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

[Ingo Molnar <mingo@elte.hu>]

* Yinghai Lu <yinghai@kernel.org> wrote:

> please check
> 
> [PATCH 01/20] x86: add find_e820_area_node
> 
> 
> [RFC PATCH] x86: use lmb to replace early_res
> 
> still keep kernel/early_res.c for the extension.
> 
> should move those file to lib/lmb.c later?
> 
> Signed-off-by: Yinghai Lu <yinghai@kernel.org>
> 
> ---
>  arch/x86/Kconfig               |    1 
>  arch/x86/include/asm/e820.h    |   38 +-
>  arch/x86/include/asm/lmb.h     |    8 
>  arch/x86/kernel/e820.c         |  163 +----------
>  arch/x86/kernel/head.c         |    2 
>  arch/x86/kernel/head32.c       |    4 
>  arch/x86/kernel/head64.c       |    2 
>  arch/x86/kernel/setup.c        |    2 
>  arch/x86/kernel/setup_percpu.c |    6 
>  include/linux/early_res.h      |    9 
>  include/linux/lmb.h            |    5 
>  kernel/early_res.c             |  594 ++++++++++++++++-------------------------
>  lib/lmb.c                      |    9 
>  mm/page_alloc.c                |    2 
>  mm/sparse-vmemmap.c            |    4 
>  15 files changed, 321 insertions(+), 528 deletions(-)

That looks like a very promising direction!

There's several things to do to make the approach fully clean:

1)

I think we want to shape this as a series of simpler (and bisectable) patches.

2)

I think we also need to concentrate the changes back into LMB:

> Index: linux-2.6/include/linux/early_res.h
> ===================================================================
> --- linux-2.6.orig/include/linux/early_res.h
> +++ linux-2.6/include/linux/early_res.h
> @@ -5,15 +5,18 @@
>  extern void reserve_early(u64 start, u64 end, char *name);
>  extern void reserve_early_overlap_ok(u64 start, u64 end, char *name);
>  extern void free_early(u64 start, u64 end);
> -void free_early_partial(u64 start, u64 end);
>  extern void early_res_to_bootmem(u64 start, u64 end);
>  
> -void reserve_early_without_check(u64 start, u64 end, char *name);
>  u64 find_early_area(u64 ei_start, u64 ei_last, u64 start, u64 end,
>  			 u64 size, u64 align);
>  u64 find_early_area_size(u64 ei_start, u64 ei_last, u64 start,
>  			 u64 *sizep, u64 align);
> -u64 find_fw_memmap_area(u64 start, u64 end, u64 size, u64 align);
> +u64 find_lmb_area(u64 start, u64 end, u64 size, u64 align);
> +u64 find_lmb_area_size(u64 start, u64 *sizep, u64 align);
> +u64 find_lmb_area_node(int nid, u64 start, u64 end, u64 size, u64 align);
> +void lmb_register_active_regions(int nid, unsigned long start_pfn,
> +					 unsigned long last_pfn);
> +u64 lmb_hole_size(u64 start, u64 end);
>  u64 get_max_mapped(void);
>  #include <linux/range.h>
>  int get_free_all_memory_range(struct range **rangep, int nodeid);

those new lmb_*() APIs should go into lmb.h.

3) 

Furthermore, i think all of early_res.c should move into lmb.c as well and we 
should eliminate kernel/early_res.c.

early_res.h will go away as well and all the new APIs will be in lmb.h.

4)

Also, we should move lib/lmb.c to mm/lmb.c, as now it's not just some optional 
library but _the_ main early-reserve memory subsystem used by the biggest 
Linux architectures.

5)

Could we perhaps also try to eliminate e820_*() method uses in arch/x86/, and 
replace them by lmb_*() API uses? (that too should be a step by step method, 
for bisectability)

> +++ linux-2.6/include/linux/lmb.h
> @@ -26,7 +26,8 @@ struct lmb_property {
>  struct lmb_region {
>  	unsigned long cnt;
>  	u64 size;
> -	struct lmb_property region[MAX_LMB_REGIONS+1];
> +	struct lmb_property *region;
> +	unsigned long region_array_size;
>  };

I suspect this should keep current LMB architectures still working, right?

	Ingo