Re: Fixed Mirage

[Anil Madhavapeddy <anil@xxxxxxxxxx>]

On 12 Jul 2013, at 00:56, "Vincent B." <vb@xxxxxxxxxxxxxx> wrote:
> On 11/07/2013 23:48, Anil Madhavapeddy wrote:
>> Hm, this is still pretty broken and crashes immediately on boot with
>> master/:
>> 
>> https://github.com/mirage/mirage-platform/commit/ab88ef636fe3ff06c8638d5ea765ad25bf9610e2
> 
> Whas crashes exactly ? the basic example with DHCP was working well for
> me when I commited those changes…

Are you sure you tested with the changes you committed or with a slightly 
earlier version?

>> Why did you remove the zeroing of the Io_pages?  That breaks
>> everything and the domain crashes as so> on as any shared rings are
>> used, since there is garbage on the ring at start.  I added a
>> memset(block, 0, len) into alloc_pages, and the domain now boots ok,
>> but I still see a lot of packet loss.  Investigating...
> 
> But I added the zeroing in the first place!!! Previously there was no zeroing 
> at all, I then added it but removed it after I observed that the pages 
> returned by memalign were actually always zeroed.

The bug is in a changeset marked "cosmetics" :-)
https://github.com/mirage/mirage-platform/commit/41f26e337beeb6799f9e04c304480ecca98f0fef

You removed this hunk from allocate_ring:

-  for i = 0 to Cstruct.len x - 1 do
-    Cstruct.set_uint8 x i 0
-  done;

which zeros the ring, and Io_pages are also no longer zeroed since you reverted 
that earlier.  One or the other is needed.  I'd prefer that pages not be zeroed 
by default, since only the rings really need it.

> I had the impression that Xen would always zero the pages in order for 
> domains not to be able to see other domains memory…

Right, Xen does that with physical memory, but remember that we have a memory 
slab allocator ("alloc_page" from the xen/runtime/kernel) that is recycling the 
OS pages for the OCaml runtime to use.  That will only zero pages out if you 
specifically do so.

-anil