Re: [PATCH v5 10/10] x86, efi: EFI boot stub support

[Matt Fleming <matt@console-pimps.org>]

On Thu, 2011-11-24 at 15:43 +0100, Maarten Lankhorst wrote:
> Hey Matt,
> 
> On 11/23/2011 08:27 PM, Matt Fleming wrote:
> 
> > On Wed, 2011-11-23 at 01:44 +0100, Maarten Lankhorst wrote:
> >> When I tested this with v3.2-rc2 it didn't boot, it hung before it
> >> initialized the kernel.
> >> Without initrd it works fine, though.
> > Bah, so this change actually makes booting worse? You said before that
> > you almost made it to userspace but this seems to hang much earlier now.
> > Is that correct?
> >
> > ... back to the drawing board.
> I was looking at why grub2 could boot, seems to be it reads in chunks of
> 256 kilobytes. I seem to be able to get it to boot with chunks of 4 mb
> as well, but didn't test beyond that.
> 
> So the fix is to simply read the file in parts, otherwise efi hangs..
> As a nice side effect, short reads are also handled, but the efi
> firmware seems to choke over huge reads and dies.

Urrggghh.....

Thanks a lot for diagnosing this, Maarten! The copy of grub that I have
(fedora's version) looks like it uses the disk_io protocol to read from
disks, I wonder if there's a reason for this. I Cc'd Peter Jones in case
he's seen anything like this before.

Aha, read_file() in the elilo source contains this helpful comment,

	/*                                                                      
         * We load by chunks rather than a single big read because              
         * early versions of EFI had troubles loading files                     
         * from floppies in a single big request.  Breaking                     
         * the read down into chunks of 4KB fixed that                          
         * problem. While this problem has been fixed, we still prefer          
         * this method because it tells us whether or not we're making          
         * forward progress.                                                    
         */

(but reading the code it's clear it uses a chunk size of 16KB, not 4KB)

So, it looks like it's not just reading from floppies that hits this bug
in the firmware.

Maarten, could you send a patch against the 'x86/efi-stub' branch at
git://git.kernel.org/pub/scm/linux/kernel/git/mfleming/linux.git that
splits the read up into 16KB chunks and explains,

	- why we need to do the reading in chunks
	- why we chose the 16KB chunk size (because elilo said so)

That way, the bug fix will be properly documented and attributed to you.

-- 
Matt Fleming, Intel Open Source Technology Center