Re: [PATCH v3] migration: Support gtree migration

[Auger Eric <eric.auger@redhat.com>]

Hi peter,

On 10/10/19 2:35 PM, Peter Xu wrote:
> On Thu, Oct 10, 2019 at 02:11:46PM +0200, Auger Eric wrote:
>>>>> Also, should we avoid using UINT in all cases?  But of course if we
>>>>> start to use VMSTATE_UINT32_V then we don't have this issue.
>>>> Depending on the clarification of above point, maybe I can rename
>>>> VMSTATE_GTREE_DIRECT_KEY_V into VMSTATE_GTREE_DIRECT_UINT_KEY_V
>>>>
>>>> direct keys seem to be more common for hash tables actually.
>>>> https://developer.gnome.org/glib/stable/glib-Hash-Tables.html#g-hash-table-new-full
>>>>
>>>> There are stand conversion macros to/from int, uint, size
>>>> https://developer.gnome.org/glib/stable/glib-Type-Conversion-Macros.html
>>>
>>> Yeh it's fine to use direct keys.  Though my question was more about
>>> "unsigned int" - here when we put, we cast a pointer into unsigned
>>> int, which can be 2/4 bytes, IIUC.  I'm thinking whether at least we
>>> should use direct cast (e.g., (uint32_t)ptr) to replace
>>> GPOINTER_TO_UINT() to make sure it's 4 bytes.  Futher, maybe we should
>>> start with uint64_t here in the migration stream, otherwise we should
>>> probably drop the high 32 bits if we migrate a gtree whose key is 64
>>> bits long (and since we're working with migration we won't be able to
>>> change that in the future for compatibility reasons...).
>>>
>>> Summary:
>>>
>>> Maybe we can replace:
>>>
>>>     qemu_put_be32(f, GPOINTER_TO_UINT(key)); /* direct key */
>>>
>>> To:
>>>
>>>     qemu_put_be64(f, (uint64_t)key); /* direct key */
>>>
>>> And apply similar thing to get() side?
>>
>> This was my first idea as well but I got stuck with a mingw compilation
>> issues if I remember correctly, trying to cast pointers to a wrong sized
>> uint. This got removed by using the GPOINTER_TO_UINT conversion functions.
> 
> #define GPOINTER_TO_UINT(p) ((guint) (gulong) (p))
> 
> Could "(uint64_t)(uintptr_t)pointer" do the work?
the problem rather is on the get side, when you cast the uint64_t value
into the pointer. it does not compile with

make docker-test-mingw@fedora


/tmp/qemu-test/src/migration/vmstate-types.c:800:19: error: cast to
pointer from integer of different size [-Werror=int-to-pointer-cast]
             key = (void *)(uint64_t)qemu_get_be64(f);

I would be tempted to rename the macro to emphasize the key is an
uint32. It should cover most of the use cases, no?

Thanks

Eric


> 
> Thanks,
>