Hi,

Sorry for the formatting.

On Tue, 23 Apr 2019, 18:34 Stefano Stabellini, <sstabellini@kernel.org>
wrote:

> On Tue, 23 Apr 2019, Julien Grall wrote:
> > Hi Stefano,
> >
> > On 4/22/19 11:42 PM, Stefano Stabellini wrote:
> > > On Tue, 26 Feb 2019, Julien Grall wrote:
> > > I am not sure about the suggestion of re-using the libfdt concept of
> > > "mem_rsv", which is meant to be for the old /memreserve/. Today, libfdt
> > > (at least our version of it) is not able to parse the new
> > > reserved-memory bindings. I don't think it is a good idea to modify
> > > libfdt for that. Also, the way we want to handle the old memreserve
> > > regions is quite different from the way we want to handle
> > > reserved-memory, right? I cannot see a way to improve this code using
> > > mem_rsv at the moment.
> >
> > I didn't mean to extend mem_rsv in libfdt but extend consider_modules and
> > dt_unreserved_regions to deal with /reserved-memory. Otherwise you
> > may miss some cases (for instance you left out discard_initial_modules).
> >
> > By extending those two functions you don't have to teach everyone how to
> skip
> > /reserved-memory.
>
> I think I get your point now. Although I don't think it should be
> correct for a bootloader to use a reserved-memory area to store a boot
> module, I wouldn't be suprised if that happens, so it is better to be
> prepared and extend dt_unreserved_regions. I'll do that.
>

Why wouldn't this be correct? It is nothing different /mem-reserve.


> However, we would still need something like check_reserved_memory,
> because we don't want setup_xenheap_mappings to be called on the
> reserved-memory area (or a memory region including the reserved memory
> area) in setup_mm. I don't think we can get away without it, but I can
> simplify it.
>

Hmmm, setup_xenheap_mappings is only doing the mapping in page-tables
allowing direct access in Xen. Are you worried of the memory attributes to
be different in Xen?

This would makes sense however setup_xenheap_mappings may still map the
reserved-memory because it is using 1G mapping... This is pretty wrong and
I have patches that should help to fix it.

 Also if you are concerned with /reserved-memory, then it should also be
fixed for /mem-reserve as they are not different. However, this may break
free_initmem as because we try to give back page to xenheap even if they
are reserved.

 The memory management is quite a mess today. I hope to make it better with
my upcoming series.

Cheers,