On Thu, Sep 04, 2014 at 06:07:32AM -0600, Eric Blake wrote: > On 09/04/2014 02:58 AM, Stefan Hajnoczi wrote: > > On-disk structures should be marked packed so the compiler does not > > insert padding for field alignment. Padding should be explicit so > > on-disk layout is obvious and we don't rely on the architecture-specific > > ABI for alignment rules. > > > > The pahole(1) diff shows that the padding is now explicit and offsets > > are unchanged: > > > > char backing_file[1024]; /* 8 1024 */ > > /* --- cacheline 16 boundary (1024 bytes) was 8 bytes ago --- */ > > int32_t mtime; /* 1032 4 */ > > - > > - /* XXX 4 bytes hole, try to pack */ > > - > > + uint32_t padding; /* 1036 4 */ > > uint64_t size; /* 1040 8 */ > > Was a 32-bit build also inserting this padding, or do we have historical > differences where 32-bit and 64-bit cow files are actually different, > and we may need to be prepared to parse files from both sources?
Good point. Let's not merge this patch since it breaks 32-bit hosts. The fact that no one hit problems when exchanging files between 32-bit and 64-bit machines shows that the cow format is rarely used. At this point we have 2 different formats: one without padding (i386-style) and one with padding (x86_64-style). The chance of more variants is small but who knows, maybe some other host architecture ABI has yet another alignment rule for uint64_t. I'd like to git rm block/cow.c but I suppose the backwards-compatible thing to do is to introduce subformats to support both variants. Opinions? Stefan
pgpMMqvimdAFZ.pgp
Description: PGP signature
