On Sun, Mar 1, 2015 at 12:07 PM, Simon Glass sjg@chromium.org wrote:
Hi Joe,
On 24 February 2015 at 17:02, Joe Hershberger joe.hershberger@ni.com
wrote:
The 'lo' interface on Linux doesn't support thinks like ARP or link-layer access like we use to talk to a normal network interface. A higher-level network API must be used to access localhost.
As written, this interface is limited to not supporting ICMP since the API doesn't allow the socket to be opened for all IP traffic and be able to receive at the same time. UDP is far more useful to test with, so it was selected over ICMP. Ping won't work, but things like TFTP should work.
Signed-off-by: Joe Hershberger joe.hershberger@ni.com
Changes in v4: -Added support for the 'lo' network interface
Changes in v3: None Changes in v2: None
arch/sandbox/cpu/eth-raw-os.c | 152
+++++++++++++++++++++++++++-------
arch/sandbox/dts/sandbox.dts | 10 +++ arch/sandbox/include/asm/eth-raw-os.h | 10 ++- drivers/net/sandbox-raw.c | 62 +++++++++++++- 4 files changed, 203 insertions(+), 31 deletions(-)
diff --git a/arch/sandbox/cpu/eth-raw-os.c
b/arch/sandbox/cpu/eth-raw-os.c
index 9218f94..acb150a 100644 --- a/arch/sandbox/cpu/eth-raw-os.c +++ b/arch/sandbox/cpu/eth-raw-os.c @@ -11,6 +11,8 @@ #include <errno.h> #include <net/if.h> #include <netinet/in.h> +#include <netinet/ip.h> +#include <netinet/udp.h> #include <stdio.h> #include <stdlib.h> #include <string.h> @@ -19,51 +21,139 @@ #include <sys/socket.h> #include <unistd.h>
+#include <arpa/inet.h> #include <linux/if_ether.h> #include <linux/if_packet.h>
int sandbox_eth_raw_os_init(const char *ifname, unsigned char *ethmac, struct eth_sandbox_raw_priv *priv) {
struct sockaddr_ll *device;struct packet_mreq mr;/* Prepare device struct */priv->device = malloc(sizeof(struct sockaddr_ll));device = priv->device;memset(device, 0, sizeof(struct sockaddr_ll));device->sll_ifindex = if_nametoindex(ifname);device->sll_family = AF_PACKET;memcpy(device->sll_addr, ethmac, 6);device->sll_halen = htons(6);/* Open socket */priv->sd = socket(PF_PACKET, SOCK_RAW, htons(ETH_P_ALL));if (priv->sd < 0) {printf("Failed to open socket: %d %s\n", errno,strerror(errno));return -errno;
if (priv->local) {Can you put these two blocks of code (if and else) in separate functions and call them from here? This function is too long.
OK
struct sockaddr_in *device;int ret;struct timeval tv;int one = 1;/* Prepare device struct */priv->device = malloc(sizeof(struct sockaddr_in));return -ENOMEM if NULL.
device = priv->device;memset(device, 0, sizeof(struct sockaddr_in));device->sin_family = AF_INET;ret = inet_pton(AF_INET, "127.0.0.1",Is this INADDR_LOOPBACK? Maybe you can just assign it here?
(struct in_addr *)&device->sin_addr.s_addr);if (ret < 0) {printf("Failed to convert address: %d %s\n",
errno,
strerror(errno));return -errno;}/*** Open socket* Since we specify UDP here, any incoming ICMP
packets will
* not be received, so things like ping will not work
on this
* localhost interface.*/priv->sd = socket(AF_INET, SOCK_RAW, IPPROTO_UDP);if (priv->sd < 0) {printf("Failed to open socket: %d %s\n", errno,strerror(errno));return -errno;}/* Allow the receive to timeout after a millisecond */tv.tv_sec = 0;tv.tv_usec = 1000;ret = setsockopt(priv->sd, SOL_SOCKET, SO_RCVTIMEO,
(char *)&tv,
sizeof(struct timeval));if (ret < 0) {printf("Failed to set opt: %d %s\n", errno,strerror(errno));return -errno;}/* Include the UDP/IP headers on send and receive */ret = setsockopt(priv->sd, IPPROTO_IP, IP_HDRINCL, &one,sizeof(one));if (ret < 0) {printf("Failed to set opt: %d %s\n", errno,strerror(errno));return -errno;}priv->local_bind_sd = -1;priv->local_bind_udp_port = 0;} else {struct sockaddr_ll *device;struct packet_mreq mr;/* Prepare device struct */priv->device = malloc(sizeof(struct sockaddr_ll));device = priv->device;memset(device, 0, sizeof(struct sockaddr_ll));device->sll_ifindex = if_nametoindex(ifname);device->sll_family = AF_PACKET;memcpy(device->sll_addr, ethmac, 6);device->sll_halen = htons(6);/* Open socket */priv->sd = socket(PF_PACKET, SOCK_RAW,
