Hi,
I'm trying to remove the current page handling in binder and switch to
using ->fault() and other mm/ infrastructure. See, we currently allocate
and insert pages manually into vmas, so dropping this boilerplate code
seems like a great idea to me. Before starting on this though, I could
use some pointers on some of the implementation details.
In binder the messages are not buffered. What happens is that a server
task mmaps a region and then allows clients to write their user data
directly into this _remote_ vma. This is currently achieved by caching
pointers to every page allocated in the server's region.
Instead of caching page pointers, I believe binder could make use of
get_user_pages_remote() and trigger a page fault as needed. Basically,
implement something similar to access_remote_vm() but that instead does
a copy_from_user(). However, I don't see many in-tree users of these
routines, so I wonder if I'm overlooking something in this approach?
One more reason for the page caching in binder is being able to free
pages from the shrinker's callback. Unused pages are added to an LRU
list and we manually zap/free them when they are reclaimed. It is not
evident to me how can binder mark a range within the vma region as
reclaimable?
Any pointers are appreciated!
--
Carlos Llamas
On Thu, Apr 20, 2023 at 1:34 PM Carlos Llamas <[email protected]> wrote:
>
> Hi,
>
> I'm trying to remove the current page handling in binder and switch to
> using ->fault() and other mm/ infrastructure. See, we currently allocate
> and insert pages manually into vmas, so dropping this boilerplate code
> seems like a great idea to me. Before starting on this though, I could
> use some pointers on some of the implementation details.
>
> In binder the messages are not buffered. What happens is that a server
> task mmaps a region and then allows clients to write their user data
> directly into this _remote_ vma. This is currently achieved by caching
> pointers to every page allocated in the server's region.
>
> Instead of caching page pointers, I believe binder could make use of
> get_user_pages_remote() and trigger a page fault as needed. Basically,
> implement something similar to access_remote_vm() but that instead does
> a copy_from_user(). However, I don't see many in-tree users of these
> routines, so I wonder if I'm overlooking something in this approach?
Sounds doable but there might be GUP details I'm missing...
>
> One more reason for the page caching in binder is being able to free
> pages from the shrinker's callback. Unused pages are added to an LRU
> list and we manually zap/free them when they are reclaimed. It is not
> evident to me how can binder mark a range within the vma region as
> reclaimable?
I think you would have to release the individual pages using
put_page(). Also note that get_user_pages_remote() is being
deprecated, see:
https://elixir.bootlin.com/linux/latest/source/mm/gup.c#L2171
>
> Any pointers are appreciated!
>
> --
> Carlos Llamas
On Thu, Apr 20, 2023 at 02:35:00PM -0700, Suren Baghdasaryan wrote:
> On Thu, Apr 20, 2023 at 1:34 PM Carlos Llamas <[email protected]> wrote:
> >
> > Instead of caching page pointers, I believe binder could make use of
> > get_user_pages_remote() and trigger a page fault as needed. Basically,
> > implement something similar to access_remote_vm() but that instead does
> > a copy_from_user(). However, I don't see many in-tree users of these
> > routines, so I wonder if I'm overlooking something in this approach?
>
> Sounds doable but there might be GUP details I'm missing...
>
> >
> > One more reason for the page caching in binder is being able to free
> > pages from the shrinker's callback. Unused pages are added to an LRU
> > list and we manually zap/free them when they are reclaimed. It is not
> > evident to me how can binder mark a range within the vma region as
> > reclaimable?
>
> I think you would have to release the individual pages using
> put_page(). Also note that get_user_pages_remote() is being
> deprecated, see:
> https://elixir.bootlin.com/linux/latest/source/mm/gup.c#L2171
Yeah, I remember reading that comment but it doesn't make sense since
the suggested alternatives do not support operations on _remote_ mm.
I actually tracked this down and it seems the original comment was made
for get_user_pages() instead, in commit f0818f472d8d ("mm: gup: add
get_user_pages_locked and get_user_pages_unlocked"). The whole comment
block was then carried over without update into get_user_pages_remote()
in commit 1e9877902dc7 ("mm/gup: Introduce get_user_pages_remote()").
The comment block was finally switched to use the "_remote" portion in
commit adc8cb406e52 ("mm/gup.c: update the documentation"). In reality,
this comment should be relocated to the get_user_pages() section.
>
> >
> > Any pointers are appreciated!
> >
> > --
> > Carlos Llamas