Re: + bootmem-node-setup-agnostic-free_bootmem.patch added to -mm tree

["Yinghai Lu" <yhlu.kernel@gmail.com>]

On Tue, Apr 15, 2008 at 12:55 PM, Johannes Weiner <hannes@saeurebad.de> wrote:
> Hi,
>
>
>
>  "Yinghai Lu" <yhlu.kernel@gmail.com> writes:
>
>  > On Tue, Apr 15, 2008 at 5:51 AM, Johannes Weiner <hannes@saeurebad.de> wrote:
>  >> Hi Ingo,
>  >>
>  >>
>  >>
>  >>  Ingo Molnar <mingo@elte.hu> writes:
>  >>
>  >>  > * akpm@linux-foundation.org <akpm@linux-foundation.org> wrote:
>  >>  >
>  >>  >> Subject: bootmem: node-setup agnostic free_bootmem()
>  >>  >> From: Johannes Weiner <hannes@saeurebad.de>
>  >>  >>
>  >>  >> Make free_bootmem() look up the node holding the specified address
>  >>  >> range which lets it work transparently on single-node and multi-node
>  >>  >> configurations.
>  >>  >
>  >>  > this patch does not fix the bug Yinghai's (now dropped) patches solved:
>  >>  > reserve_early() allocations. So NAK until the full problem has been
>  >>  > sorted out ...
>  >>
>  >>  Okay, NAK on -mm and -x86 for sure.  The patch was meant for mainline
>  >>  where there is no need for free_bootmem() going across nodes, right?
>  >>
>  >>  But I still object to the way Yinghai implemented it.
>  >>  free_bootmem_core() should not be twisted like this.
>  >>
>  >>  How about the following (untested, even uncompiled, but you should get
>  >>  the idea) proposal which would replace the patch discussed in this
>  >>  thread:
>  >>
>  >>  --- tree-linus.orig/mm/bootmem.c
>  >>  +++ tree-linus/mm/bootmem.c
>  >>  @@ -421,7 +421,25 @@ int __init reserve_bootmem(unsigned long
>  >>
>  >>
>  >>   void __init free_bootmem(unsigned long addr, unsigned long size)
>  >>   {
>  >>  -       free_bootmem_core(NODE_DATA(0)->bdata, addr, size);
>  >>  +       bootmem_data_t *bdata;
>  >>  +
>  >>  +       list_for_each_entry(bdata, &bdata_list, list) {
>  >>  +               unsigned long remainder = 0;
>  >>
>  >> +
>  >>  +               if (addr < bdata->node_boot_start)
>  >>  +                       continue;
>  >>  +
>  >>  +               if (PFN_DOWN(addr + size) > bdata->node_low_pfn)
>  >>  +                       remainder = PFN_DOWN(addr + size) - bdata->node_low_pfn;
>  >>  +
>  >>  +               size -= PFN_PHYS(remainder);
>  >>
>  >> +               free_bootmem_core(bdata, addr, size)
>  >>  +
>  >>  +               if (!remainder)
>  >>  +                       break;
>  >>  +
>  >>  +               addr = PFN_PHYS(bdata->node_low_pfn + 1);
>  >>  +       }
>  >>
>  >>  }
>  >>
>  >>   unsigned long __init free_all_bootmem(void)
>  >
>  > how about
>  > 1. bdata is not sorted?
>
>  They are kept in a sorted list.  How could they be unsorted?
>
>
>  > 2. intel cross node box: node0: 0g-2g, 4g-6g, node1: 2g-4g, 6g-8g. i
>  > don't think they have two bdata struct for every node.
>
>  How do the bdata structures represent this setup right now?  Are you
>  sure that there is not a node descriptor for every contiguous region?

http://lkml.org/lkml/2008/3/25/233

Subject	[patch] srat, x86_64: Add support for nodes spanning other nodes

For example, If the physical address layout on a two node system with 8 GB
memory is something like:
node 0: 0-2GB, 4-6GB
node 1: 2-4GB, 6-8GB

Current kernels fail to boot/detect this NUMA topology.

ACPI SRAT tables can expose such a topology which needs to be supported.

Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>

YH