Re: [bug report] mm: avoid leaving partial pfn mappings around in error case

[Lorenzo Stoakes <lorenzo.stoakes@oracle.com>]

On Sun, Sep 15, 2024 at 04:14:21PM GMT, Dan Carpenter wrote:
> On Sun, Sep 15, 2024 at 01:38:40PM +0100, Lorenzo Stoakes wrote:
> > + get_maintainers.pl people for drivers/misc/sgi-gru/grumain.c
> >
> > On Sun, Sep 15, 2024 at 03:09:35PM GMT, Dan Carpenter wrote:
> > > On Sun, Sep 15, 2024 at 01:01:43PM +0100, Lorenzo Stoakes wrote:
> > > > On Sun, Sep 15, 2024 at 01:08:27PM GMT, Dan Carpenter wrote:
> > > > > Hi Linus,
> > > > >
> > > > > Commit 79a61cc3fc04 ("mm: avoid leaving partial pfn mappings around in
> > > > > error case") from Sep 11, 2024 (linux-next), leads to the following
> > > > > Smatch static checker warning:
> > > > >
> > > > > 	mm/memory.c:2709 remap_pfn_range_notrack()
> > > > > 	warn: sleeping in atomic context
> > > > >
> > > > > mm/memory.c
> > > > >     2696 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
> > > > >     2697                 unsigned long pfn, unsigned long size, pgprot_t prot)
> > > > >     2698 {
> > > > >     2699         int error = remap_pfn_range_internal(vma, addr, pfn, size, prot);
> > > > >     2700
> > > > >     2701         if (!error)
> > > > >     2702                 return 0;
> > > > >     2703
> > > > >     2704         /*
> > > > >     2705          * A partial pfn range mapping is dangerous: it does not
> > > > >     2706          * maintain page reference counts, and callers may free
> > > > >     2707          * pages due to the error. So zap it early.
> > > > >     2708          */
> > > > > --> 2709         zap_page_range_single(vma, addr, size, NULL);
> > > > >
> > > > > The lru_add_drain() function at the start of zap_page_range_single() takes a
> > > > > mutext.
> > > >
> > > > Hm does it? I see a local lock, and some folio batch locking which are
> > > > local locks too?
> > >
> > > Ah...  No it doesn't.  It's the mmu_notifier_invalidate_range_start() which is
> > > a might_sleep() function.  Sorry for the confusion.
> >
> > OK so in conclusion it seems to be that Linus's commit introducing
> > zap_page_range_single() accidentally had smatch hit a might_sleep() via
> > mmu_notifier_invalidate_range_start(), but it should, in theory, have fired
> > due to page table allocations invoking the page allocator that might sleep,
> > but didn't, because smatch misses the below might_alloc() path...
> >
> > -> prepare_alloc_pages()
> >    -> might_alloc()
> >       -> might_sleep_if(gfpflags_allow_blocking(gfp_mask))
> >
> > ...as a result of get_zeroed_page() tripping it up *breathes*. :)
> >
> > (please correct me if I am wrong here).
>
> That's an accurate summary...

Thanks!

>
> >
> > The preempt_disable() is introduced in commit fe5bb6b00c3a9 ("sgi-gru: misc
> > GRU cleanup") from... 2009, but it fixed it from the far far more broken
> > 'disable preemption before taking a mutex' situation that existed before.
> >
> > So fix seems to me to not invoke remap_pfn_range() with preemption disabled
> > and a mutex held? gru_fault() maintainers added for input...
>
> Every time I get a response to this bug report I feel dumber.  How did I not
> see that this was a bug in drivers/misc/sgi-gru/?  Here is another one from the
> same driver:
>
> drivers/misc/sgi-gru/grukservices.c:262 gru_get_cpu_resources() warn: sleeping in atomic context

Nothing to feel dumb about, this stuff is confounding by nature, if I had a
penny for every time I felt dumb doing kernel work I'd be very rich by now! ;)

>
> regards,
> dan carpenter

Cheers for report! This means we can now get this thing fixed...