OFFS) << IP_OFFS_SHIFT;if ((off != 0) || (ip->ip_off & htons(IP_FLAGS_MFRAG))) {int size, idx, complete;char *start;/* New fragmented packet arrived, clear data. */if (ntohs(ip->ip_id) != ip_id) {ip_id = ntohs(ip->ip_id);memset(NetFragBufUsed, 0, sizeof(NetFragBufUsed));udp_len = udp_src = udp_dst = max_idx = 0;}idx = off / NET_UDP_FRAG_SIZE;/* Packet does not fit into IP/UDP fragmentation buf */if (idx >= NET_FRAG_BUF_USED) {return;}NetFragBufUsed[idx] = 1;/* Copy the UDP hdr with the data for 1stfragment, else copy just payload */if (off == 0) {udp_len = ntohs(ip->udp_len);udp_src = ntohs(ip->udp_src);udp_dst = ntohs(ip->udp_dst);}size = ntohs(ip->ip_len) - IP_HDR_SIZE_NO_UDP;start = (char *)ip + IP_HDR_SIZE_NO_UDP;memcpy(NetFragBuf + off, start, size);/** When last fragement has been received we
s/fragement/fragment/
* know the number of fragments we expect. If* all have arrived we process the packet.*/if (((off != 0) && !(ip->ip_off & htons(IP_FLAGS_MFRAG))))max_idx = idx;if (max_idx == 0)return;complete = 1;for (idx = 0; idx < max_idx; idx++) {if (NetFragBufUsed[idx] == 0) {complete = 0;break;}}if (!complete)return;+#ifdef CONFIG_UDP_CHECKSUM
if (ip->udp_xsum != 0) {ulong xsum = ip->ip_p;uint16_t *sumptr;xsum += udp_len;xsum += (ntohl(ip->ip_src) >> 16) & 0xffff;xsum += (ntohl(ip->ip_src) >> 0) & 0xffff;xsum += (ntohl(ip->ip_dst) >> 16) & 0xffff;xsum += (ntohl(ip->ip_dst) >> 0) & 0xffff;sumptr = (ushort *)NetFragBuf;if (udp_checksum(sumptr, udp_len, xsum)) {putc('U');return;}}+#endif /* CONFIG_UDP_CHECKSUM */
(*packetHandler)(NetFragBuf + 8,udp_dst,udp_src,udp_len - 8);
Can some of these arguments share lines?
return;}- /*
- watch for ICMP host redirects
@@ -1502,26 +1636,8 @@ NetReceive(volatile uchar * inpkt, int l sumlen = ntohs(ip->udp_len); sumptr = (ushort *) &(ip->udp_src);
while (sumlen > 1) {ushort sumdata;sumdata = *sumptr++;xsum += ntohs(sumdata);sumlen -= 2;}if (sumlen > 0) {ushort sumdata;sumdata = *(unsigned char *) sumptr;sumdata = (sumdata << 8) & 0xff00;xsum += sumdata;}while ((xsum >> 16) != 0) {xsum = (xsum & 0x0000ffff) + ((xsum >> 16) & 0x0000ffff);}if ((xsum != 0x00000000) && (xsum != 0x0000ffff)) {printf(" UDP wrong checksum %08lx %08x\n",xsum, ntohs(ip->udp_xsum));
if (udp_checksum(sumptr, sumlen, xsum)) {putc('U'); return; }--- u-boot.git.orig/net/tftp.c +++ u-boot.git/net/tftp.c @@ -456,6 +456,7 @@ TftpTimeout (void) void TftpStart (void) {
- char *s, *err;
#ifdef CONFIG_TFTP_PORT char *ep; /* Environment pointer */ #endif @@ -518,6 +519,27 @@ TftpStart (void)
puts ("Loading: *\b");
- /* Get alternate tftp_block_size */
- if ((s = getenv("tftp_block_size")) != NULL) {
Kind of a long environment variable name. Maybe "tftp_bs"?