htons(ETH_P_ALL));
if (priv->sd < 0) {printf("Failed to open socket: %d %s\n", errno,strerror(errno));return -errno;}/* Bind to the specified interface */setsockopt(priv->sd, SOL_SOCKET, SO_BINDTODEVICE,
ifname,
strlen(ifname) + 1);/* Enable promiscuous mode to receive responses meant
for us */
mr.mr_ifindex = device->sll_ifindex;mr.mr_type = PACKET_MR_PROMISC;setsockopt(priv->sd, SOL_PACKET, PACKET_ADD_MEMBERSHIP,&mr, sizeof(mr)); }
/* Bind to the specified interface */setsockopt(priv->sd, SOL_SOCKET, SO_BINDTODEVICE, ifname,strlen(ifname) + 1);/* Enable promiscuous mode to receive responses meant for us */mr.mr_ifindex = device->sll_ifindex;mr.mr_type = PACKET_MR_PROMISC;setsockopt(priv->sd, SOL_PACKET, PACKET_ADD_MEMBERSHIP,&mr, sizeof(mr));
return 0;}
int sandbox_eth_raw_os_send(void *packet, int length,
const struct eth_sandbox_raw_priv *priv)
struct eth_sandbox_raw_priv *priv){ int retval;
struct udphdr *udph = packet + sizeof(struct iphdr); if (!priv->sd || !priv->device) return -EINVAL;if (priv->local && (priv->local_bind_sd == -1 ||priv->local_bind_udp_port != udph->source))
{
What does this block of code do? Comment?
OK, I'll add a comment. This block of code came about when testing tftp on the localhost interface. When using the RAW AF_INET API, the network stack is still in play responding to incoming traffic based on open "ports". Since it is raw (at the IP layer, no Ethernet) the network stack tells the TFTP server that the port it responded to is closed. This causes the TFTP transfer to be aborted. This block of code inspects the outgoing packet as formulated by the u-boot network stack to determine the source port (that the TFTP server will send packets back to) and opens a typical UDP socket on that port, thus preventing the network stack from sending that ICMP message claiming that the port has no bound socket.
struct iphdr *iph = packet;struct sockaddr_in addr;if (priv->local_bind_sd != -1)close(priv->local_bind_sd);/* A normal UDP socket is required to bind */priv->local_bind_sd = socket(AF_INET, SOCK_DGRAM, 0);if (priv->local_bind_sd < 0) {printf("Failed to open bind sd: %d %s\n", errno,strerror(errno));return -errno;}priv->local_bind_udp_port = udph->source;/*** Bind the UDP port that we intend to use as our
source port
* so that the kernel will not send ICMP port
unreachable
Do you mean return the 'ICMP port unreachable' error?
I guess it is an error. I just thought of it as guidance to the sender.
*/addr.sin_family = AF_INET;addr.sin_port = udph->source;addr.sin_addr.s_addr = iph->saddr;retval = bind(priv->local_bind_sd, &addr, sizeof(addr));if (retval < 0)printf("Failed to bind: %d %s\n", errno,strerror(errno));}retval = sendto(priv->sd, packet, length, 0, (struct sockaddr *)priv->device, sizeof(struct sockaddr_ll));@@ -99,4 +189,10 @@ void sandbox_eth_raw_os_halt(struct
eth_sandbox_raw_priv *priv)
priv->device = NULL; close(priv->sd); priv->sd = -1;
if (priv->local) {if (priv->local_bind_sd != -1)close(priv->local_bind_sd);priv->local_bind_sd = -1;priv->local_bind_udp_port = 0;}} diff --git a/arch/sandbox/dts/sandbox.dts b/arch/sandbox/dts/sandbox.dts index b6762f4..9851bda 100644 --- a/arch/sandbox/dts/sandbox.dts +++ b/arch/sandbox/dts/sandbox.dts @@ -4,6 +4,10 @@ #address-cells = <1>; #size-cells = <0>;
aliases {eth5 = "/eth@90000000";};chosen { stdout-path = "/serial"; };@@ -192,4 +196,10 @@ reg = <0x80000000 0x1000>; host-raw-interface = "eth0"; };
eth@90000000 {compatible = "sandbox,eth-raw";reg = <0x90000000 0x1000>;host-raw-interface = "lo";};}; diff --git a/arch/sandbox/include/asm/eth-raw-os.h
b/arch/sandbox/include/asm/eth-raw-os.h
index d92b72c..44e7050 100644 --- a/arch/sandbox/include/asm/eth-raw-os.h +++ b/arch/sandbox/include/asm/eth-raw-os.h @@ -15,16 +15,24 @@
- sd: socket descriptor - the open socket during a session
- device: struct sockaddr_ll - the host interface packets move to/from
- local: 1 or 0 to select a local interface or not
Should expand this a bit. The local interface is the loopback device....why do you want one over the other?
