Re: [PATCH] memory_hotplug: fix the panic when memory end is not on the section boundary

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi Michal,

It is tricky, but probably can be done. Either change
memmap_init_zone() or its caller to also cover the ends and starts of
unaligned sections to initialize and reserve pages.

The same thing would also need to be done in deferred_init_memmap() to
cover the deferred init case.

For hotplugged memory we do not need to worry, as that one is always
section aligned.

Do you think this would be a better approach?

Thank you,
Pavel
On Mon, Sep 10, 2018 at 10:00 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Mon 10-09-18 13:46:45, Pavel Tatashin wrote:
> >
> >
> > On 9/10/18 9:17 AM, Michal Hocko wrote:
> > > [Cc Pavel]
> > >
> > > On Mon 10-09-18 14:35:27, Mikhail Zaslonko wrote:
> > >> If memory end is not aligned with the linux memory section boundary, such
> > >> a section is only partly initialized. This may lead to VM_BUG_ON due to
> > >> uninitialized struct pages access from is_mem_section_removable() or
> > >> test_pages_in_a_zone() function.
> > >>
> > >> Here is one of the panic examples:
> > >>  CONFIG_DEBUG_VM_PGFLAGS=y
> > >>  kernel parameter mem=3075M
> > >
> > > OK, so the last memory section is not full and we have a partial memory
> > > block right?
> > >
> > >>  page dumped because: VM_BUG_ON_PAGE(PagePoisoned(p))
> > >
> > > OK, this means that the struct page is not fully initialized. Do you
> > > have a specific place which has triggered this assert?
> > >
> > >>  ------------[ cut here ]------------
> > >>  Call Trace:
> > >>  ([<000000000039b8a4>] is_mem_section_removable+0xcc/0x1c0)
> > >>   [<00000000009558ba>] show_mem_removable+0xda/0xe0
> > >>   [<00000000009325fc>] dev_attr_show+0x3c/0x80
> > >>   [<000000000047e7ea>] sysfs_kf_seq_show+0xda/0x160
> > >>   [<00000000003fc4e0>] seq_read+0x208/0x4c8
> > >>   [<00000000003cb80e>] __vfs_read+0x46/0x180
> > >>   [<00000000003cb9ce>] vfs_read+0x86/0x148
> > >>   [<00000000003cc06a>] ksys_read+0x62/0xc0
> > >>   [<0000000000c001c0>] system_call+0xdc/0x2d8
> > >>
> > >> This fix checks if the page lies within the zone boundaries before
> > >> accessing the struct page data. The check is added to both functions.
> > >> Actually similar check has already been present in
> > >> is_pageblock_removable_nolock() function but only after the struct page
> > >> is accessed.
> > >>
> > >
> > > Well, I am afraid this is not the proper solution. We are relying on the
> > > full pageblock worth of initialized struct pages at many other place. We
> > > used to do that in the past because we have initialized the full
> > > section but this has been changed recently. Pavel, do you have any ideas
> > > how to deal with this partial mem sections now?
> >
> > We have:
> >
> > remove_memory()
> >       BUG_ON(check_hotplug_memory_range(start, size))
> >
> > That supposed to safely check for this condition: if [start, start +
> > size) not block size aligned (and we know block size is section
> > aligned), hot remove is not allowed. The problem is this check is late,
> > and only happens when invalid range has already passed through previous
> > checks.
> >
> > We could add check_hotplug_memory_range() to is_mem_section_removable():
> >
> > is_mem_section_removable(start_pfn, nr_pages)
> >  if (check_hotplug_memory_range(PFN_PHYS(start_pfn), PFN_PHYS(nr_pages)))
> >   return false;
> >
> > I think it should work.
>
> I do not think we want to sprinkle these tests over all pfn walkers. Can
> we simply initialize those uninitialized holes as well and make them
> reserved without handing them over to the page allocator? That would be
> much more robust approach IMHO.
> --
> Michal Hocko
> SUSE Labs
>




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux