Re: [External] Re: [RFC PATCH 00/24] mm/hugetlb: Free some vmemmap pages of hugetlb page

[Matthew Wilcox]

On Wed, Sep 16, 2020 at 02:03:15AM +0800, Muchun Song wrote:
> On Wed, Sep 16, 2020 at 1:39 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >
> > On Wed, Sep 16, 2020 at 01:32:46AM +0800, Muchun Song wrote:
> > > On Tue, Sep 15, 2020 at 11:42 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > >
> > > > On Tue, Sep 15, 2020 at 11:28:01PM +0800, Muchun Song wrote:
> > > > > On Tue, Sep 15, 2020 at 10:32 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Tue, Sep 15, 2020 at 08:59:23PM +0800, Muchun Song wrote:
> > > > > > > This patch series will free some vmemmap pages(struct page structures)
> > > > > > > associated with each hugetlbpage when preallocated to save memory.
> > > > > >
> > > > > > It would be lovely to be able to do this.  Unfortunately, it's completely
> > > > > > impossible right now.  Consider, for example, get_user_pages() called
> > > > > > on the fifth page of a hugetlb page.
> > > > >
> > > > > Can you elaborate on the problem? Thanks so much.
> > > >
> > > > OK, let's say you want to do a 2kB I/O to offset 0x5000 of a 2MB page
> > > > on a 4kB base page system.  Today, that results in a bio_vec containing
> > > > {head+5, 0, 0x800}.  Then we call page_to_phys() on that (head+5) struct
> > > > page to get the physical address of the I/O, and we turn it into a struct
> > > > scatterlist, which similarly has a reference to the page (head+5).
> > >
> > > As I know, in this case, the get_user_pages() will get a reference
> > > to the head page (head+0) before returning such that the hugetlb
> > > page can not be freed. Although get_user_pages() returns the
> > > page (head+5) and the scatterlist has a reference to the page
> > > (head+5), this patch series can handle this situation. I can not
> > > figure out where the problem is. What I missed? Thanks.
> >
> > You freed pages 4-511 from the vmemmap so they could be used for
> > something else.  Page 5 isn't there any more.  So if you return head+5,
> > then when we complete the I/O, we'll look for the compound_head() of
> > head+5 and we won't find head.
> 
> We do not free pages 4-511 from the vmemmap. Actually, we only
> free pages 128-511 from the vmemmap.
> 
> The 512 struct pages occupy 8 pages of physical memory. We only
> free 6 physical page frames to the buddy. But we will create a new
> mapping just like below. The virtual address of the freed pages will
> remap to the second page frame. So the second page frame is
> reused.

Oh!  I get what you're doing now.

For the vmemmap case, you free the last N-2 physical pages but map the
second physical page multiple times.  So for the 512 pages case, we
see pages:

abcdefgh | ijklmnop | ijklmnop | ijklmnop | ijklmnop | ijklmnop | ijklmnop ...

Huh.  I think that might work, except for PageHWPoison.  I'll go back
to your patch series and look at that some more.