Not sure what you're getting at here. You want local if you want to talk to a local server (TFTP is what I tested with). You want non-local if you want to talk over a real Ethernet interface to a different machine than the one running sandbox.
- local_bindsd: socket descriptor to prevent the kernel from sending
a message to the server claiming the port is
unreachable*/
- local_bind_udp_port: The UDP port number that we bound to
struct eth_sandbox_raw_priv { int sd; void *device;
int local;int local_bind_sd;unsigned short local_bind_udp_port;};
int sandbox_eth_raw_os_init(const char *ifname, unsigned char *ethmac, struct eth_sandbox_raw_priv *priv); int sandbox_eth_raw_os_send(void *packet, int length,
const struct eth_sandbox_raw_priv *priv);
struct eth_sandbox_raw_priv *priv);int sandbox_eth_raw_os_recv(void *packet, int *length, const struct eth_sandbox_raw_priv *priv); void sandbox_eth_raw_os_halt(struct eth_sandbox_raw_priv *priv); diff --git a/drivers/net/sandbox-raw.c b/drivers/net/sandbox-raw.c index 01b33a9..4b0f836 100644 --- a/drivers/net/sandbox-raw.c +++ b/drivers/net/sandbox-raw.c @@ -15,6 +15,8 @@
DECLARE_GLOBAL_DATA_PTR;
+static int reply_arp; +static IPaddr_t arp_ip;
static int sb_eth_raw_start(struct udevice *dev) { @@ -28,6 +30,12 @@ static int sb_eth_raw_start(struct udevice *dev) interface = fdt_getprop(gd->fdt_blob, dev->of_offset, "host-raw-interface", NULL);
if (strcmp(interface, "lo") == 0) {priv->local = 1;setenv("ipaddr", "127.0.0.1");setenv("serverip", "127.0.0.1");}retval = sandbox_eth_raw_os_init(interface, pdata->enetaddr,
priv);
return retval;@@ -39,19 +47,69 @@ static int sb_eth_raw_send(struct udevice *dev,
void *packet, int length)
debug("eth_sandbox_raw: Send packet %d\n", length);
if (priv->local) {struct ethernet_hdr *eth = packet;if (ntohs(eth->et_protlen) == PROT_ARP) {struct arp_hdr *arp = packet + ETHER_HDR_SIZE;/*** localhost works on a higher level API in
Linux than
nit: higher-level
* ARP packets, so fake itfake what?
Fake an ARP response. The u-boot network stack is sending an ARP request (to find the MAC address to address the actual packet to) and requires an ARP response to continue. Since this is the localhost interface, there is no Etherent level traffic at all, so there is no way to send an ARP request or to get a response. For this reason we fake the response to make the u-boot network stack happy.
*/arp_ip = NetReadIP(&arp->ar_tpa);reply_arp = 1;return 0;}packet += ETHER_HDR_SIZE;length -= ETHER_HDR_SIZE;} return sandbox_eth_raw_os_send(packet, length, priv);}
static int sb_eth_raw_recv(struct udevice *dev) {
struct eth_pdata *pdata = dev_get_platdata(dev); struct eth_sandbox_raw_priv *priv = dev_get_priv(dev);
int retval;
int retval = 0; uchar buffer[PKTSIZE]; int length;
retval = sandbox_eth_raw_os_recv(buffer, &length, priv);
if (reply_arp) {struct arp_hdr *arp = (void *)buffer + ETHER_HDR_SIZE;/* Formulate a fake ARP */arp->ar_hrd = htons(ARP_ETHER);arp->ar_pro = htons(PROT_IP);arp->ar_hln = ARP_HLEN;arp->ar_pln = ARP_PLEN;arp->ar_op = htons(ARPOP_REPLY);/* Any non-zero MAC address will work */memset(&arp->ar_sha, 0x01, ARP_HLEN);/* Use whatever IP we were looking for (always
127.0.0.1?) */
NetWriteIP(&arp->ar_spa, arp_ip);memcpy(&arp->ar_tha, pdata->enetaddr, ARP_HLEN);NetWriteIP(&arp->ar_tpa, NetOurIP);length = ARP_HDR_SIZE;} else {/* If local, the Ethernet header won't be included;
skip it */
uchar *pktptr = priv->local ? buffer + ETHER_HDR_SIZE :
buffer;
retval = sandbox_eth_raw_os_recv(pktptr, &length, priv);} if (!retval && length) {if (priv->local) {struct ethernet_hdr *eth = (void *)buffer;/* Fill in enough of the missing Ethernet
header */
memcpy(eth->et_dest, pdata->enetaddr, ARP_HLEN);memset(eth->et_src, 0x01, ARP_HLEN);eth->et_protlen = htons(reply_arp ? PROT_ARP :
PROT_IP);
reply_arp = 0;length += ETHER_HDR_SIZE;}debug("eth_sandbox_raw: received packet %d\n", length); NetReceive(buffer, length);-- 1.7.11.5
Regards, Simon _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de http://lists.denx.de/mailman/listinfo/u-boot