Re: [PATCH uprobe, thp 3/4] uprobe: support huge page by only splitting the pmd

[Song Liu]

> On May 30, 2019, at 5:14 AM, Kirill A. Shutemov  wrote:
> 
> On Wed, May 29, 2019 at 02:20:48PM -0700, Song Liu wrote:
>> Instead of splitting the compound page with FOLL_SPLIT, this patch allows
>> uprobe to only split pmd for huge pages.
>> 
>> A helper function mm_address_trans_huge(mm, address) was introduced to
>> test whether the address in mm is pointing to THP.
> 
> Maybe it would be cleaner to have FOLL_SPLIT_PMD which would strip
> trans_huge PMD if any and then set pte using get_locked_pte()?

FOLL_SPLIT_PMD sounds like a great idea! Let me try it. 

Thanks,
Song

> 
> This way you'll not need any changes in split_huge_pmd() path. Clearing
> PMD will be fine.
> 
> -- 
> Kirill A. Shutemov

Re: [PATCH uprobe, thp 3/4] uprobe: support huge page by only splitting the pmd

[Song Liu]

> On May 30, 2019, at 4:08 AM, William Kucharski  
> wrote:
> 
> 
> Is there any reason to worry about supporting PUD-sized uprobe pages if
> CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD is defined? I would prefer
> not to bake in the assumption that "huge" means PMD-sized and more than
> it already is.
> 
> For example, if CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD is configured,
> mm_address_trans_huge() should either make the call to pud_trans_huge()
> if appropriate, or a VM_BUG_ON_PAGE should be added in case the routine
> is ever called with one.
> 
> Otherwise it looks pretty reasonable to me.
> 
>-- Bill
> 

I will try that in v2. 

Thanks,
Song

Re: [PATCH uprobe, thp 3/4] uprobe: support huge page by only splitting the pmd

[Kirill A. Shutemov]

On Wed, May 29, 2019 at 02:20:48PM -0700, Song Liu wrote:
> Instead of splitting the compound page with FOLL_SPLIT, this patch allows
> uprobe to only split pmd for huge pages.
> 
> A helper function mm_address_trans_huge(mm, address) was introduced to
> test whether the address in mm is pointing to THP.

Maybe it would be cleaner to have FOLL_SPLIT_PMD which would strip
trans_huge PMD if any and then set pte using get_locked_pte()?

This way you'll not need any changes in split_huge_pmd() path. Clearing
PMD will be fine.

-- 
 Kirill A. Shutemov

Re: [PATCH uprobe, thp 3/4] uprobe: support huge page by only splitting the pmd

[William Kucharski]

Is there any reason to worry about supporting PUD-sized uprobe pages if
CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD is defined? I would prefer
not to bake in the assumption that "huge" means PMD-sized and more than
it already is.

For example, if CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD is configured,
mm_address_trans_huge() should either make the call to pud_trans_huge()
if appropriate, or a VM_BUG_ON_PAGE should be added in case the routine
is ever called with one.

Otherwise it looks pretty reasonable to me.

-- Bill