1
2
3
4
5
6
7
8
9
10
11
12
13
|
<think>
Okay, let's see. The user is reporting an issue with netlink being broken on big-endian MIPS when using qemu-user. The problem causes 'ip route' to hang, and the strace shows that the sendto call has a byteswapped len field. They mention that removing fd_trans_unregister in the NR_socket syscall fixes it, but they're unsure why it was there.
Hmm, the key part here is the syscall handling for NR_socket. The code in do_syscall for TARGET_NR_socket calls do_socket and then immediately unregisters the fd translation. The user modified that by commenting out the fd_trans_unregister line, which fixed the issue.
So, the problem seems to be related to how file descriptors are being translated or managed during the socket syscall. The fact that the len is byteswapped suggests that there's an endianness issue, possibly due to incorrect handling of the data structures involved in the syscall. Since the fd translation is being unregistered right after the socket is created, maybe the translation is needed for subsequent operations on that socket, like sending or receiving data. Without the translation, the kernel might not be handling the endianness correctly, leading to the byteswapped value.
The root cause likely lies in the syscall handling for socket, specifically the management of file descriptor translations. The fd_trans_unregister is part of the QEMU's mechanism to handle file descriptors between the target (MIPS) and host (probably x86_64). If the translation is unregistered too early, the subsequent syscalls (like sendto, recvmsg) might not have the correct translation, leading to incorrect data being passed, such as the byteswapped len field.
Therefore, this is a syscall-related issue. The problem arises from how the syscall is processed, specifically the handling of file descriptor translations after the socket is created. The error isn't in the instruction set or runtime environment but in the syscall's implementation in QEMU.
</think>
syscall
|