mm: honor FOLL_GET flag in follow_hugetlb_page
Subsystems (2)
PATCH 1/3 mm: honor FOLL_GET flag in follow_hugetlb_page
From: Jerome Glisse <[email protected]>
Do not increase page count if FOLL_GET is not set.
--- mm/hugetlb.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
mm/hugetlb.c +3 -1
@@ -2991,7 +2991,9 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, same_page:if (pages) { pages[i] =mem_map_offset (page, pfn_offset); -get_page (pages[i]); +if (flags &FOLL_GET ) { +get_page_foll (pages[i]); + } }if (vmas)
Signature · 7 lines
1.8.2.1 -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 15 lines · On Tue, May 07, 2013 at 04:45:54PM -0400, [email protected] wrote:
From: Jerome Glisse <[email protected]>
Do not increase page count if
FOLL_GETis not set.Signed-off-by Jerome Glisse <[email protected]>--- mm/hugetlb.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)mm/hugetlb.c +0
-0
mm/hugetlb.c long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, · line 2999
Quoted 10 lines
mm/hugetlb.c +3
-1@@ -2991,7 +2991,9 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, same_page:if (pages) { pages[i] =mem_map_offset (page, pfn_offset); -get_page (pages[i]); +if (flags &FOLL_GET ) { +get_page_foll (pages[i]); + } }if (vmas)
Hi Jerome,
I think that we need to be careful in handling tail pages, because __get_page_tail_foll() uses page->_mapcount as refcount. When you get refcount on a tail page and free the hugepage without putting the *mapcount*, you will hit BUG_ON() in free_huge_page(). Yes, this is a very tricky workaround for thp, so to avoid making things too complicated, I think either of the following is better:
- to get refcount only for head pages, or
- to introduce a hugetlbfs variant of get_page_foll().
BTW, who do you expect is the caller of follow_hugetlb_page() with FOLL_GET (I can't find your subsequent patches 2/3 or 3/3)? I'm interested in this change because in my project it's necessary to implement this for hugepage migration
(see https://lkml.org/lkml/2013/3/22/553).
Thanks,
Naoya Horiguchi
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 29 lines · Naoya Horiguchi
Quoted 15 lines · On Tue, May 07, 2013 at 04:45:54PM -0400, [email protected] wrote:
From: Jerome Glisse <[email protected]>
Do not increase page count if
FOLL_GETis not set.Signed-off-by Jerome Glisse <[email protected]>--- mm/hugetlb.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)mm/hugetlb.c +0
-0@@ -2991,7 +2991,9 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, same_page:mm/hugetlb.c long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
Quoted 10 lines
mm/hugetlb.c +3
-2if (pages) { pages[i] =mem_map_offset (page, pfn_offset); -get_page (pages[i]); +if (flags &FOLL_GET ) { +get_page_foll (pages[i]); + } }if (vmas) --Hi Jerome,
Quoted 8 lines
I think that we need to be careful in handling tail pages, because
__get_page_tail_foll()usespage->_mapcountas refcount. When you get refcount on a tail page and free the hugepage without putting the *mapcount*, you will hitBUG_ON()infree_huge_page(). Yes, this is a very tricky workaround for thp, so to avoid making things too complicated, I think either of the following is better:
- to get refcount only for head pages, or
- to introduce a hugetlbfs variant ofget_page_foll().
Maybe a simpler variant is to just not take any refcount, ie like current code if FOLL_GET is set then take refcount on all page wether they are head/tail or not. I will resend with that.
Quoted 5 lines
BTW, who do you expect is the caller of
follow_hugetlb_page()withFOLL_GET(I can't find your subsequent patches 2/3 or 3/3)? I'm interested in this change because in my project it's necessary to implement this for hugepage migration
(see https://lkml.org/lkml/2013/3/22/553).
I can not talk about the patchset yet (and it's not fully cook) but i need to be able to get the page without taking reference so without the FOLL_GET flag set but i need splitting, well no real splitting, i need pfn for each fake sub page of huge page (interested in physical address not in the page struct).
