Re: [PATCH] mm/page_alloc: Fix freeing of failed-split poisoned compound pages

[Boudewijn van der Heide <boudewijn@delta-utec.com>]

> > free_pages_prepare() only handles poisoned order-0 pages.
> > In memory_failure() (hard offline), pages
> > are poisoned before attempting to split huge pages. If the split fails,
> > the page remains a compound (order > 0) but is already poisoned. However,
> > Soft-offline pages are always poisoned as order-0 after migration, so
> > they are unaffected.
> >
> > The '!order' check causes these poisoned compound pages to skip
> > poison handling, leaving them in the buddy allocator.
> >
> > Worst case, a poisoned compound page could be reallocated,
> > potentially leading to crashes, silent data corruption,
> > or unwanted memory containment actions before the poison bit is detected.
> >
> > This patch removes the '&& !order' restriction. Cleanup functions in the
> > poison-handling block correctly handle non-zero order pages, making
> > this change safe.

> This is not a fix. IIUC, for >0 order free pages, memory failure uses
> take_page_off_buddy() in a different code path.
>

Thanks again for the quick response and clarification! 
From my understanding, 
you correctly noted that take_page_off_buddy() handles already-free pages, 
removing them from the buddy lists and setting SetPageHWPoisonTakenOff(). 
This prevents those pages from re-entering the buddy allocator.

My concern is about in-use THP-backed compound pages:
1. A compound page is in use. 
2. memory_failure() marks it poisoned (TestSetPageHWPoison).
3. try_to_split_thp_page() fails.
4. The process using the THP may be killed; 
        the page remains compound and poisoned.
5. Later, when the page is finally freed, it reaches free_pages_prepare();
        'take_page_off_buddy()' is not invoked in this path.

At this point, the current check:

'if (unlikely(PageHWPoison(page)) && !order)'

will not trigger, because the order > 0.

> Miaohe (cc’d) should be able to elaborate more on it.
Thanks for Cc'ing Miaohe, hopefully Miaohe can provide some more insights!

Thanks,
Boudewijn