Re: [PATCH v2 2/2] mm/page_alloc: add some comments to explain the possible hole in __pageblock_pfn_to_page()

[Michal Hocko]

On Mon 24-04-23 20:48:32, Baolin Wang wrote:
> 
> 
> On 4/24/2023 8:08 PM, Michal Hocko wrote:
> > On Mon 24-04-23 19:40:30, Baolin Wang wrote:
> > > 
> > > 
> > > On 4/24/2023 7:34 PM, Michal Hocko wrote:
> > > > On Mon 24-04-23 19:20:43, Baolin Wang wrote:
> > > > > 
> > > > > 
> > > > > On 4/24/2023 5:54 PM, Michal Hocko wrote:
> > > > > > On Sun 23-04-23 18:59:11, Baolin Wang wrote:
> > > > > > > Now the __pageblock_pfn_to_page() is used by set_zone_contiguous(), which
> > > > > > > checks whether the given zone contains holes, and uses pfn_to_online_page()
> > > > > > > to validate if the start pfn is online and valid, as well as using pfn_valid()
> > > > > > > to validate the end pfn.
> > > > > > > 
> > > > > > > However, the __pageblock_pfn_to_page() function may return non-NULL even
> > > > > > > if the end pfn of a pageblock is in a memory hole in some situations. For
> > > > > > > example, if the pageblock order is MAX_ORDER, which will fall into 2
> > > > > > > sub-sections, and the end pfn of the pageblock may be hole even though
> > > > > > > the start pfn is online and valid.
> > > > > > > 
> > > > > > > This did not break anything until now, but the zone continuous is fragile
> > > > > > > in this possible scenario. So as previous discussion[1], it is better to
> > > > > > > add some comments to explain this possible issue in case there are some
> > > > > > > future pfn walkers that rely on this.
> > > > > > > 
> > > > > > > [1] https://lore.kernel.org/all/87r0sdsmr6.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
> > > > > > 
> > > > > > Do I remember correctly you've had a specific configuration that would
> > > > > > trigger this case?
> > > > > 
> > > > > Yes, I provided an example in previous thread [2] so show the
> > > > > __pageblock_pfn_to_page() is fragile in some cases.
> > > > > 
> > > > > [2] https://lore.kernel.org/all/52dfdd2e-9c99-eac4-233e-59919a24323e@xxxxxxxxxxxxxxxxx/
> > > > 
> > > > Please make it a part of the changelog.
> > > 
> > > Sure.
> > > 
> > > > > > 
> > > > > > > Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> > > > > > > ---
> > > > > > > Changes from v1:
> > > > > > >     - Update the comments per Ying and Mike, thanks.
> > > > > > > ---
> > > > > > >     mm/page_alloc.c | 7 +++++++
> > > > > > >     1 file changed, 7 insertions(+)
> > > > > > > 
> > > > > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > > > > > index 6457b64fe562..9756d66f471c 100644
> > > > > > > --- a/mm/page_alloc.c
> > > > > > > +++ b/mm/page_alloc.c
> > > > > > > @@ -1502,6 +1502,13 @@ void __free_pages_core(struct page *page, unsigned int order)
> > > > > > >      * interleaving within a single pageblock. It is therefore sufficient to check
> > > > > > >      * the first and last page of a pageblock and avoid checking each individual
> > > > > > >      * page in a pageblock.
> > > > > > > + *
> > > > > > > + * Note: the function may return non-NULL even if the end pfn of a pageblock
> > > > > > > + * is in a memory hole in some situations. For example, if the pageblock
> > > > > > > + * order is MAX_ORDER, which will fall into 2 sub-sections, and the end pfn
> > > > > > > + * of the pageblock may be hole even though the start pfn is online and valid.
> > > > > > > + * This did not break anything until now, but be careful about this possible
> > > > > > > + * issue when checking whether all pfns of a pageblock are valid.
> > > > > > 
> > > > > > It is not really clear what you should be doing (other than to be
> > > > > > careful which is not helpful much TBH) when you encounter this
> > > > > > situation. If the reality changes and this would break in the future
> > > > > > what would breakage look like? What should be done about that?
> > > > > 
> > > > > That depends on what the future pfn walkers do, which may access some hole
> > > > > memory with zero-init page frame. For example, if checking the
> > > > > __PageMovable() for a zero-init page frame, that will crash the system. But
> > > > > I can not list all the possible cases.
> > > > > 
> > > > > So how about below words?
> > > > > 
> > > > >    * Note: the function may return non-NULL even if the end pfn of a pageblock
> > > > >    * is in a memory hole in some situations. For example, if the pageblock
> > > > >    * order is MAX_ORDER, which will fall into 2 sub-sections, and the end pfn
> > > > >    * of the pageblock may be hole even though the start pfn is online and
> > > > > valid.
> > > > >    * This did not break anything until now, but be careful about this possible
> > > > >    * issue when checking whether all pfns of a pageblock are valid, that may
> > > > >    * lead to accessing empty page frame, and the worst case can crash the
> > > > > system.
> > > > >    * So you should use pfn_to_onlie_page() instead of pfn_valid() to valid the
> > > > >    * pfns in a pageblock if such case happens.
> > > > 
> > > > Does that mean that struct page is not initialized and PagePoisoned will
> > > > trigger or it is just zero-prefilled?
> > > 
> > > In the example I provided[2], these page frames of the hole memory are
> > > zero-prefilled.
> > 
> > OK, so make _that_ explicit in the comment. Essentially you want to say
> > that there are cases where we have zero-initialized struct pages for
> > memory holes. In general no pfn walker should touch a physical memory
> > range for pfn where the struct page doesn't contain any metadata it
> > recognizes. Zero fill struct pages do not contain any distinguishable
> > state so that makes it less of a problem.
> > 
> > All that being said I would reformulate the comment as follows:
> > 
> > 	* Note: the function may return non-NULL struct page even for a
> > 	* page block which contains a memory hole (i.e. there is no
> > 	* physical memory for a subset of the pfn range). This should be
> > 	* safe most of the time because struct pages are still zero
> > 	* pre-filled and pfn walkers shouldn't touch any physical memory
> > 	* range for which they do not recognize any specific metadata in
> > 	* struct pages.
> 
> Thanks. That makes sense to me. A trivial thing is I still want to add the
> example in the comments to make it clear. Are you okay with below
> description?
> 
> + * Note: the function may return non-NULL struct page even for a page block
> + * which contains a memory hole (i.e. there is no physical memory for a
> subset
> + * of the pfn range). For example, if the pageblock order is MAX_ORDER,
> which
> + * will fall into 2 sub-sections, and the end pfn of the pageblock may be
> hole
> + * even though the start pfn is online and valid. This should be safe most
> of
> + * the time because struct pages are still zero pre-filled and pfn walkers
> + * shouldn't touch any physical memory range for which they do not
> recognize
> + * any specific metadata in struct pages.

No objections of course. I do not see an additional value, quite
honestly but if somebody does then it doesn't hurt.

-- 
Michal Hocko
SUSE Labs