Date:	Mon, 05 May 2008 12:58:31 +0200
From:	Johannes Weiner <hannes@...urebad.de>
To:	"Yinghai Lu" <yhlu.kernel@...il.com>
Cc:	"Ingo Molnar" <mingo@...e.hu>, linux-kernel@...r.kernel.org,
	linux-mm@...ck.org, "Andi Kleen" <andi@...stfloor.org>,
	"Andrew Morton" <akpm@...ux-foundation.org>,
	"Linus Torvalds" <torvalds@...ux-foundation.org>
Subject: Re: [RFC 0/2] Rootmem: boot-time memory allocator

Hi Yinghai,

"Yinghai Lu" <yhlu.kernel@...il.com> writes:

> On Sun, May 4, 2008 at 8:34 AM, Johannes Weiner <hannes@...urebad.de> wrote:
>>
>> Hi,
>>
>>  Johannes Weiner <hannes@...urebad.de> writes:
>>
>>  > Hi Yinghai,
>>  >
>>  > Johannes Weiner <hannes@...urebad.de> writes:
>>  >
>>  >> Hi,
>>  >>
>>  >> "Yinghai Lu" <yhlu.kernel@...il.com> writes:
>>  >>
>>  >>> On Sat, May 3, 2008 at 10:54 AM, Ingo Molnar <mingo@...e.hu> wrote:
>>  >>>>
>>  >>>>  * Johannes Weiner <hannes@...urebad.de> wrote:
>>  >>>>
>>  >>>>  > I was spending some time and work on the bootmem allocator the last
>>  >>>>  > few weeks and came to the conclusion that its current design is not
>>  >>>>  > appropriate anymore.
>>  >>>>  >
>>  >>>>  > As Ingo said in another email, NUMA technologies will become weirder,
>>  >>>>  > nodes whose PFNs span other nodes for example and it makes bootmem
>>  >>>>  > code become an unreadable mess.
>>  >>>>  >
>>  >>>>  > So I sat down two days ago and rewrote the allocator, here is the
>>  >>>>  > result: rootmem!
>>  >>>>
>>  >>>>  hehe :-)
>>  >>>>
>>  >>>>
>>  >>>>  > The biggest difference to the old design is that there is only one
>>  >>>>  > bitmap for all PFNs of all nodes together, so the overlapping PFN
>>  >>>>  > problems simply dissolve and fun like allocations crossing node
>>  >>>>  > boundaries work implicitely.  The new API requires every node used by
>>  >>>>  > the allocator to be registered and after that the bitmap gets
>>  >>>>  > allocated and the allocator enabled.
>>  >>>>  >
>>  >>>>  > I chose to add a new allocator rather than replacing bootmem at once
>>  >>>>  > because that would have required all callsites to switch in one go,
>>  >>>>  > which would be a lot.  The new allocator can be adopted more slowly
>>  >>>>  > and I added a compatibility API for everything besides actually
>>  >>>>  > setting up the allocator.  When the last user dies, bootmem can be
>>  >>>>  > dropped completely (including pgdat->bdata, whee..)
>>  >>>>  >
>>  >>>>  > The main ideas from bootmem have been stolen^W preserved but the new
>>  >>>>  > design allowed me to shrink the code a lot and express things more
>>  >>>>  > simple and clear:
>>  >>>>  >
>>  >>>>  > $ sloc.awk < mm/bootmem.c
>>  >>>>  > 455 lines of code, 65 lines of comments (520 lines total)
>>  >>>>  >
>>  >>>>  > $ sloc.awk < mm/rootmem.c
>>  >>>>  > 243 lines of code, 96 lines of comments (339 lines total)
>>  >>>>
>>  >>>>  amazing!
>>  >>>>
>>  >>>>  i'd still suggest to keep it all named bootmem though :-/ How about
>>  >>>>  bootmem2.c and then renaming it back to bootmem.c, once the last user is
>>  >>>>  gone? That would save people from having to rename whole chapters in
>>  >>>>  entire books ;-)
>>  >>>
>>  >>> for spanning support node0:0-2g, 4-6g; node1: 2-4g, 6-8g, could have
>>  >>> some problem.
>>  >>
>>  >> Could you eleborate on that?
>>  >>
>>  >>> +/*
>>  >>> + * rootmem_register_node - register a node to rootmem
>>  >>> + * @nid: node id
>>  >>> + * @start: first pfn on the node
>>  >>> + * @end: first pfn after the node
>>  >>> + *
>>  >>> + * This function must not be called anymore if the allocator
>>  >>> + * is already up and running (rootmem_setup() has been called).
>>  >>> + */
>>  >>> +void __init rootmem_register_node(int nid, unsigned long start,
>>  >>> +                       unsigned long end)
>>  >>> +{
>>  >>> +       BUG_ON(rootmem_functional);
>>  >>> +
>>  >>> +       if (start < rootmem_min_pfn)
>>  >>> +               rootmem_min_pfn = start;
>>  >>> +       if (end > rootmem_max_pfn)
>>  >>> +               rootmem_max_pfn = end;
>>  >>> +
>>  >>> +       rootmem_node_pages[nid] = end - start;
>>  >>> +       rootmem_node_offsets[nid] = start;
>>  >>> +       rootmem_nr_nodes++;
>>  >>> +}
>>  >>>
>>  >>> could change rootmem_node_pages/offsets to be struct array with
>>  >>> offset, pages, and nid. and every node could several struct. and whole
>>  >>> array should be sorted with nid.
>>
>>  In the long term, this would have to be implemented no matter if
>>  rootmem/bootmem2 gets merged or not, because bootmem suffers the same
>>  problem, right?
>>
>>
>>  >> The whole point is to be agnostic about weird NUMA configs.  Right now,
>>  >> I am pretty proud of the simple data structures and I would avoid
>>  >> blowing them up again unless there is a hard reason to do so.
>>
>>  This is non-helping crap, please excuse me.
>>
>>
>>  > One thing I have found is that __rootmem_alloc_node can not garuantee
>>  > that the memory it returns is on the requested node right now.
>>
>>  Hm, we have two choices: Either we introduce a new API that requests the
>>  arch code to register not only node ranges but also subranges on that
>>  node, or we won't garuantee that you get all memory on the node you
>>  specified.  Correct?
>>
>>  The first option would be what you have proposed, I think.
>
> 1. current bootmem, add not_used_map to bdata.
> 2. or in bootmem2, use pages_offset struct for every range... so one
> node could have several ranges.

I think I found a solution, please have a look at the bootmem2 patches
(coming soon).

	Hannes
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists