Re: [Qemu-ppc] [PATCH v4 2/3] target/ppc: Add GDB callbacks for SPRs

[Alexey Kardashevskiy]

On 01/02/2019 08:57, Fabiano Rosas wrote:
> Alexey Kardashevskiy <address@hidden> writes:
> 
>> On 31/01/2019 03:30, Fabiano Rosas wrote:
>>> Alexey Kardashevskiy <address@hidden> writes:
>>>
>>>>
>>>> but this is a register which does not have endianness, the endianness
>>>> appears here because the interface between gdb and qemu is
>>>> uint8_t*==bytestream but this interface should have fixed endianness
>>>> imho (now it is bigendian afaict).
>>>>
>>>> Something is not right here...
>>>
>>> Having a fixed endianness would not work because GDB have no way of
>>> knowing how to represent what comes from the remote end.
>>
>> It definitely would. A register is stored as "unsigned long" in QEMU and
>> all gdb has to do is printf("%lx") and that is it. 
> 
> OK, but something is not clear to me. Even if GDB just printf("%lx") the
> value, we would still have to bswap when the host is LE, right?

Not for %lx, this should just print a correct value.

> QEMU BE:
>  (gdb) x/8xb &env->spr[287]
>  0x11391760: 0x00    0x00    0x00    0x00    0x00    0x4e    0x12    0x00
> 
> QEMU LE:
>  (gdb) x/8xb &env->spr[287]
>  0x7ffff5bd98c0: 0x01    0x02    0x4b    0x00    0x00    0x00    0x00    0x00
> 
>> The problem is that
>> we want to pass it as a byte stream from the gdb_read_register() hook
>> all the way to gdb and for some reason gdb does not define endianness of
>> that stream but rather tries guessing the endianness which is broken.
> 
> GDB does define the endianness of the stream (in a way):
> 
>  "Each byte of register data is described by two hex digits. The bytes
>  with the register are transmitted in target byte order."
> 
> https://sourceware.org/gdb/current/onlinedocs/gdb/Packets.html#Packets


Target byte order changes all the time so we need a endian-agnostic way
of knowing the current endianness.


> 
>> Today I was debugging rtas/clientinterface calls which a little endian
>> kernel makes and these calls need to switch to the big endian first. And
>> gdb goes nuts when this switch happens (note that I did not give an ELF
>> to gdb this time so it picked LE itself). Even if it could fetch the
>> endianness from QEMU, it would fail as it is an LE bit in MSR which is a
>> register which is parsed according to the gdb's idea of endianness :)
> 
> I think it would be possible to define another packet for the remote
> protocol that informs the endianness explicitly at each time the guest
> stops. If you provide more info on how to reproduce that issue I could
> put in my list or go bug GDB folks about it.
> 
>>> It will
>>> always check the target endianness before printing a value, even if it
>>> refers to a register:
>>>
>>> https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=blob;f=gdb/findvar.c;hb=HEAD#l49
>>>
>>> So in our case the contents of mem_buf need to match both the guest
>>> endianness *and* what GDB has set for 'show endian' because it will
>>> detect it automatically from the ELF. If it guesses incorrectly because
>>> there is no ELF, we need to use the 'set endian' command.
>>>
>>> By the way, this is already the behavior for the registers that are
>>> already implemented (e.g. $msr). Here's the commit that introduced
>>> that:
>>>
>>> https://git.qemu.org/?p=qemu.git;a=commitdiff;h=8a286ce4502356ce0b97a2424a2cb7
>>>
>>> Now, what might be a source of confusion here is the fact that we
>>> *always* do a bswap when the host is LE because QEMU thinks that the ppc
>>> guest is always BE. That requires the maybe_bswap function to make
>>> things right in the end.
>>>
>>> What I could do is try to improve this by only swapping when the
>>> guest's actual endianness (msr_le) is different from the host's.
>>
>> The bytestream for registers should have fixed endianness. But looking
>> at the gdb code makes me think it is not going to happen :(
> 
> Yes, I can't think of a way to fix that without changing the way GDB
> exhibits the values or the remote protocol.
> 
> May I proceed with this series as it is with the bswaps?
> 
>>> That
>>> is not entirely within the scope of this patch, though.
>>
>> True. But since you are touching this, may be you could fix gdb too :)
>>
>> Does gdb tell QEMU about what endianness it thinks that QEMU is using?
>> Or can it read it from QEMU? I cannot easily spot this in QEMU...
> 
> GDB currently does not tell and does not ask about endianness. So I
> believe there is room for improvement here. I could not find it in the
> documentation but I think that GDB supports file transfers and it asks
> for the ELF in some scenarios. This approach could be one way of
> informing it about the endianness, although it has its own shortcomings.


There is no ELF in my scenario really. What I did was:

1. hack qemu to not load slof.bin but load vmlinux instead and changed
starting pc to where I loaded the kernel (I did not have to but it is a
lot easier to ditch slof and set breakpoint in gdb before starting the
guest).

2. start a guest, connect external gdb and observe garbage in $pc and
disassembly until the guest switched to LE.

3. trace till here:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/kernel/entry_64.S?h=v4.20#n1299
and set breakpoint to whatever $r4 was at this point; $r4 points to RTAS
which executes in BE as we remove LE bit from SRR1 (lines 1296..1298).

When we stop at this new breakpoint - we get same biteswapped garbage as
in step2. Well, this is SLOF and there is an ELF somewhere but do we
really want to load ELF and switch endianness every time the guest
changes it and jumps to different chunks of code? :)


It also seems like we cannot really debug slof via the gdb stub as it is
big endian so it will be executing instructions but the UI will look
byteswapped.


-- 
Alexey