Re: [mm/gup] 9857a17f20: kernel_BUG_at_include/linux/pagemap.h

[John Hubbard]

On 9/7/21 11:14 AM, Linus Torvalds wrote:
> On Tue, Sep 7, 2021 at 8:20 AM kernel test robot <oliver.sang@xxxxxxxxx> wrote:
> > FYI, we noticed the following commit (built with clang-14):
> >
> > commit: 9857a17f206f ("mm/gup: remove try_get_page(), call try_get_compound_head() directly")
> > caused below changes (please refer to attached dmesg/kmsg for entire log/backtrace):
> >
> > [  143.908513][ T3260] kernel BUG at include/linux/pagemap.h:223!
>
> Ahh, well, yes.
>
> That commit is clearly buggy, in that the try_get_compound_head() code
> really doesn't work at all for us.
>
> __page_cache_add_speculative() is not at all the same as
> try_get_page(), and I should have caught on to this as I applied it. I
> just read the explanation, and it sounded believable, but it was
> entirely wrong.
>
> try_get_page() is literally about that "page ref overflow" case, but
> try_get_compound_head() uses page_cache_add_speculative() which has
> different logic and has those extra "this only works in RCU context"
> logic.
>
> So that commit was completely bogus, and the "lack of maintenance" was
> not lack of maintenance at all, it was all about entirely different
> semantics.
>
> Reverted.
>
>                   Linus

Apologies for the bug! There is a lesson in here, somewhere...


thanks,
--
John Hubbard
NVIDIA