kmalloc() question - migration/remap of kmalloc()'ed pages

kmalloc() question - migration/remap of kmalloc()'ed pages

[Mounesh Badiger]

[-- Attachment #1: Type: text/plain, Size: 440 bytes --]

Hi,

We have a device driver that allocates physical pages using kmalloc() and
these pages are mapped to the userspace process via mmap.

Does the kmalloc()'ed memory mapped to userspace can migrate or get
remapped due to any of kernel mm operations?

we are looking to keep VA->PA mapping intact from a userspace process point
of view.

-- 
Mounesh Badiger
Support for education
Visit*: www.vidyaposhak.org <http://www.vidyaposhak.org/> *

[-- Attachment #2: Type: text/html, Size: 967 bytes --]

Re: kmalloc() question - migration/remap of kmalloc()'ed pages

[David Hildenbrand]

On 25.04.24 11:21, Mounesh Badiger wrote:
> Hi,

Hi,

> 
> We have a device driver that allocates physical pages using kmalloc() 
> and these pages are mapped to the userspace process via mmap.

Which interface are you using for inserting these pages into the page 
tables?

remap_pfn_range? vm_insert_page() ? ...

Using kmalloc sound weird ... why not simply alloc_pages()?

> 
> Does the kmalloc()'ed memory mapped to userspace can migrate or get 
> remapped due to any of kernel mm operations?

No.

> 
> we are looking to keep VA->PA mapping intact from a 
> userspace process point of view.

That will work.

-- 
Cheers,

David / dhildenb

Re: kmalloc() question - migration/remap of kmalloc()'ed pages

[Mounesh Badiger]

[-- Attachment #1: Type: text/plain, Size: 1026 bytes --]

Thanks for the confirmation.
I am using remap_pfn_range(). I can use alloc_pages() instead of kmalloc().

On Thu, Apr 25, 2024 at 3:06 PM David Hildenbrand <david@redhat.com> wrote:

> On 25.04.24 11:21, Mounesh Badiger wrote:
> > Hi,
>
> Hi,
>
> >
> > We have a device driver that allocates physical pages using kmalloc()
> > and these pages are mapped to the userspace process via mmap.
>
> Which interface are you using for inserting these pages into the page
> tables?
>
> remap_pfn_range? vm_insert_page() ? ...
>
> Using kmalloc sound weird ... why not simply alloc_pages()?
>
> >
> > Does the kmalloc()'ed memory mapped to userspace can migrate or get
> > remapped due to any of kernel mm operations?
>
> No.
>
> >
> > we are looking to keep VA->PA mapping intact from a
> > userspace process point of view.
>
> That will work.
>
> --
> Cheers,
>
> David / dhildenb
>
>

-- 
Mounesh Badiger
Support for education
Visit*: www.vidyaposhak.org <http://www.vidyaposhak.org/> *

[-- Attachment #2: Type: text/html, Size: 1822 bytes --]

Re: kmalloc() question - migration/remap of kmalloc()'ed pages

[David Hildenbrand]

On 25.04.24 11:48, Mounesh Badiger wrote:
> Thanks for the confirmation.
> I am using remap_pfn_range(). I can use alloc_pages() instead of kmalloc().
> 

Okay, for remap_pfn_range() it doesn't matter, because we'll never touch 
the "struct page" metadata (pages are considered not refcounted).

vm_insert_page() would likely have refused the kmalloc'ed pages because 
validate_page_before_insert() would refuse mapping any slab folios.

Should be working for you as expected.

-- 
Cheers,

David / dhildenb