Re: Commiting files larger than 4 GB on Windows

> On Wed, Mar 15, 2017 at 11:59:52AM -0400, Jeff King wrote:
> 
>> I agree that detecting the situation in the meantime is a good idea.
>> The patch above probably handles the bulk-checkin code path, I'd guess.
>> It might be nice to have similar checks in other places, too:
>>
>>   - when reading from an existing packfile
>>
>>     Looks like we may already have such a check in
>>     unpack_object_header_buffer().
>>
>>   - when taking in new objects via index-pack or unpack-objects (to
>>     catch a fetch of a too-big object)
>>
>>     I think index-pack.c:unpack_raw_entry() would want a similar check
>>     to what is in unpack_object_header_buffer().
> 
> Here are the results of a few quick experiments using two versions of
> git, one built for 32-bit and one for 64-bit:
> 
>   $ git init
>   $ dd if=/dev/zero of=foo.zero bs=1M count=4097
>   $ git32 add foo.zero
>   fatal: Cannot handle files this big
> 
> That comes from the xsize_t() wrapper. I guess it wouldn't trigger on
> Windows, though, because it is measuring size_t, not "unsigned long" (on
> my 32-bit build they are the same, of course).
> 
>   $ git64 add foo.zero
>   $ git32 cat-file blob :foo.zero
>   error: bad object header
>   fatal: packed object df6f032f301d1ce40477eefa505f2fac1de5e243 (stored in .git/objects/pack/pack-57d422f19904e9651bec43d10b7a9cd882de48ac.pack) is corrupt
> 
> So we notice, which is good. This is the message from
> unpack_object_header_buffer(). It might be worth improving the error
> message to mention the integer overflow.
> 
> And here's what index-pack looks like:
> 
>   $ git32 index-pack --stdin <.git/objects/pack/*.pack
>   fatal: pack has bad object at offset 12: inflate returned -5
> 
> It's good that we notice, but the error message isn't great. What
> happens is that we overflow the size integer, allocate a too-small
> buffer, and then zlib complains when we run out of buffer but there's
> still content to inflate. We probably ought to notice the integer
> overflow in the first place and complain there.