Re: [PATCH -next] mm/hotplug: skip bad PFNs from pfn_to_online_page()

[Dan Williams <dan.j.williams@intel.com>]

On Wed, Jun 12, 2019 at 2:47 PM Qian Cai <cai@lca.pw> wrote:
>
> On Wed, 2019-06-12 at 12:38 -0700, Dan Williams wrote:
> > On Wed, Jun 12, 2019 at 12:37 PM Dan Williams <dan.j.williams@intel.com>
> > wrote:
> > >
> > > On Wed, Jun 12, 2019 at 12:16 PM Qian Cai <cai@lca.pw> wrote:
> > > >
> > > > The linux-next commit "mm/sparsemem: Add helpers track active portions
> > > > of a section at boot" [1] causes a crash below when the first kmemleak
> > > > scan kthread kicks in. This is because kmemleak_scan() calls
> > > > pfn_to_online_page(() which calls pfn_valid_within() instead of
> > > > pfn_valid() on x86 due to CONFIG_HOLES_IN_ZONE=n.
> > > >
> > > > The commit [1] did add an additional check of pfn_section_valid() in
> > > > pfn_valid(), but forgot to add it in the above code path.
> > > >
> > > > page:ffffea0002748000 is uninitialized and poisoned
> > > > raw: ffffffffffffffff ffffffffffffffff ffffffffffffffff ffffffffffffffff
> > > > raw: ffffffffffffffff ffffffffffffffff ffffffffffffffff ffffffffffffffff
> > > > page dumped because: VM_BUG_ON_PAGE(PagePoisoned(p))
> > > > ------------[ cut here ]------------
> > > > kernel BUG at include/linux/mm.h:1084!
> > > > invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN PTI
> > > > CPU: 5 PID: 332 Comm: kmemleak Not tainted 5.2.0-rc4-next-20190612+ #6
> > > > Hardware name: Lenovo ThinkSystem SR530 -[7X07RCZ000]-/-[7X07RCZ000]-,
> > > > BIOS -[TEE113T-1.00]- 07/07/2017
> > > > RIP: 0010:kmemleak_scan+0x6df/0xad0
> > > > Call Trace:
> > > >  kmemleak_scan_thread+0x9f/0xc7
> > > >  kthread+0x1d2/0x1f0
> > > >  ret_from_fork+0x35/0x4
> > > >
> > > > [1] https://patchwork.kernel.org/patch/10977957/
> > > >
> > > > Signed-off-by: Qian Cai <cai@lca.pw>
> > > > ---
> > > >  include/linux/memory_hotplug.h | 1 +
> > > >  1 file changed, 1 insertion(+)
> > > >
> > > > diff --git a/include/linux/memory_hotplug.h
> > > > b/include/linux/memory_hotplug.h
> > > > index 0b8a5e5ef2da..f02be86077e3 100644
> > > > --- a/include/linux/memory_hotplug.h
> > > > +++ b/include/linux/memory_hotplug.h
> > > > @@ -28,6 +28,7 @@
> > > >         unsigned long ___nr = pfn_to_section_nr(___pfn);           \
> > > >                                                                    \
> > > >         if (___nr < NR_MEM_SECTIONS && online_section_nr(___nr) && \
> > > > +           pfn_section_valid(__nr_to_section(___nr), pfn) &&      \
> > > >             pfn_valid_within(___pfn))                              \
> > > >                 ___page = pfn_to_page(___pfn);                     \
> > > >         ___page;                                                   \
> > >
> > > Looks ok to me:
> > >
> > > Acked-by: Dan Williams <dan.j.williams@intel.com>
> > >
> > > ...but why is pfn_to_online_page() a multi-line macro instead of a
> > > static inline like all the helper routines it invokes?
> >
> > I do need to send out a refreshed version of the sub-section patchset,
> > so I'll fold this in and give you a Reported-by credit.
>
> BTW, not sure if your new version will fix those two problem below due to the
> same commit.
>
> https://patchwork.kernel.org/patch/10977957/
>
> 1) offline is busted [1]. It looks like test_pages_in_a_zone() missed the same
> pfn_section_valid() check.
>
> 2) powerpc booting is generating endless warnings [2]. In vmemmap_populated() at
> arch/powerpc/mm/init_64.c, I tried to change PAGES_PER_SECTION to
> PAGES_PER_SUBSECTION, but it alone seems not enough.

Yes, I was just sending you another note about this. I don't think
your proposed fix is sufficient. The original intent of
pfn_valid_within() was to use it as a cheaper lookup after already
validating that the first page in a MAX_ORDER_NR_PAGES range satisfied
pfn_valid(). Quoting commit  14e072984179 "add pfn_valid_within helper
for sub-MAX_ORDER hole detection":

    Add a pfn_valid_within() helper which should be used when scanning pages
    within a MAX_ORDER_NR_PAGES block when we have already checked the
validility
    of the block normally with pfn_valid().  This can then be
optimised away when
    we do not have holes within a MAX_ORDER_NR_PAGES block of pages.

So, with that insight I think the complete fix is this:

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 6dd52d544857..9d15ec793330 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -1428,7 +1428,7 @@ void memory_present(int nid, unsigned long
start, unsigned long end);
 #ifdef CONFIG_HOLES_IN_ZONE
 #define pfn_valid_within(pfn) pfn_valid(pfn)
 #else
-#define pfn_valid_within(pfn) (1)
+#define pfn_valid_within(pfn) pfn_section_valid(pfn)
 #endif

 #ifdef CONFIG_ARCH_HAS_HOLES_MEMORYMODEL