err = NULL;TftpBlkSizeOption = simple_strtoul(s, &err, 10);if (*err) {printf("ERR: \"tftp_block_size\" is not a number\n");TftpBlkSizeOption = TFTP_BLOCK_SIZE;}/** Reject values which require extensive handling.* block size of 1428 octets (Ethernet MTU, less* the TFTP, UDP and IP header lengths).*/if (TftpBlkSizeOption > TFTP_BLOCK_SIZE_MAX) {printf("ERR: tftp_block_sizes larger than %d not ""supported\n", TFTP_BLOCK_SIZE_MAX);TftpBlkSizeOption = TFTP_BLOCK_SIZE;}- }
- NetSetTimeout (TIMEOUT * CFG_HZ, TftpTimeout); NetSetHandler (TftpHandler);
--- u-boot.git.orig/net/tftp.h +++ u-boot.git/net/tftp.h @@ -8,11 +8,21 @@ #ifndef __TFTP_H__ #define __TFTP_H__
+#include <net.h>
/**********************************************************************/ /*
- Global functions and variables.
*/
+/*
- Maximum TFTP block size bound to max size of fragmented IP/UDP
- packets minus TFTP and UDP/IP overhead. TFTP overhead is 2 byte
- opcode and 2 byte block-number.
- */
+#define TFTP_BLOCK_SIZE_MAX (NET_FRAG_BUF_SIZE - sizeof(IP_t) - 4)
/* tftp.c */ extern void TftpStart (void); /* Begin TFTP get */
regards, Ben
Subject: U-Boot: Add simple IP/UDP fragmentation support From: Frank Haverkamp haver@vnet.ibm.com
http://tools.ietf.org/html/rfc2348 describes the TFTP block size option which allows larger packets than the 512 byte default. This reduces the number of TFTP ACKs significantly and improves performance.
To get the most benefit out of the tftp block size option the support of defragmentation of IP/UDP packet is helpful. The current implementation should work even with packets received out of order. To enable the large packet size the user should set "tftp_bsize" to a value like 16352.
We experimented with different packet sizes and found that more than those 16KiB do not contribute much to the performance anymore. Therefore I limited the defragmentation buffer to 16KiB as not to not waste memory.
Multicast TFTP is not yet tested with this patch. Would be nice if someone who has set this up could give it a try.
Signed-off-by: Frank Haverkamp haver@vnet.ibm.com Signed-off-by: Josh Boyer jwboyer@linux.vnet.ibm.com
--- include/net.h | 17 ++++++ net/net.c | 155 +++++++++++++++++++++++++++++++++++++++++++++++++--------- net/tftp.c | 22 ++++++++ net/tftp.h | 10 +++ 4 files changed, 182 insertions(+), 22 deletions(-)
--- u-boot.git.orig/include/net.h +++ u-boot.git/include/net.h @@ -200,6 +200,13 @@ typedef struct { ushort udp_xsum; /* Checksum */ } IP_t;
+#define IP_OFFS 0x1FFF /* ip offset *= 8 */ +#define IP_OFFS_SHIFT 3 /* in 8 byte steps */ +#define IP_FLAGS 0xE000 /* first 3 bits */ +#define IP_FLAGS_RES 0x8000 /* reserved */ +#define IP_FLAGS_DFRAG 0x4000 /* don't fragments */ +#define IP_FLAGS_MFRAG 0x2000 /* more fragments */ + #define IP_HDR_SIZE_NO_UDP (sizeof (IP_t) - 8) #define IP_HDR_SIZE (sizeof (IP_t))
@@ -282,6 +289,16 @@ typedef struct icmphdr { #define PKTSIZE_ALIGN 1536 /*#define PKTSIZE 608*/
+ /* + * IP/UDP Fragmentation support + * See: http://en.wikipedia.org/wiki/IPv4#Fragmentation_and_reassembly + * MAX possible UDP packet size is 64 KiB, if there is memory available. + */ +#define NET_ETH_MTU 1500 +#define NET_FRAG_BUF_SIZE (16 * 1024) /* MAX is 64 KiB */ +#define NET_UDP_FRAG_SIZE (NET_ETH_MTU - IP_HDR_SIZE_NO_UDP) /* 1480 */ +#define NET_FRAG_BUF_USED (NET_FRAG_BUF_SIZE / NET_UDP_FRAG_SIZE + 1) + /* * Maximum receive ring size; that is, the number of packets * we can buffer before overflow happens. Basically, this just --- u-boot.git.orig/net/net.c +++ u-boot.git/net/net.c @@ -192,6 +192,15 @@ volatile uchar PktBuf[(PKTBUFSRX+1) * PK
volatile uchar *NetRxPackets[PKTBUFSRX]; /* Receive packets */
+/* Packet fragmentation support */ +static uint16_t ip_id = 0; /* sequence number */ +static uint16_t udp_len = 0; +static uint16_t udp_src = 0; +static uint16_t udp_dst = 0; +static int max_idx = 0; +static uchar NetFragBuf[NET_FRAG_BUF_SIZE]; +static char NetFragBufUsed[NET_FRAG_BUF_USED] = { 0, }; + static rxhand_f *packetHandler; /* Current RX packet handler */ static thand_f *timeHandler; /* Current timeout handler */ static ulong timeStart; /* Time base value */ @@ -288,6 +297,13 @@ NetLoop(proto_t protocol) { bd_t *bd = gd->bd;
+ /* Packet fragmentation support */ + ip_id = udp_len = udp_src = udp_dst = max_idx = 0; + memset(NetFragBuf, 0xFF, sizeof(NetFragBuf)); + memset(NetFragBufUsed, 0, sizeof(NetFragBufUsed)); + printf("NetFragBuf @ %08x max tftp_bsize=%d udp_frag_size=%d\n", + NetFragBuf, TFTP_BLOCK_SIZE_MAX, NET_UDP_FRAG_SIZE); + #ifdef CONFIG_NET_MULTI NetRestarted = 0; NetDevExists = 0; @@ -1150,6 +1166,39 @@ static void CDPStart(void) } #endif
+#ifdef CONFIG_UDP_CHECKSUM +/* + * @sumptr: Points to UDP data + * @sumlen: Size of UDP data + * @xsum: UDP checksum across IP source, destination address, protocol and size + * + * Returns 0 when checksum is correct and -1 if it is not. + */ +static int udp_checksum(ushort *sumptr, ushort sumlen, ulong xsum) +{ + while (sumlen > 1) { + ushort sumdata; + + sumdata = *sumptr++; + xsum += ntohs(sumdata); + sumlen -= 2; + } + if (sumlen > 0) { + ushort sumdata; + + sumdata = *(unsigned char *) sumptr; + sumdata = (sumdata << 8) & 0xff00; + xsum += sumdata; + } + while ((xsum >> 16) != 0) { + xsum = (xsum & 0x0000ffff) + ((xsum >> 16) & 0x0000ffff); + } + if ((xsum != 0x00000000) && (xsum != 0x0000ffff)) + return -1; + + return 0; +} +#endif /* CONFIG_UDP_CHECKSUM */
void NetReceive(volatile uchar * inpkt, int len) @@ -1164,6 +1213,7 @@ NetReceive(volatile uchar * inpkt, int l int iscdp; #endif ushort cti = 0, vlanid = VLAN_NONE, myvlanid, mynvlanid; + uint32_t ip_off; /* ip_off for fragmentation */
#ifdef ET_DEBUG printf("packet received\n"); @@ -1404,9 +1454,7 @@ NetReceive(volatile uchar * inpkt, int l if ((ip->ip_hl_v & 0xf0) != 0x40) { return; } - if (ip->ip_off & htons(0x1fff)) { /* Can't deal w/ fragments */ - return; - } + /* can't deal with headers > 20 bytes */ if ((ip->ip_hl_v & 0x0f) > 0x05) { return; @@ -1422,6 +1470,86 @@ NetReceive(volatile uchar * inpkt, int l #endif return; } + + /* + * Fragmentation support. We need to check the ip_id + * and if all fragments were received correctly. + */ + ip_off = (ntohs(ip->ip_off) & IP_OFFS) << IP_OFFS_SHIFT; + if ((ip_off != 0) || (ip->ip_off & htons(IP_FLAGS_MFRAG))) { + int size, idx, complete; + char *start; + + /* New fragmented packet arrived, clear data. */ + if (ntohs(ip->ip_id) != ip_id) { + ip_id = ntohs(ip->ip_id); + memset(NetFragBufUsed, 0, sizeof(NetFragBufUsed)); + udp_len = udp_src = udp_dst = max_idx = 0; + } + + idx = ip_off / NET_UDP_FRAG_SIZE; + + /* Packet does not fit into IP/UDP fragmentation buf */ + if (idx >= NET_FRAG_BUF_USED) { + return; + } + + NetFragBufUsed[idx] = 1; + + /* Copy the UDP hdr with the data for 1st + fragment, else copy just payload */ + if (ip_off == 0) { + udp_len = ntohs(ip->udp_len); + udp_src = ntohs(ip->udp_src); + udp_dst = ntohs(ip->udp_dst); + } + size = ntohs(ip->ip_len) - IP_HDR_SIZE_NO_UDP; + start = (char *)ip + IP_HDR_SIZE_NO_UDP; + memcpy(NetFragBuf + ip_off, start, size); + + /* + * When last fragment has been received we + * know the number of fragments we expect. If + * all have arrived we process the packet. + */ + if (((ip_off != 0) && !(ip->ip_off & htons(IP_FLAGS_MFRAG)))) + max_idx = idx; + + if (max_idx == 0) + return; + + complete = 1; + for (idx = 0; idx < max_idx; idx++) { + if (NetFragBufUsed[idx] == 0) { + complete = 0; + break; + } + } + if (!complete) + return; +#ifdef CONFIG_UDP_CHECKSUM + if (ip->udp_xsum != 0) { + ulong xsum = ip->ip_p; + uint16_t *sumptr; + + xsum += udp_len; + xsum += (ntohl(ip->ip_src) >> 16) & 0xffff; + xsum += (ntohl(ip->ip_src) >> 0) & 0xffff; + xsum += (ntohl(ip->ip_dst) >> 16) & 0xffff; + xsum += (ntohl(ip->ip_dst) >> 0) & 0xffff; + sumptr = (ushort *)NetFragBuf; + + if (udp_checksum(sumptr, udp_len, xsum)) { + putc('U'); + return; + } + } +#endif /* CONFIG_UDP_CHECKSUM */ + (*packetHandler)(NetFragBuf + 8, udp_dst, udp_src, + udp_len - 8); + return; + } + /* * watch for ICMP host redirects * @@ -1502,26 +1630,9 @@ NetReceive(volatile uchar * inpkt, int l sumlen = ntohs(ip->udp_len); sumptr = (ushort *) &(ip->udp_src);
- while (sumlen > 1) { - ushort sumdata; - - sumdata = *sumptr++; - xsum += ntohs(sumdata); - sumlen -= 2; - } - if (sumlen > 0) { - ushort sumdata;
- sumdata = *(unsigned char *) sumptr; - sumdata = (sumdata << 8) & 0xff00; - xsum += sumdata; - } - while ((xsum >> 16) != 0) { - xsum = (xsum & 0x0000ffff) + ((xsum >> 16) & 0x0000ffff); - } - if ((xsum != 0x00000000) && (xsum != 0x0000ffff)) { - printf(" UDP wrong checksum %08lx %08x\n", - xsum, ntohs(ip->udp_xsum)); + if (udp_checksum(sumptr, sumlen, xsum)) { + putc('U'); return; } } --- u-boot.git.orig/net/tftp.c +++ u-boot.git/net/tftp.c @@ -456,6 +456,7 @@ TftpTimeout (void) void TftpStart (void) { + char *s, *err; #ifdef CONFIG_TFTP_PORT char *ep; /* Environment pointer */ #endif @@ -518,6 +519,27 @@ TftpStart (void)
puts ("Loading: *\b");
+ /* Get alternate tftp_bsize */ + if ((s = getenv("tftp_bsize")) != NULL) { + err = NULL; + + TftpBlkSizeOption = simple_strtoul(s, &err, 10); + if (*err) { + printf("ERR: "tftp_bsize" is not a number\n"); + TftpBlkSizeOption = TFTP_BLOCK_SIZE; + } + /* + * Reject values which require extensive handling. + * block size of 1428 octets (Ethernet MTU, less + * the TFTP, UDP and IP header lengths). + */ + if (TftpBlkSizeOption > TFTP_BLOCK_SIZE_MAX) { + printf("ERR: tftp_bsize larger than %d not supported\n", + TFTP_BLOCK_SIZE_MAX); + TftpBlkSizeOption = TFTP_BLOCK_SIZE; + } + } + NetSetTimeout (TIMEOUT * CFG_HZ, TftpTimeout); NetSetHandler (TftpHandler);
--- u-boot.git.orig/net/tftp.h +++ u-boot.git/net/tftp.h @@ -8,11 +8,21 @@ #ifndef __TFTP_H__ #define __TFTP_H__
+#include <net.h> + /**********************************************************************/ /* * Global functions and variables. */
+/* + * Maximum TFTP block size bound to max size of fragmented IP/UDP + * packets minus TFTP and UDP/IP overhead. TFTP overhead is 2 byte + * opcode and 2 byte block-number. + */ +#define TFTP_BLOCK_SIZE_MAX (NET_FRAG_BUF_SIZE - sizeof(IP_t) - 4) + + /* tftp.c */ extern void TftpStart (void); /* Begin TFTP get */
Dear Ben,
In message 1220460631.17453.3.camel@august Frank Haverkamp wrote:
Subject: U-Boot: Add simple IP/UDP fragmentation support From: Frank Haverkamp haver@vnet.ibm.com
You replied:
| I know you guys posted this one a couple of months ago and I apologize | for not being more active. I've put this in the net/testing repo with | the hope that people will smack it around a bit.
What to do with this now?
Best regards,
Wolfgang Denk
Wolfgang Denk wrote:
Dear Ben,
In message 1220460631.17453.3.camel@august Frank Haverkamp wrote:
Subject: U-Boot: Add simple IP/UDP fragmentation support From: Frank Haverkamp haver@vnet.ibm.com
You replied:
| I know you guys posted this one a couple of months ago and I apologize | for not being more active. I've put this in the net/testing repo with | the hope that people will smack it around a bit.
What to do with this now?
Best regards,
Wolfgang Denk
We may as well pull it in. I have quite a bit of net-related stuff in the next few days and will include this patch.
regards, Ben
Hi Josh,
Josh Boyer wrote:
From: Frank Haverkamp haver@vnet.ibm.com
http://tools.ietf.org/html/rfc2348 describes the TFTP block size option which allows larger packtes than the 512 byte default. This reduces the number of TFTP ACKs significantly and improves performance.
To get the most benefit out of the tftp block size option the support of defragementation of IP/UDP packet is helpful. The current implemenation should work even with packets received out of order. To enable the large packet size the user should set "tftp_block_size" so a value like 16352.
We experimented with different packet sizes and found that more than those 16KiB do not contribute much to the performance anymore. Therefor I limited the defragmentation buffer to 16KiB no too waste memory.
Signed-off-by: Frank Haverkamp haver@vnet.ibm.com Signed-off-by: Josh Boyer jwboyer@linux.vnet.ibm.com
I know you guys posted this one a couple of months ago and I apologize for not being more active. I've put this in the net/testing repo with the hope that people will smack it around a bit.
regards, Ben
Josh & Frank,
Sorry again for dropping the ball on this one. This is a useful feature that we should add to U-boot, but I'm concerned about the memory footprint.
Josh Boyer wrote:
From: Frank Haverkamp haver@vnet.ibm.com
http://tools.ietf.org/html/rfc2348 describes the TFTP block size option which allows larger packtes than the 512 byte default. This reduces the number of TFTP ACKs significantly and improves performance.
To get the most benefit out of the tftp block size option the support of defragementation of IP/UDP packet is helpful. The current implemenation should work even with packets received out of order. To enable the large packet size the user should set "tftp_block_size" so a value like 16352.
