Re: [Qemu-arm] recvfrom error with linux-user emulating armhf on aarch64

[Philippe Mathieu-Daudé]

Hi Guido,

On 01/23/2018 11:42 AM, Philippe Mathieu-Daudé wrote:
> Hi Guido, Laurent,
> 
> On 01/23/2018 09:05 AM, Guido Günther wrote:
>> Hi,
>> Running the attached program⁰ in a armhf chroot on a arm64 host¹ like:
>>
>>     chroot . /usr/bin/qemu-arm-static tmp/nl-bad-addr
>>
>> fails with "Bad address" when invoking audit_log_acct_message. strace looks 
>> like:
>>
>>   [..snip..]
>>   31572 socket(AF_NETLINK, SOCK_RAW, NETLINK_AUDIT) = 3
>>   ...
>>   31572 sendto(3, {{len=124, type=0x454 /* NLMSG_??? */, 
>> flags=NLM_F_REQUEST|NLM_F_ACK, seq=1, pid=0}, "op=test:message acct=\"?\" 
>> exe=\"/tmp/nl-bad-addr\" hostname=localhost addr=? terminal=/dev/pts/2 
>> res=success\0\0\0"}, 124, 0, 0xfffffa3897d0, 0) = 124
> 
> the do_sendto() seems to correctly handle the u64 target_addr
> 
>>   31572 ppoll([{fd=3, events=POLLIN}], 1, {tv_sec=0, tv_nsec=500000000}, 
>> NULL, 0) = 1 ([{fd=3, revents=POLLIN}], left {tv_sec=0, tv_nsec=499993180})
>>   31572 recvfrom(3, 0x112a50eb4, 8988, MSG_PEEK|MSG_DONTWAIT, 
>> 0xfffffa3897e0, 0x42) = -1 EFAULT (Bad address)
> 
> but the recvfrom() code is:
> 
> static abi_long do_recvfrom(int fd, abi_ulong msg, size_t len,
>                             int flags,
>                             abi_ulong target_addr,
>                             abi_ulong target_addrlen)
> {
>     [...]
>     if (target_addr) {
>         if (get_user_u32(addrlen, target_addrlen)) {
>             ret = -TARGET_EFAULT;
>             goto fail;
>         }
> 
> so current code won't work on 64-bit host, and we are probably lucky it
> works on x86_64.
> 
>>   [..snip..]
>>
>> And it fails in
>>
>>    ret = get_errno(safe_recvfrom(fd, host_msg, len, flags,
>>                                       addr, &addrlen));
>>
>> in linux-user/syscall.c:do_recvfrom but I have no idea yet what causes this.
>>
>> However the same works without problems when the host is a x86_64 box.
>> Any pointers in debugging this further would be appreciated.

I proposed a fix for this, do you mind testing it?

http://lists.nongnu.org/archive/html/qemu-devel/2018-01/msg05670.html

Regards,

Phil.

> 
> Laurent, in do_syscall() should we swap the pid back?
> 
> @@ -10628,4 +10628,5 @@ abi_long do_syscall(...
>          target_header->version = tswap32(header.version);
> +        target_header->pid = tswap32(header.pid);
>          unlock_user_struct(target_header, arg1, 1);
> 
>>
>> The issue is also present in qemu master. The overall goal is to get
>>
>>    qemu-debootstrap --arch armhf buster chroot
>>
>> to succeed on an aarch64 host as it does on x86_64.
>> Cheers,
>>  -- Guido
>>
>> ⁹: which is a stripped down testcase for /usr/bin/chfn failing
>> ¹: the scaleway server this runs on doesn't support armhf.
>>
>

signature.asc
Description: OpenPGP digital signature