On Fri 25-01-19 19:15:49, Michal Hocko wrote:
> On Fri 25-01-19 18:33:15, Michal Hocko wrote:
> > On Fri 25-01-19 17:39:38, Michal Hocko wrote:
> > > On Fri 25-01-19 11:16:30, robert shteynfeld wrote:
> > > > Attached is the dmesg from patched kernel.
> > >
> > > Your Node1 physical memory range precedes Node0 which is quite unusual
> > > but it shouldn't be a huge problem on its own. But memory ranges are
> > > not aligned to the memory section
> > >
> > > [ 0.286954] Early memory node ranges
> > > [ 0.286955] node 1: [mem 0x0000000000001000-0x0000000000090fff]
> > > [ 0.286955] node 1: [mem 0x0000000000100000-0x00000000dbdf8fff]
> > > [ 0.286956] node 1: [mem 0x0000000100000000-0x0000001423ffffff]
> > > [ 0.286956] node 0: [mem 0x0000001424000000-0x0000002023ffffff]
> > >
> > > As you can see the last pfn for the node1 is inside the section and
> > > Node0 starts right after. This is quite unusual as well. If for no other
> > > reasons then the memmap of those struct pages will be remote for one or
> > > the other. Actually I am not even sure we can handle that properly
> > > because we do expect 1:1 mapping between sections and nodes.
> > >
> > > Now it also makes some sense why 2830bf6f05fb ("mm, memory_hotplug:
> > > initialize struct pages for the full memory section") made any
> > > difference. We simply write over a potentially initialized struct page
> > > and blow up on that. I strongly suspect that the commit just uncovered
> > > a pre-existing problem. Let me think what we can do about that.
> >
> > Appart from force aligning node's start the only other option is to
> > revert 2830bf6f05fb and handling the underlying issue in the hotplug
> > code.
>
> We cannot really align because we have things like ZONE_DMA starting at
> 0x1000 and who knows what else. So let's go with the revert. Hutplug
> simply needs a larger surgery to get rid of the PAGES_PER_SECTION
> inherent assumptions.
>
> Linus, could you take the revert please?
or should I post the patch as a reply to make your life easier?
> From 817b18d3db36a6900ca9043af8c1416c56358be3 Mon Sep 17 00:00:00 2001
> From: Michal Hocko <mho...@suse.com>
> Date: Fri, 25 Jan 2019 19:08:58 +0100
> Subject: [PATCH] Revert "mm, memory_hotplug: initialize struct pages for the
> full memory section"
>
> This reverts commit 2830bf6f05fb3e05bc4743274b806c821807a684.
>
> The underlying assumption that one sparse section belongs into a single
> numa node doesn't hold really. Robert Shteynfeld has reported a boot
> failure. The boot log was not captured but his memory layout is as
> follows:
> [ 0.286954] Early memory node ranges
> [ 0.286955] node 1: [mem 0x0000000000001000-0x0000000000090fff]
> [ 0.286955] node 1: [mem 0x0000000000100000-0x00000000dbdf8fff]
> [ 0.286956] node 1: [mem 0x0000000100000000-0x0000001423ffffff]
> [ 0.286956] node 0: [mem 0x0000001424000000-0x0000002023ffffff]
>
> This means that node0 starts in the middle of a memory section which is
> also in node1. memmap_init_zone tries to initialize padding of a section
> even when it is outside of the given pfn range because there are code
> paths (e.g. memory hotplug) which assume that the full worth of memory
> section is always initialized. In this particular case, though, such a
> range is already intialized and most likely already managed by the page
> allocator. Scribbling over those pages corrupts the internal state and
> likely blows up when any of those pages gets used.
>
> Reported-by: Robert Shteynfeld <robert.shteynf...@gmail.com>
> Fixes: 2830bf6f05fb ("mm, memory_hotplug: initialize struct pages for the
> full memory section")
> Cc: stable
> Signed-off-by: Michal Hocko <mho...@suse.com>
> ---
> mm/page_alloc.c | 12 ------------
> 1 file changed, 12 deletions(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index d295c9bc01a8..35fdde041f5c 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5701,18 +5701,6 @@ void __meminit memmap_init_zone(unsigned long size,
> int nid, unsigned long zone,
> cond_resched();
> }
> }
> -#ifdef CONFIG_SPARSEMEM
> - /*
> - * If the zone does not span the rest of the section then
> - * we should at least initialize those pages. Otherwise we
> - * could blow up on a poisoned page in some paths which depend
> - * on full sections being initialized (e.g. memory hotplug).
> - */
> - while (end_pfn % PAGES_PER_SECTION) {
> - __init_single_page(pfn_to_page(end_pfn), end_pfn, zone, nid);
> - end_pfn++;
> - }
> -#endif
> }
>
> #ifdef CONFIG_ZONE_DEVICE
> --
> 2.20.1
>
> --
> Michal Hocko
> SUSE Labs
--
Michal Hocko
SUSE Labs
