[PATCH v2 4/5] ARM: init: add support for reserved memory defined by device tree

[Grant Likely]

On Tue, 11 Feb 2014 11:52:42 +0100, Marek Szyprowski <m.szyprowski at samsung.com> wrote:
> 
> On 2014-02-10 22:59, Grant Likely wrote:
> > On Thu, 06 Feb 2014 14:26:13 +0100, Marek Szyprowski <m.szyprowski at samsung.com> wrote:
> > > Hello,
> > >
> > > On 2014-02-05 11:15, Grant Likely wrote:
> > > > On Tue, 04 Feb 2014 13:09:32 +0100, Marek Szyprowski <m.szyprowski at samsung.com> wrote:
> > > > > Enable reserved memory initialization from device tree.
> > > > >
> > > > > Cc: Benjamin Herrenschmidt <benh at kernel.crashing.org>
> > > > > Cc: Laura Abbott <lauraa at codeaurora.org>
> > > > > Signed-off-by: Marek Szyprowski <m.szyprowski at samsung.com>
> > > > > ---
> > > > >  arch/arm/mm/init.c |    3 +++
> > > > >  1 file changed, 3 insertions(+)
> > > > >
> > > > > diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> > > > > index 804d61566a53..ebafdb479410 100644
> > > > > --- a/arch/arm/mm/init.c
> > > > > +++ b/arch/arm/mm/init.c
> > > > > @@ -17,6 +17,7 @@
> > > > >  #include <linux/nodemask.h>
> > > > >  #include <linux/initrd.h>
> > > > >  #include <linux/of_fdt.h>
> > > > > +#include <linux/of_reserved_mem.h>
> > > > >  #include <linux/highmem.h>
> > > > >  #include <linux/gfp.h>
> > > > >  #include <linux/memblock.h>
> > > > > @@ -323,6 +324,8 @@ void __init arm_memblock_init(struct meminfo *mi,
> > > > >  	if (mdesc->reserve)
> > > > >  		mdesc->reserve();
> > > > >
> > > > > +	early_init_dt_scan_reserved_mem();
> > > > > +
> > > >
> > > > The new binding is being made fundamental. If the reserved-memory node
> > > > is present, then it needs to be honored, even if the kernel doesn't know
> > > > how to use the regions. Therefore, This needs to be unconditional for
> > > > all architectures. The hook should be called in early_init_dt_scan()
> > > > (drivers/of/fdt.c) immediately after the early_init_dt_scan_memory()
> > > > hook.
> > >
> > > In theory this will be the best solution, but it practice there is a
> > > problem. early_init_dt_scan() is called as the first function from kernel
> > > booting code. That time there is no memory yet added to the system, so it
> > > would be really hard to reserve anything. Memory nodes are being added
> > > later either with memblock_add() or by some other arch specific way.
> >
> > Hmmm, depends on the architecture. On ARM the memory is loaded into the
> > meminfo structure first, and it isn't until arm_memblock_init() that
> > memblock_add() gets called on all the regions. Some architectures do the
> > memblock_add() directly from early_init_dt_add_memory_arch() function.
> >
> > The default early_init_dt_add_memory_arch() in drivers/of/fdt.c is
> > overridden by ARM and a number of other architectures. However...
> >
> > > Finally, once all memory has been added to the system we can parse and
> > > reserve all regions defined in the device tree. This really requires
> > > creating another function which will be called by arch specific code.
> >
> > ...Or it means getting rid of meminfo entirely so that memblock is
> > available earlier. Laura Abbott has just posted v2 of her series to do
> > exactly that. If you base on that then you should be able to do exactly
> > what I suggested.
> 
> I've checked Laura's patches and in fact it is possible to do memory
> reservation as a last step in early_init_dt_scan_memory(). However still
> see some problem which I have no idea how to resolve. Right now I focus
> only on ARM, so I have no idea how it is solved by other architectures.

It isn't at the moment because no other architectures are doing dynamic
allocation of regions in this way.

> On of the key features of the new binding is the ability to automatically
> allocate reserved regions of the given size. However kernel, initrd, dt
> and other sub-arch specific critical regions are marked/allocated in
> arm_memblock_init(), which is called after setup_machine_fdt(). This
> might lead to some serious failures when automatically reserved region
> overlaps with some critical resources. Do you have any idea how to solve
> this without a new callback?

Not without moving the kernel/initrd/fdt reservation earlier, which can
certainly be done after Laura's series is applied. It would also be
possible to split up the fixed regions from the dynamic regions so that
the dynamic stuff would be done later.

However, I don't want to block this series on Laura's rework. Go ahead
and keep it as a separate hook, but make the other changes I suggested,
particularly with regard to making the hook architecture independent.
Further rework can be done later after Laura's series is merged.

g.