Re: [Qemu-devel] [PATCH 4/5] block: cow - used QEMU_PACKED for on-disk structures

[Markus Armbruster]

Jeff Cody <address@hidden> writes:

> On Thu, Sep 19, 2013 at 12:01:24PM -0700, Richard Henderson wrote:
>> On 09/19/2013 11:43 AM, Jeff Cody wrote:
>> > cow_header_v2 is read and written directly from the image file
>> > with bdrv_pread()/bdrv_pwrite(), and as such should be packed to
>> > avoid unintentional padding.
>> > 
>> > Also change struct cow_header_v2 to a typedef, and some minor
>> > code style changes to keep checkpatch.pl happy.
>> > 
>> > Signed-off-by: Jeff Cody <address@hidden>
>> > ---
>> >  block/cow.c | 21 +++++++++++----------
>> >  1 file changed, 11 insertions(+), 10 deletions(-)
>> > 
>> > diff --git a/block/cow.c b/block/cow.c
>> > index 909c3e7..9c15afb 100644
>> > --- a/block/cow.c
>> > +++ b/block/cow.c
>> > @@ -32,14 +32,14 @@
>> >  #define COW_MAGIC 0x4f4f4f4d  /* MOOO */
>> >  #define COW_VERSION 2
>> >  
>> > -struct cow_header_v2 {
>> > +typedef struct QEMU_PACKED cow_header_v2 {
>> >      uint32_t magic;
>> >      uint32_t version;
>> >      char backing_file[1024];
>> >      int32_t mtime;
>> >      uint64_t size;
>> >      uint32_t sectorsize;
>> > -};
>> > +} COWHeaderV2;
>> 
>> This changes the layout of this struct.  In particular, there's padding
>> (depending on the host) between mtime and size.
>> 
>
> You are right, and that poses a problem for this patch.
>
>> I don't know what the right solution is: COWHeaderV3 with the bug fix, 
>> leaving
>> V2 alone; adding an int32_t dummy there where the padding was; nothing,
>> considering the padding to be gone a good thing.
>> 
>
> I'm not sure either.  I don't think the right thing is to take the
> patch as-is, because that will likely break a lot of existing COW
> images (I just checked, and on x86_64, it is 1056 bytes unpacked, or
> 1048 bytes packed).
>
> Unfortunately, this means that theoretically, image files with this
> format may not be portable, depending on the hosts' compiler and
> alignment.  In reality, it likely is not a problem.
>
> I'll drop this one for v2.

Possible solutions:

* Declare format "cow" non-portable.  To move a cow to another system,
  you have to convert to a portable format.

* Keep using the non-portable header.  When opening an existing image,
  guess which of the two header variants it got: the padding should be
  zero, size and sectorsize sane.  Perhaps provide an option to overrule
  the guess.

Who's still using format "cow"?