Re: a question of split_huge_page

Re: a question of split_huge_page

[Alex Shi]

在 2020/7/10 下午1:28, Mika Penttilä 写道:
> 
> 
> On 10.7.2020 7.51, Alex Shi wrote:
>>
>> 在 2020/7/10 上午12:07, Kirill A. Shutemov 写道:
>>> On Thu, Jul 09, 2020 at 04:50:02PM +0100, Matthew Wilcox wrote:
>>>> On Thu, Jul 09, 2020 at 11:11:11PM +0800, Alex Shi wrote:
>>>>> Hi Kirill & Matthew,
>>>>>
>>>>> In the func call chain, from split_huge_page() to lru_add_page_tail(),
>>>>> Seems tail pages are added to lru list at line 963, but in this scenario
>>>>> the head page has no lru bit and isn't set the bit later. Why we do this?
>>>>> or do I miss sth?
>>>> I don't understand how we get to split_huge_page() with a page that's
>>>> not on an LRU list.  Both anonymous and page cache pages should be on
>>>> an LRU list.  What am I missing?> 
>>
>> Thanks a lot for quick reply!
>> What I am confusing is the call chain: __iommu_dma_alloc_pages()
>> to split_huge_page(), in the func, splited page,
>> 	page = alloc_pages_node(nid, alloc_flags, order);
>> And if the pages were added into lru, they maybe reclaimed and lost,
>> that would be a panic bug. But in fact, this never happened for long time.
>> Also I put a BUG() at the line, it's nevre triggered in ltp, and run_vmtests
> 
> 
> In  __iommu_dma_alloc_pages, after split_huge_page(),  who is taking a
> reference on tail pages? Seems tail pages are freed and the function
> errornously returns them in pages[] array for use?
> 

CC Joerg and iommu list,

That's a good question. seems the split_huge_page was never triggered here,
since the func would check the PageLock first. and have page->mapping and PageAnon
check, any of them couldn't be matched for the alloced page.

Hi Joerg,
would you like look into this? do we still need the split_huge_page() here?

Thanks
Alex

int split_huge_page_to_list(struct page *page, struct list_head *list)
{
        struct page *head = compound_head(page);
        struct deferred_split *ds_queue = get_deferred_split_queue(head);
        struct anon_vma *anon_vma = NULL;
        struct address_space *mapping = NULL;
        int count, mapcount, extra_pins, ret;
        pgoff_t end;

        VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
        VM_BUG_ON_PAGE(!PageLocked(head), head);	<==
> 
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu

Re: a question of split_huge_page

[Joerg Roedel]

Adding Robin.

On Fri, Jul 10, 2020 at 05:34:52PM +0800, Alex Shi wrote:
> 在 2020/7/10 下午1:28, Mika Penttilä 写道:
> > 
> > 
> > On 10.7.2020 7.51, Alex Shi wrote:
> >>
> >> 在 2020/7/10 上午12:07, Kirill A. Shutemov 写道:
> >>> On Thu, Jul 09, 2020 at 04:50:02PM +0100, Matthew Wilcox wrote:
> >>>> On Thu, Jul 09, 2020 at 11:11:11PM +0800, Alex Shi wrote:
> >>>>> Hi Kirill & Matthew,
> >>>>>
> >>>>> In the func call chain, from split_huge_page() to lru_add_page_tail(),
> >>>>> Seems tail pages are added to lru list at line 963, but in this scenario
> >>>>> the head page has no lru bit and isn't set the bit later. Why we do this?
> >>>>> or do I miss sth?
> >>>> I don't understand how we get to split_huge_page() with a page that's
> >>>> not on an LRU list.  Both anonymous and page cache pages should be on
> >>>> an LRU list.  What am I missing?> 
> >>
> >> Thanks a lot for quick reply!
> >> What I am confusing is the call chain: __iommu_dma_alloc_pages()
> >> to split_huge_page(), in the func, splited page,
> >> 	page = alloc_pages_node(nid, alloc_flags, order);
> >> And if the pages were added into lru, they maybe reclaimed and lost,
> >> that would be a panic bug. But in fact, this never happened for long time.
> >> Also I put a BUG() at the line, it's nevre triggered in ltp, and run_vmtests
> > 
> > 
> > In  __iommu_dma_alloc_pages, after split_huge_page(),  who is taking a
> > reference on tail pages? Seems tail pages are freed and the function
> > errornously returns them in pages[] array for use?
> > 
> 
> CC Joerg and iommu list,
> 
> That's a good question. seems the split_huge_page was never triggered here,
> since the func would check the PageLock first. and have page->mapping and PageAnon
> check, any of them couldn't be matched for the alloced page.
> 
> Hi Joerg,
> would you like look into this? do we still need the split_huge_page() here?
> 
> Thanks
> Alex
> 
> int split_huge_page_to_list(struct page *page, struct list_head *list)
> {
>         struct page *head = compound_head(page);
>         struct deferred_split *ds_queue = get_deferred_split_queue(head);
>         struct anon_vma *anon_vma = NULL;
>         struct address_space *mapping = NULL;
>         int count, mapcount, extra_pins, ret;
>         pgoff_t end;
> 
>         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
>         VM_BUG_ON_PAGE(!PageLocked(head), head);	<==
> > 
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu

Re: a question of split_huge_page

[Yang Shi]

On Fri, Jul 10, 2020 at 2:35 AM Alex Shi <alex.shi@linux.alibaba.com> wrote:
>
> 在 2020/7/10 下午1:28, Mika Penttilä 写道:
> >
> >
> > On 10.7.2020 7.51, Alex Shi wrote:
> >>
> >> 在 2020/7/10 上午12:07, Kirill A. Shutemov 写道:
> >>> On Thu, Jul 09, 2020 at 04:50:02PM +0100, Matthew Wilcox wrote:
> >>>> On Thu, Jul 09, 2020 at 11:11:11PM +0800, Alex Shi wrote:
> >>>>> Hi Kirill & Matthew,
> >>>>>
> >>>>> In the func call chain, from split_huge_page() to lru_add_page_tail(),
> >>>>> Seems tail pages are added to lru list at line 963, but in this scenario
> >>>>> the head page has no lru bit and isn't set the bit later. Why we do this?
> >>>>> or do I miss sth?
> >>>> I don't understand how we get to split_huge_page() with a page that's
> >>>> not on an LRU list.  Both anonymous and page cache pages should be on
> >>>> an LRU list.  What am I missing?>
> >>
> >> Thanks a lot for quick reply!
> >> What I am confusing is the call chain: __iommu_dma_alloc_pages()
> >> to split_huge_page(), in the func, splited page,
> >>      page = alloc_pages_node(nid, alloc_flags, order);
> >> And if the pages were added into lru, they maybe reclaimed and lost,
> >> that would be a panic bug. But in fact, this never happened for long time.
> >> Also I put a BUG() at the line, it's nevre triggered in ltp, and run_vmtests
> >
> >
> > In  __iommu_dma_alloc_pages, after split_huge_page(),  who is taking a
> > reference on tail pages? Seems tail pages are freed and the function
> > errornously returns them in pages[] array for use?
> >
>
> CC Joerg and iommu list,
>
> That's a good question. seems the split_huge_page was never triggered here,
> since the func would check the PageLock first. and have page->mapping and PageAnon
> check, any of them couldn't be matched for the alloced page.
>
> Hi Joerg,
> would you like look into this? do we still need the split_huge_page() here?

I think this is the same problem which has been discussed a couple of
weeks ago. Please refer to:
https://lore.kernel.org/linux-mm/20200619001938.GA135965@carbon.dhcp.thefacebook.com/

I think the conclusion is split_huge_page() can't be used in this path
at all. But we didn't reach a fix yet.

>
> Thanks
> Alex
>
> int split_huge_page_to_list(struct page *page, struct list_head *list)
> {
>         struct page *head = compound_head(page);
>         struct deferred_split *ds_queue = get_deferred_split_queue(head);
>         struct anon_vma *anon_vma = NULL;
>         struct address_space *mapping = NULL;
>         int count, mapcount, extra_pins, ret;
>         pgoff_t end;
>
>         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
>         VM_BUG_ON_PAGE(!PageLocked(head), head);        <==
> >
>
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu