Messages in this thread | | | Date | Wed, 21 May 2008 16:57:35 -0700 | From | Andrew Morton <> | Subject | Re: [RFC 0/3] bootmem rewrite |
| |
On Wed, 21 May 2008 03:37:35 +0200 Johannes Weiner <hannes@saeurebad.de> wrote:
> Hi, > > This is a complete overhaul of the bootmem allocator while preserving > its original functionality, excluding bugs.
Where angels fear to tread.
> free_bootmem and reserve_bootmem become a bit stricter than they are > right now, callsites have to make sure that the PFN range is > contiguous but it might go across node boundaries. > > alloc_bootmem satisfying the allocation goal is more likely as the > routines will try to allocate on the node holding the goal first > before falling back as opposed to the original behaviour that > satisfies the goal only if it is on the first node. > > All in all, I think the code has become simpler and cleaner. All > public interfaces have been documented, too. > > The first patch moves the bootmem node descriptor definitions into > bootmem.c where they belong. > > The second patch is the new allocator itself. > > The third patch converts all users of ->node_boot_start to > ->node_min_pfn as this is what they really use. It then removes the > unused ->node_boot_start. > > Compile and runtime tested on X86_32, therefor RFC only. > > arch/alpha/mm/numa.c | 8 +- > arch/arm/mm/discontig.c | 34 +- > arch/arm/plat-omap/fb.c | 4 +- > arch/avr32/mm/init.c | 3 +- > arch/ia64/mm/discontig.c | 30 +- > arch/m32r/mm/discontig.c | 4 +- > arch/m32r/mm/init.c | 4 +- > arch/m68k/mm/init.c | 4 +- > arch/mips/sgi-ip27/ip27-memory.c | 3 +- > arch/mn10300/mm/init.c | 6 +- > arch/parisc/mm/init.c | 3 +- > arch/powerpc/mm/numa.c | 3 +- > arch/sh/mm/init.c | 2 +- > arch/sh/mm/numa.c | 5 +- > arch/sparc64/mm/init.c | 3 +- > arch/x86/mm/discontig_32.c | 3 +- > arch/x86/mm/numa_64.c | 6 +- > include/linux/bootmem.h | 115 ++--- > mm/bootmem.c | 914 +++++++++++++++++++------------------- > mm/page_alloc.c | 4 +-
Oh gee.
bootmem is an area where large numbers of people have done hit-and-run jobs over a lot of years. Nobody owns it and I'm sure that you are now the world's expert. We just need to push ahead with this, I guess.
I expect there will be problems - so many architectures which do such different things, and all the configuration options churning things around.
So how to move ahead with this?
- I think I'd prefer not to drop
mm-fix-free_all_bootmem_core-alignment-check.patch mm-normalize-internal-argument-passing-of-bootmem-data.patch mm-unexport-__alloc_bootmem_core.patch
because those are small, simple things which are on track for 2.6.27 whereas a massive rewrite may take longer to get merged, and may never get there at all, in which case we lost those little fixes.
- It would suit my purposes to have these patches right at the tail of the -mm patch queue so that I can drop them easily if problems occur, and so that others can revert them easily when diagnosing problems.
- It would be nice to get some review attention from architecture guys, but I can understand them finding other things to do, when bootmem is presumably good-enough-for-now.
- Is x86_32 the only test platform which you have available? Awkward.
Anyway, if you can redo these patches against most-recent-mm or, better, against http://userweb.kernel.org/~akpm/mmotm then it would make things easier for me to handle. I can then at least test it all on my seven-odd test boxes. Please feel free to ping me if you want a single rolled-up patch - that's always trivial and I can do it in three minutes.
Finally, if you haven't done so, I'd encourage you to stuff as many handy debugging printks into this code as you possibly can. Just fill 'er up with them. So that when people start running it and it goes boom, they can send you their debug output _without_ having to go through another handful of email-email-patch-rebuild-retest cycles. We can pull them all out later on.
| |