mbox series

[v4,0/3] Cleanup for PAT

Message ID 20240220034835.507022-1-mawupeng1@huawei.com
Headers
Series Cleanup for PAT |

Message

mawupeng Feb. 20, 2024, 3:48 a.m. UTC 
  From: Ma Wupeng <mawupeng1@huawei.com>

Patch #1 move follow_phys to memtype.c since only pat use this.
Patch #2 cleanup parameter in follow_phys.
Patch #3 drop the unnecessary WARN_ON_ONCE if follow_phys fails.

Changelog since v3:
- rebase to latest linux
- fix compile warnings

Changelog since v2:
- rebase to latest linux

Changelog since v1:
- split patch #1 into two patches based on Boris's advise

Ma Wupeng (3):
  x86/mm/pat: Move follow_phys to pat-related file
  x86/mm/pat: Cleanup unused parameter in follow_phys
  x86/mm/pat: Remove WARN_ON_ONCE if follow_phys fails

 arch/x86/mm/pat/memtype.c | 32 ++++++++++++++++++++++++++------
 include/linux/mm.h        |  2 --
 mm/memory.c               | 28 ----------------------------
 3 files changed, 26 insertions(+), 36 deletions(-)

Comments

Xin Li (Intel) Feb. 20, 2024, 8:37 a.m. UTC | #1 
On 2/19/2024 7:48 PM, Wupeng Ma wrote:
> From: Ma Wupeng <mawupeng1@huawei.com>
> 

This patch set is all about follow_phys() cleanups, so "Cleanup for PAT"
seems too generic.

> Patch #1 move follow_phys to memtype.c since only pat use this.
> Patch #2 cleanup parameter in follow_phys.
> Patch #3 drop the unnecessary WARN_ON_ONCE if follow_phys fails.

I'm more curious why follow_phys() ended up this way?

follow_phys() was introduced in commit 28b2ee20c7cba ("access_process_vm
device memory infrastructure") in 2008 for getting a physical page address
for a virtual address, and used in generic_access_phys(). And later it's
used in x86 PAT code.

Commit 03668a4debf4f ("mm: use generic follow_pte() in follow_phys()") made
follow_phys() more of a wrapper of follow_pte(), and commit 96667f8a4382d
("mm: Close race in generic_access_phys") replaced follow_phys() with
follow_pte() in generic_access_phys(). And the end result is that
follow_phys() is used in x86 PAT code only.

As follow_phys() in untrack_pfn() can be replaced with follow_pfn(), then
maybe we don't have to keep follow_phys(), and just use follow_pte() in
track_pfn_copy()?

Thanks!
     Xin

> 
> Changelog since v3:
> - rebase to latest linux
> - fix compile warnings
> 
> Changelog since v2:
> - rebase to latest linux
> 
> Changelog since v1:
> - split patch #1 into two patches based on Boris's advise
> 
> Ma Wupeng (3):
>    x86/mm/pat: Move follow_phys to pat-related file
>    x86/mm/pat: Cleanup unused parameter in follow_phys
>    x86/mm/pat: Remove WARN_ON_ONCE if follow_phys fails
> 
>   arch/x86/mm/pat/memtype.c | 32 ++++++++++++++++++++++++++------
>   include/linux/mm.h        |  2 --
>   mm/memory.c               | 28 ----------------------------
>   3 files changed, 26 insertions(+), 36 deletions(-)
>
mawupeng Feb. 20, 2024, 9:06 a.m. UTC | #2 
On 2024/2/20 16:37, Xin Li wrote:
> On 2/19/2024 7:48 PM, Wupeng Ma wrote:
>> From: Ma Wupeng <mawupeng1@huawei.com>
>>
> 
> This patch set is all about follow_phys() cleanups, so "Cleanup for PAT"
> seems too generic.
> 
>> Patch #1 move follow_phys to memtype.c since only pat use this.
>> Patch #2 cleanup parameter in follow_phys.
>> Patch #3 drop the unnecessary WARN_ON_ONCE if follow_phys fails.
> 
> I'm more curious why follow_phys() ended up this way?
> 
> follow_phys() was introduced in commit 28b2ee20c7cba ("access_process_vm
> device memory infrastructure") in 2008 for getting a physical page address
> for a virtual address, and used in generic_access_phys(). And later it's
> used in x86 PAT code.
> 
> Commit 03668a4debf4f ("mm: use generic follow_pte() in follow_phys()") made
> follow_phys() more of a wrapper of follow_pte(), and commit 96667f8a4382d
> ("mm: Close race in generic_access_phys") replaced follow_phys() with
> follow_pte() in generic_access_phys(). And the end result is that
> follow_phys() is used in x86 PAT code only.

Thanks for the explanation. I have a better understanding of the history of
this function.

> 
> As follow_phys() in untrack_pfn() can be replaced with follow_pfn(), then

Yes, this can be replaced with follow_pfn().

> maybe we don't have to keep follow_phys(), and just use follow_pte() in
> track_pfn_copy()?

As follow_phys() will return unsigned long *prot which is need in track_pfn_copy(),
we need to do something with this.

Can we replace follow_pfn with follow_phys()?

Thanks!
Ma

> 
> Thanks!
>     Xin
> 
>>
>> Changelog since v3:
>> - rebase to latest linux
>> - fix compile warnings
>>
>> Changelog since v2:
>> - rebase to latest linux
>>
>> Changelog since v1:
>> - split patch #1 into two patches based on Boris's advise
>>
>> Ma Wupeng (3):
>>    x86/mm/pat: Move follow_phys to pat-related file
>>    x86/mm/pat: Cleanup unused parameter in follow_phys
>>    x86/mm/pat: Remove WARN_ON_ONCE if follow_phys fails
>>
>>   arch/x86/mm/pat/memtype.c | 32 ++++++++++++++++++++++++++------
>>   include/linux/mm.h        |  2 --
>>   mm/memory.c               | 28 ----------------------------
>>   3 files changed, 26 insertions(+), 36 deletions(-)
>>
> 
>
Xin Li (Intel) Feb. 20, 2024, 6:46 p.m. UTC | #3 
On 2/20/2024 1:06 AM, mawupeng wrote:

> On 2024/2/20 16:37, Xin Li wrote:
>> On 2/19/2024 7:48 PM, Wupeng Ma wrote:
>> follow_phys() was introduced in commit 28b2ee20c7cba ("access_process_vm
>> device memory infrastructure") in 2008 for getting a physical page address
>> for a virtual address, and used in generic_access_phys(). And later it's
>> used in x86 PAT code.
>>
>> Commit 03668a4debf4f ("mm: use generic follow_pte() in follow_phys()") made
>> follow_phys() more of a wrapper of follow_pte(), and commit 96667f8a4382d
>> ("mm: Close race in generic_access_phys") replaced follow_phys() with
>> follow_pte() in generic_access_phys(). And the end result is that
>> follow_phys() is used in x86 PAT code only.
> 
> Thanks for the explanation. I have a better understanding of the history of
> this function.
> 

"git blame" tells the story.

>>
>> As follow_phys() in untrack_pfn() can be replaced with follow_pfn(), then
> 
> Yes, this can be replaced with follow_pfn().
> 
>> maybe we don't have to keep follow_phys(), and just use follow_pte() in
>> track_pfn_copy()?
> 
> As follow_phys() will return unsigned long *prot which is need in track_pfn_copy(),
> we need to do something with this.

Commit 96667f8a4382d did that already.

> Can we replace follow_pfn with follow_phys()?

Sorry, I don't get your point.

Thanks!
     Xin