Re: [PATCH] mm: hugepage: mark splitted page dirty when needed

[Peter Xu <peterx@redhat.com>]

On Wed, Sep 05, 2018 at 08:49:20AM -0400, Zi Yan wrote:
> On 5 Sep 2018, at 3:30, Peter Xu wrote:
> 
> > On Tue, Sep 04, 2018 at 10:00:28AM -0400, Zi Yan wrote:
> >> On 4 Sep 2018, at 4:01, Kirill A. Shutemov wrote:
> >>
> >>> On Tue, Sep 04, 2018 at 03:55:10PM +0800, Peter Xu wrote:
> >>>> When splitting a huge page, we should set all small pages as dirty if
> >>>> the original huge page has the dirty bit set before.  Otherwise we'll
> >>>> lose the original dirty bit.
> >>>
> >>> We don't lose it. It got transfered to struct page flag:
> >>>
> >>> 	if (pmd_dirty(old_pmd))
> >>> 		SetPageDirty(page);
> >>>
> >>
> >> Plus, when split_huge_page_to_list() splits a THP, its subroutine __split_huge_page()
> >> propagates the dirty bit in the head page flag to all subpages in __split_huge_page_tail().
> >
> > Hi, Kirill, Zi,
> >
> > Thanks for your responses!
> >
> > Though in my test the huge page seems to be splitted not by
> > split_huge_page_to_list() but by explicit calls to
> > change_protection().  The stack looks like this (again, this is a
> > customized kernel, and I added an explicit dump_stack() there):
> >
> >   kernel:  dump_stack+0x5c/0x7b
> >   kernel:  __split_huge_pmd+0x192/0xdc0
> >   kernel:  ? update_load_avg+0x8b/0x550
> >   kernel:  ? update_load_avg+0x8b/0x550
> >   kernel:  ? account_entity_enqueue+0xc5/0xf0
> >   kernel:  ? enqueue_entity+0x112/0x650
> >   kernel:  change_protection+0x3a2/0xab0
> >   kernel:  mwriteprotect_range+0xdd/0x110
> >   kernel:  userfaultfd_ioctl+0x50b/0x1210
> >   kernel:  ? do_futex+0x2cf/0xb20
> >   kernel:  ? tty_write+0x1d2/0x2f0
> >   kernel:  ? do_vfs_ioctl+0x9f/0x610
> >   kernel:  do_vfs_ioctl+0x9f/0x610
> >   kernel:  ? __x64_sys_futex+0x88/0x180
> >   kernel:  ksys_ioctl+0x70/0x80
> >   kernel:  __x64_sys_ioctl+0x16/0x20
> >   kernel:  do_syscall_64+0x55/0x150
> >   kernel:  entry_SYSCALL_64_after_hwframe+0x44/0xa9
> >
> > At the very time the userspace is sending an UFFDIO_WRITEPROTECT ioctl
> > to kernel space, which is handled by mwriteprotect_range().  In case
> > you'd like to refer to the kernel, it's basically this one from
> > Andrea's (with very trivial changes):
> >
> >   https://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git userfault
> >
> > So... do we have two paths to split the huge pages separately?
> >
> > Another (possibly very naive) question is: could any of you hint me
> > how the page dirty bit is finally applied to the PTEs?  These two
> > dirty flags confused me for a few days already (the SetPageDirty() one
> > which sets the page dirty flag, and the pte_mkdirty() which sets that
> > onto the real PTEs).
> 
> change_protection() only causes splitting a PMD entry into multiple PTEs
> but not the physical compound page, so my answer does not apply to your case.
> It is unclear how the dirty bit makes your QEMU get a SIGBUS. I think you
> need to describe your problem with more details.

Hi, Zi,

I explained with some more details on my problem in my other reply to
Kirill.  Please have a look.

> 
> AFAIK, the PageDirty bit will not apply back to any PTEs. So for your case,
> when reporting a pagea??s dirty bit information, some function in the kernel only checks
> the PTEa??s dirty bit but not the dirty bit in the struct page flags, which
> might provide a wrong answer.

Are you suggesting that we should always check both places (the PTE
dirty bit) and also the page flag to know whether a page is dirty
(hence, either of the bit set should mean the page is dirty)?

Thanks,

-- 
Peter Xu