We experimented with different packet sizes and found that more than those 16KiB do not contribute much to the performance anymore. Therefor I limited the defragmentation buffer to 16KiB no too waste memory.
Signed-off-by: Frank Haverkamp haver@vnet.ibm.com Signed-off-by: Josh Boyer jwboyer@linux.vnet.ibm.com
include/net.h | 17 ++++++ net/net.c | 156 ++++++++++++++++++++++++++++++++++++++++++++++++++-------- net/tftp.c | 22 ++++++++ net/tftp.h | 10 +++ 4 files changed, 185 insertions(+), 20 deletions(-)
--- u-boot.git.orig/include/net.h +++ u-boot.git/include/net.h @@ -200,6 +200,13 @@ typedef struct { ushort udp_xsum; /* Checksum */ } IP_t;
+#define IP_OFFS 0x1FFF /* ip offset *= 8 */ +#define IP_OFFS_SHIFT 3 /* in 8 byte steps */ +#define IP_FLAGS 0xE000 /* first 3 bits */ +#define IP_FLAGS_RES 0x8000 /* reserved */ +#define IP_FLAGS_DFRAG 0x4000 /* don't fragments */ +#define IP_FLAGS_MFRAG 0x2000 /* more fragments */
#define IP_HDR_SIZE_NO_UDP (sizeof (IP_t) - 8) #define IP_HDR_SIZE (sizeof (IP_t))
@@ -282,6 +289,16 @@ typedef struct icmphdr { #define PKTSIZE_ALIGN 1536 /*#define PKTSIZE 608*/
- /*
- IP/UDP Fragmentation support
- MAX possible UDP packet size is 64 KiB, if there is memory available.
- */
+#define NET_ETH_MTU 1500 +#define NET_FRAG_BUF_SIZE (16 * 1024) /* MAX is 64 KiB */ +#define NET_UDP_FRAG_SIZE (NET_ETH_MTU - IP_HDR_SIZE_NO_UDP) /* 1480 */ +#define NET_FRAG_BUF_USED (NET_FRAG_BUF_SIZE / NET_UDP_FRAG_SIZE + 1)
/*
- Maximum receive ring size; that is, the number of packets
- we can buffer before overflow happens. Basically, this just
--- u-boot.git.orig/net/net.c +++ u-boot.git/net/net.c @@ -192,6 +192,15 @@ volatile uchar PktBuf[(PKTBUFSRX+1) * PK
volatile uchar *NetRxPackets[PKTBUFSRX]; /* Receive packets */
+/* Packet fragmentation support */ +static uint16_t ip_id = 0; /* sequence number */ +static uint16_t udp_len = 0; +static uint16_t udp_src = 0; +static uint16_t udp_dst = 0; +static int max_idx = 0; +static uchar NetFragBuf[NET_FRAG_BUF_SIZE];
Allocating a 16k buffer should be optional. Please consider wrapping this feature in a CONFIG.
+static char NetFragBufUsed[NET_FRAG_BUF_USED] = { 0, };
static rxhand_f *packetHandler; /* Current RX packet handler */ static thand_f *timeHandler; /* Current timeout handler */ static ulong timeStart; /* Time base value */ @@ -288,6 +297,13 @@ NetLoop(proto_t protocol) { bd_t *bd = gd->bd;
- /* Packet fragmentation support */
- ip_id = udp_len = udp_src = udp_dst = max_idx = 0;
- memset(NetFragBuf, 0xFF, sizeof(NetFragBuf));
- memset(NetFragBufUsed, 0, sizeof(NetFragBufUsed));
- printf("NetFragBuf @ %08x max tftp_block_size=%d udp_frag_size=%d\n",
NetFragBuf, TFTP_BLOCK_SIZE_MAX, NET_UDP_FRAG_SIZE);
I get a compiler warning here that a simple cast should fix: net.c: In function 'NetLoop': net.c:300: warning: format '%08x' expects type 'unsigned int', but argument 2 has type 'uchar *'
regards, Ben
participants (4)
-
Ben Warren -
Frank Haverkamp -
Josh Boyer -
Wolfgang Denk