mm/swap: Fix release_pages() when releasing devmap pages
PATCH v2 1/1 mm/swap: Fix release_pages() when releasing devmap pages
From: Ira Weiny <[email protected]>
Device pages can be more than type MEMORY_DEVICE_PUBLIC.
Handle all device pages within release_pages()
This was found via code inspection while determining if release_pages() and the new put_user_pages() could be interchangeable.
---
Changes from V1:
Add comment clarifying that put_devmap_managed_page() can still fail.
Add Reviewed-by tags.
mm/swap.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)
mm/swap.c +7 -4
@@ -740,15 +740,18 @@ void release_pages(struct page **pages, int nr)if (is_huge_zero_page (page))continue ; -/* Device public page can not be huge page */ -if (is_device_public_page (page)) { +if (is_zone_device_page (page)) {if (locked_pgdat) {spin_unlock_irqrestore (&locked_pgdat->lru_lock, flags); locked_pgdat =NULL ; } -put_devmap_managed_page (page); -continue ; +/* +* zone-device-pages can still fail here and will +* therefore need put_page_testzero() +*/ +if (put_devmap_managed_page (page)) +continue ; } page =compound_head (page);
Signature · 1 line
2.20.1
Quoted 8 lines · On Fri 24-05-19 10:36:56, [email protected] wrote:
From: Ira Weiny <[email protected]>
Device pages can be more than type
MEMORY_DEVICE_PUBLIC.Handle all device pages within
release_pages()This was found via code inspection while determining if
release_pages()and the newput_user_pages()could be interchangeable.
Please expand more about who is such a user and why does it use release_pages rather than put_*page API. The above changelog doesn't really help understanding what is the actual problem. I also do not understand the fix and a failure mode from release_pages is just scary. It is basically impossible to handle the error case. So what is going on here?
Quoted 45 lines
Cc Jérôme Glisse <[email protected]>Cc Michal Hocko <[email protected]>Reviewed-by Dan Williams <[email protected]>Reviewed-by John Hubbard <[email protected]>Signed-off-by Ira Weiny <[email protected]>---
Changes from V1:
Add comment clarifying thatput_devmap_managed_page()can still fail.
Add Reviewed-by tags.mm/swap.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)mm/swap.c +7
-4@@ -740,15 +740,18 @@ void release_pages(struct page **pages, int nr)if (is_huge_zero_page (page))continue ; -/* Device public page can not be huge page */ -if (is_device_public_page (page)) { +if (is_zone_device_page (page)) {if (locked_pgdat) {spin_unlock_irqrestore (&locked_pgdat->lru_lock, flags); locked_pgdat =NULL ; } -put_devmap_managed_page (page); -continue ; +/* +* zone-device-pages can still fail here and will +* therefore need put_page_testzero() +*/ +if (put_devmap_managed_page (page)) +continue ; } page =compound_head (page);Signature · 1 line
2.20.1
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 5 lines · Michal Hocko
Quoted 4 lines · On Fri 24-05-19 10:36:56, [email protected] wrote:
From: Ira Weiny <[email protected]>
Device pages can be more than type
MEMORY_DEVICE_PUBLIC.
Quoted 7 lines
Quoted 4 lines
Handle all device pages within
release_pages()This was found via code inspection while determining if
release_pages()and the newput_user_pages()could be interchangeable.Please expand more about who is such a user and why does it use
release_pagesrather than put_*page API.
Sorry for not being more clear. The error was discovered while discussing a
proposal to change a use of release_pages() to put_user_pages()[1]
[1] https://lore.kernel.org/lkml/[email protected]/
In that thread John was saying that release_pages() was functionally equivalent to a loop around put_page(). He also suggested implementing put_user_pages() by using release_pages(). On the surface they did not seem the same to me so I did a deep dive to make sure they were and found this error.
Quoted 4 lines
The above changelog doesn't
really help understanding what is the actual problem. I also do not understand the fix and a failure mode fromrelease_pagesis just scary.
This is not failing release_pages(). The fix is that not all devmap pages are "public" type. So previous to this change devmap pages of other types would not correctly be accounted for.
The discussion about put_devmap_managed_page() "failing" is not about it failing directly but rather in how these pages should be accounted for. Only devmap pages which require pagemap ops (specifically page_free()) requireput_devmap_managed_page() processing. Because of the optimized locking inrelease_pages() the zone device check is required to release the lock even if put_devmap_managed_page() does not handle the put.
Quoted 2 lines
It is basically impossible to handle the error case. So what is going on here?
I think what has happened is the code in release_pages() and put_page() diverged at some point. I think it is worth a clean up in this area but I don't see way to do it at the moment which would be any cleaner than what is there. So I've refrained from doing so.
Does this help? Would you like to roll a V3 with some of this in the commit message?
Ira
Quoted 53 lines
Quoted 49 lines
Cc Jérôme Glisse <[email protected]>Cc Michal Hocko <[email protected]>Reviewed-by Dan Williams <[email protected]>Reviewed-by John Hubbard <[email protected]>Signed-off-by Ira Weiny <[email protected]>---
Changes from V1:
Add comment clarifying thatput_devmap_managed_page()can still fail.
Add Reviewed-by tags.mm/swap.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)mm/swap.c +7
-4@@ -740,15 +740,18 @@ void release_pages(struct page **pages, int nr)if (is_huge_zero_page (page))continue ; -/* Device public page can not be huge page */ -if (is_device_public_page (page)) { +if (is_zone_device_page (page)) {if (locked_pgdat) {spin_unlock_irqrestore (&locked_pgdat->lru_lock, flags); locked_pgdat =NULL ; } -put_devmap_managed_page (page); -continue ; +/* +* zone-device-pages can still fail here and will +* therefore need put_page_testzero() +*/ +if (put_devmap_managed_page (page)) +continue ; } page =compound_head (page);Signature · 1 line
2.20.1Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 2 lines · Ira Weiny
Would you like to roll a V3 with some of this in the commit message?
Yes please. I will re-read the whole changelog and let's see whether I can make more sense of it.
Thanks!
Signature · 2 lines
Michal Hocko SUSE Labs