1 files changed, 83 insertions, 0 deletions
diff --git a/results/classifier/gemma3:12b/network/2553 b/results/classifier/gemma3:12b/network/2553
new file mode 100644
index 000000000..5d0c7c205
--- /dev/null
+++ b/results/classifier/gemma3:12b/network/2553
@@ -0,0 +1,83 @@
+
+Joining IP multicast fails when emulating 64-bit Linux
+Description of problem:
+I have some code that joins IP multicast groups and I'd like to use QEMU to test it on big-endian and/or 32-bit platforms. But when I compile it for 64-bit big-endian platforms (e.g. PowerPC64) and run it under QEMU user-mode emulation, the setsockopt(IP_ADD_MEMBERSHIP) call fails with ENODEV.
+
+This appears to refer to the imr_ifindex ("interface index") field in struct ip_mreqn not being valid, which in turn appears to be because it's not being correctly marshalled from the binary under emulation, to the host's *actual* setsockopt system call.
+
+I *think* this may be because linux-user/syscall_defs.h (https://github.com/qemu/qemu/blob/master/linux-user/syscall_defs.h) contains the following at line 210:
+ 
+```
+struct target_ip_mreqn {
+    struct target_in_addr imr_multiaddr;
+    struct target_in_addr imr_address;
+    abi_long imr_ifindex;
+};
+```
+
+but the actual Linux ip_mreqn has imr_ifindex as an int (32-bit everywhere) not a long (64-bit on PPC64); the size of this structure is 12 on all Linux platforms.
+
+I opted to submit an issue instead of just patching it, in case there was some wider context I hadn't seen?
+Steps to reproduce:
+1. take the following C program (distilled from a larger program):
+
+```
+#include <sys/types.h>
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <arpa/inet.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+int main(int argc, char *argv[])
+{
+    int fd = socket(AF_INET, SOCK_DGRAM, 0);
+    if (fd < 0) {
+        perror("socket");
+        return 1;
+    }
+
+    struct ip_mreqn mreq;
+    mreq.imr_multiaddr.s_addr = inet_addr("239.255.255.250");
+    mreq.imr_address.s_addr = htonl(INADDR_ANY);
+    mreq.imr_ifindex = 1;
+    int size = sizeof(mreq);
+    printf("size=%u\n", size);
+    if (setsockopt(fd, IPPROTO_IP, IP_ADD_MEMBERSHIP,
+                   (char*) &mreq, sizeof(mreq)) < 0) {
+        perror("setsockopt");
+        return 1;
+    }
+    printf("OK\n");
+    return 0;
+}
+```
+
+2. confirm it works compiled native on amd64/x86_64:
+
+```
+[peter@amd64 misc]$ gcc mcast.c -o mcast
+[peter@amd64 misc]$ ./mcast
+size=12
+OK
+```
+
+3. watch it *not* work emulated:
+
+```
+[peter@amd64 misc]$ powerpc64-linux-gnu-gcc mcast.c -o mcast.ppc64
+[peter@amd64 misc]$ QEMU_LD_PREFIX=/usr/powerpc64-linux-gnu qemu-ppc64 ./mcast.ppc64 
+size=12
+setsockopt: No such device
+```
+Additional information:
+If the target_ip_mreqn issue is real, the following code in syscall.c helped conceal it:
+
+            if (optlen < sizeof (struct target_ip_mreq) ||
+                optlen > sizeof (struct target_ip_mreqn)) {
+                return -TARGET_EINVAL;
+            }
+
+Should this instead be testing for size equal to target_ip_mreq or equal to target_ip_mreqn, not anywhere in between? in this case target_ip_mreq is 8 bytes, target_ip_mreqn is 16 bytes, but optlen is 12. The end result is that QEMU passes 4 bytes of uninitialised stack memory as imr_ifindex!
+
+The actual kernel behaviour appears to be: smaller than ip_mreq, EINVAL; between ip_mreq and ip_mreqn, silently treat as ip_mreq; larger or equal to ip_mreqn, silently treat as ip_mreqn. see https://github.com/torvalds/linux/blob/b31c4492884252a8360f312a0ac2049349ddf603/net/ipv4/ip_sockglue.c#L1234