Re: [PATCH 06/20] early_res: seperate common memmap func from e820.c to fw_memmap.cy

[Ingo Molnar <mingo@elte.hu>]

* Benjamin Herrenschmidt <benh@kernel.crashing.org> wrote:

> 
> > I though one possibility would be to have LMB regions become more lists
> > than arrays, so that the static storage only needs to cover as much as
> > is needed during really early boot (and we could probably still move the
> > BSS top point on some archs to dynamically make more ... actually we
> > could be smart arses and use LMB to allocate more LMB list heads if we
> > are reaching the table limit :-)
> 
> Actually what about that:
> 
> LMB entries are linked-listed. The array is just storage for those entry
> "heads".
> 
> The initial static array only needs to be big enough for very very early
> platform specific kernel bits and pieces, so it could even be sized by a
> Kconfig option. Or it could just use a klimit moving trick to pick up a
> page right after the BSS but that may need to be arch specific.
> 
> lmb_init() queues all the entries from the initial array in a freelist
> 
> lmb_alloc() and lmb_reserve() just pop entries from that freelist to
> populate the two main linked lists (memory and reserved).
> 
> When something tries to dequeue up the last freelist entry, then under
> the hood, LMB uses it instead to allocate a new block of LMB entries
> that gets added to the freelist.
> 
> We never free blocks of LMB entries.
> 
> That way, we can fine tine the static array to be as small as we can
> realistically make it be, and we have no boundary limitations on the
> amount of entries in either the memory list or the reserved list.
> 
> I'm a bit too flat out right now to write code, but if there's no objection, 
> I might give that a go either later this week or next week, see if I can 
> replace bootmem on powerpc.

That would be fantastic! PowerPC and x86 both doing it would give it enough of 
a critical mass to make the removal of bootmem realistic.

Thanks,

	Ingo