Re: [PATCH v2] mm, zone_device: replace {get, put}_zone_device_page() with a single reference

From: Jerome Glisse
Date: Tue May 02 2017 - 09:23:05 EST


On Tue, May 02, 2017 at 02:37:46PM +0300, Kirill A. Shutemov wrote:
> On Mon, May 01, 2017 at 09:55:48AM -0400, Jerome Glisse wrote:
> > On Mon, May 01, 2017 at 01:23:59PM +0300, Kirill A. Shutemov wrote:
> > > On Sun, Apr 30, 2017 at 07:14:24PM -0400, Jerome Glisse wrote:
> > > > On Sat, Apr 29, 2017 at 01:17:26PM +0300, Kirill A. Shutemov wrote:
> > > > > On Fri, Apr 28, 2017 at 03:33:07PM -0400, Jerome Glisse wrote:
> > > > > > On Fri, Apr 28, 2017 at 12:22:24PM -0700, Dan Williams wrote:
> > > > > > > Are you sure about needing to hook the 2 -> 1 transition? Could we
> > > > > > > change ZONE_DEVICE pages to not have an elevated reference count when
> > > > > > > they are created so you can keep the HMM references out of the mm hot
> > > > > > > path?
> > > > > >
> > > > > > 100% sure on that :) I need to callback into driver for 2->1 transition
> > > > > > no way around that. If we change ZONE_DEVICE to not have an elevated
> > > > > > reference count that you need to make a lot more change to mm so that
> > > > > > ZONE_DEVICE is never use as fallback for memory allocation. Also need
> > > > > > to make change to be sure that ZONE_DEVICE page never endup in one of
> > > > > > the path that try to put them back on lru. There is a lot of place that
> > > > > > would need to be updated and it would be highly intrusive and add a
> > > > > > lot of special cases to other hot code path.
> > > > >
> > > > > Could you explain more on where the requirement comes from or point me to
> > > > > where I can read about this.
> > > > >
> > > >
> > > > HMM ZONE_DEVICE pages are use like other pages (anonymous or file back page)
> > > > in _any_ vma. So i need to know when a page is freed ie either as result of
> > > > unmap, exit or migration or anything that would free the memory. For zone
> > > > device a page is free once its refcount reach 1 so i need to catch refcount
> > > > transition from 2->1
> > >
> > > What if we would rework zone device to have pages with refcount 0 at
> > > start?
> >
> > That is a _lot_ of work from top of my head because it would need changes
> > to a lot of places and likely more hot code path that simply adding some-
> > thing to put_page() note that i only need something in put_page() i do not
> > need anything in the get page path. Is adding a conditional branch for
> > HMM pages in put_page() that much of a problem ?
>
> Well, it gets inlined everywhere. Removing zone_device code from
> get_page() and put_page() saved non-trivial ~140k in vmlinux for
> allyesconfig.
>
> Re-introducing part this bloat would be unfortunate.
>
> > > > This is the only way i can inform the device that the page is now free. See
> > > >
> > > > https://cgit.freedesktop.org/~glisse/linux/commit/?h=hmm-v21&id=52da8fe1a088b87b5321319add79e43b8372ed7d
> > > >
> > > > There is _no_ way around that.
> > >
> > > I'm still not convinced that it's impossible.
> > >
> > > Could you describe lifecycle for pages in case of HMM?
> >
> > Process malloc something, end it over to some function in the program
> > that use the GPU that function call GPU API (OpenCL, CUDA, ...) that
> > trigger a migration to device memory.
> >
> > So in the kernel you get a migration like any existing migration,
> > original page is unmap, if refcount is all ok (no pin) then a device
> > page is allocated and thing are migrated to device memory.
> >
> > What happen after is unknown. Either userspace/kernel driver decide
> > to migrate back to system memory, either there is an munmap, either
> > there is a CPU page fault, ... So from that point on the device page
> > as the exact same life as a regular page.
> >
> > Above i describe the migrate case, but you can also have new memory
> > allocation that directly allocate device memory. For instance if the
> > GPU do a page fault on an address that isn't back by anything then
> > we can directly allocate a device page. No migration involve in that
> > case.
> >
> > HMM pages are like any other pages in most respect. Exception are:
> > - no GUP
>
> Hm. How do you exclude GUP? And why is it required?

Well it is not forbiden it just can not happen simply because as device
memory is not accessible by CPU then the corresponding CPU page table
entry is a special entry and thus GUP trigger a page fault that migrate
thing back to regular memory.

The why is simply because we need to always be able to migrate back to
regular memory as device memory is not accessible by the CPU. So we can
not allow anyone to pin it.

Cheers,
Jérôme