Re: [PATCH] Wrong unpacked structure for epoll_event on qemu-or1k (openrisc user-space)

[Laurent Vivier]

Le 18/07/2023 à 17:06, Luca Bonissi a écrit :
> On 18/07/23 16:40, Peter Maydell wrote:
> > Hi; thanks for this patch. Unfortunately we need patches
> > to include a Signed-off-by: line that says you're legally
> > OK with it being contributed to QEMU, or we can't take them.
>
> Sorry for the missing "signed-off-by" line, adding it just now:
>
> ==============
> The or1k epoll_event structure - unlike other architectures - is packed, so we need to define it as 
> packed in qemu-user, otherwise it leads to infinite loop due to missing file descriptor in the 
> returned data:
>
>
> Signed-off-by: Luca Bonissi <qemu@bonslack.org>
> ---
>
> diff -up a/linux-user/syscall_defs.h b/linux-user/syscall_defs.h
> --- a/linux-user/syscall_defs.h    2023-03-27 15:41:42.000000000 +0200
> +++ b/linux-user/syscall_defs.h    2023-06-30 17:29:39.034322213 +0200
> @@ -2714,7 +2709,7 @@
>   #define FUTEX_CMD_MASK          ~(FUTEX_PRIVATE_FLAG | FUTEX_CLOCK_REALTIME)
>
>   #ifdef CONFIG_EPOLL
> -#if defined(TARGET_X86_64)
> +#if defined(TARGET_X86_64) || defined(TARGET_OPENRISC)
>   #define TARGET_EPOLL_PACKED QEMU_PACKED
>   #else
>   #define TARGET_EPOLL_PACKED

According to linux/glibc sourced, epoll is only packed for x86_64.

Did you try to check the alignment of the structure with gdb of a C program using offsetof() in an 
openrisc VM or linux-user container?

Perhaps the default alignment of long is not correctly defined in qemu for 
openrisc?

You can check with:

int main(void)
{
        printf("alignof(short) %ld\n", __alignof__(short));
        printf("alignof(int) %ld\n", __alignof__(int));
        printf("alignof(long) %ld\n", __alignof__(long));
        printf("alignof(long long) %ld\n", __alignof__(long long));
}

See include/exec/user/abitypes.h to update the value.

Thanks,
Laurent