Messages in this thread | | | From | Johannes Weiner <> | Subject | Re: [PATCH 0/3] bootmem2 III | Date | Tue, 13 May 2008 14:40:44 +0200 |
| |
Hi,
Andi Kleen <andi@firstfloor.org> writes:
> Johannes Weiner wrote: > >>> On Fri, May 09, 2008 at 05:17:13PM +0200, Johannes Weiner wrote: >>>> here is bootmem2, a memory block-oriented boot time allocator. >>>> >>>> Recent NUMA topologies broke the current bootmem's assumption that >>>> memory nodes provide non-overlapping and contiguous ranges of pages. >>> I'm still not sure that's a really good rationale for bootmem2. >>> e.g. the non continuous nodes are really special cases and there tends >>> to be enough memory at the beginning which is enough for boot time >>> use, so for those systems it would be quite reasonably to only >>> put the continuous starts of the nodes into bootmem. >> >> Hm, that would put the logic into arch-code. I have no strong opinion >> about it. > > In fact I suspect the current code will already work like that > implicitely. The aliasing is only a problem for the new "arbitary node > free_bootmem" right?
And that alloc_bootmem_node() can not garuantee node-locality which is the much worse part, I think.
>>> That said the bootmem code has gotten a little crufty and a clean >>> rewrite might be a good idea. >> >> I agree completely. > > The trouble is just that bootmem is used in early boot and early boot is > very subtle and getting it working over all architectures could be a > challenge. Not wanting to discourage you, but it's not exactly the > easiest part of the kernel to hack on.
Bootmem seemed pretty self-contained to me, at least in the beginning. The bad thing is that I can test only the most simple configuration with it.
I was wondering yesterday if it would be feasible to enforce contiguousness for nodes. So that arch-code does not create one pgdat for each node but one for each contiguous block. I have not yet looked deeper into it, but I suspect that other mm code has similar problems with nodes spanning other nodes.
Hannes
| |