Cheers,
Jerome
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 40 lines · Jerome Glisse
Quoted 29 lines · Naoya Horiguchi
Quoted 15 lines · On Tue, May 07, 2013 at 04:45:54PM -0400, [email protected] wrote:
From: Jerome Glisse <[email protected]>
Do not increase page count if
FOLL_GETis not set.Signed-off-by Jerome Glisse <[email protected]>--- mm/hugetlb.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)mm/hugetlb.c +0
-0@@ -2991,7 +2991,9 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, same_page:mm/hugetlb.c long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
Quoted 10 lines
mm/hugetlb.c +3
-2if (pages) { pages[i] =mem_map_offset (page, pfn_offset); -get_page (pages[i]); +if (flags &FOLL_GET ) { +get_page_foll (pages[i]); + } }if (vmas) --Hi Jerome,
Quoted 8 lines
I think that we need to be careful in handling tail pages, because
__get_page_tail_foll()usespage->_mapcountas refcount. When you get refcount on a tail page and free the hugepage without putting the *mapcount*, you will hitBUG_ON()infree_huge_page(). Yes, this is a very tricky workaround for thp, so to avoid making things too complicated, I think either of the following is better:
- to get refcount only for head pages, or
- to introduce a hugetlbfs variant ofget_page_foll().
Quoted 3 lines
Maybe a simpler variant is to just not take any refcount, ie like current code if
FOLL_GETis set then take refcount on all page wether they are head/tail or not. I will resend with that.
Hmm, I think that FOLL_GET flag means "do get_page on page", so the "not take any refcount" variant seems to make no sense for me. Would it be better just call without FOLL_GET?
Quoted 11 lines
Quoted 5 lines
BTW, who do you expect is the caller of
follow_hugetlb_page()withFOLL_GET(I can't find your subsequent patches 2/3 or 3/3)? I'm interested in this change because in my project it's necessary to implement this for hugepage migration
(see https://lkml.org/lkml/2013/3/22/553).I can not talk about the patchset yet (and it's not fully cook) but i need to be able to get the page without taking reference so without the
FOLL_GETflag set but i need splitting, well no real splitting, i need pfn for each fake sub page of huge page (interested in physical address not in the page struct).
If the caller knows the vma is backed by hugetlbfs, we can get pfn of the tail page by adding page offset (which can be calcurated by the virtual address with proper huge page mask) to the head page's pfn.
Thanks,
Naoya Horiguchi
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 61 lines · Naoya Horiguchi
Quoted 40 lines · Jerome Glisse
Quoted 29 lines · Naoya Horiguchi
Quoted 15 lines · On Tue, May 07, 2013 at 04:45:54PM -0400, [email protected] wrote:
From: Jerome Glisse <[email protected]>
Do not increase page count if
FOLL_GETis not set.Signed-off-by Jerome Glisse <[email protected]>--- mm/hugetlb.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)mm/hugetlb.c +0
-0@@ -2991,7 +2991,9 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, same_page:mm/hugetlb.c long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
Quoted 10 lines
mm/hugetlb.c +3
-2if (pages) { pages[i] =mem_map_offset (page, pfn_offset); -get_page (pages[i]); +if (flags &FOLL_GET ) { +get_page_foll (pages[i]); + } }if (vmas) --Hi Jerome,
Quoted 8 lines
I think that we need to be careful in handling tail pages, because
__get_page_tail_foll()usespage->_mapcountas refcount. When you get refcount on a tail page and free the hugepage without putting the *mapcount*, you will hitBUG_ON()infree_huge_page(). Yes, this is a very tricky workaround for thp, so to avoid making things too complicated, I think either of the following is better:
- to get refcount only for head pages, or
- to introduce a hugetlbfs variant ofget_page_foll().Quoted 3 lines
Maybe a simpler variant is to just not take any refcount, ie like current code if
FOLL_GETis set then take refcount on all page wether they are head/tail or not. I will resend with that.Hmm, I think that
FOLL_GETflag means "doget_pageon page", so the "not take any refcount" variant seems to make no sense for me. Would it be better just call withoutFOLL_GET?Quoted 11 lines
Quoted 5 lines
BTW, who do you expect is the caller of
follow_hugetlb_page()withFOLL_GET(I can't find your subsequent patches 2/3 or 3/3)? I'm interested in this change because in my project it's necessary to implement this for hugepage migration
(see https://lkml.org/lkml/2013/3/22/553).I can not talk about the patchset yet (and it's not fully cook) but i need to be able to get the page without taking reference so without the
FOLL_GETflag set but i need splitting, well no real splitting, i need pfn for each fake sub page of huge page (interested in physical address not in the page struct).
Quoted 3 lines
If the caller knows the vma is backed by hugetlbfs, we can get pfn of the tail page by adding page offset (which can be calcurated by the virtual address with proper huge page mask) to the head page's pfn.
Yes, i am just lazy in my code on that front. But i probably should do that anyway, it will save me some temporary allocation.
Cheers,
Jerome
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>