[PATCH v11 00/14] HMM anomymous memory migration to device memory

From: JÃrÃme Glisse
Date: Wed Oct 21 2015 - 16:14:34 EST


Minor fixes since last post, apply on top of 4.3rc6. Tree with
the patchset:

git://people.freedesktop.org/~glisse/linux hmm-v11 branch

This patchset implement anonymous memory migration for HMM.
See HMM patchset for full description of what is HMM and why
doing HMM :

https://lkml.org/lkml/2015/10/21/739

Seamless migration from system memory to device memory ie on CPU
access we migrate memory back to system memory so CPU can access
it again.

Design is simple, a new special swap type is added and CPU pte are
set to this special swap type for migrated memory. On CPU page fault
HMM use its mirror page table to find proper page into device memory
and migrate it back to system memory.

Migration to device memory involves several steps :
- First CPU page table is updated to special pte and current
pte is save to temporary array.
- We check that all pte are for normal/real pages.
- We check that no one holds an extra reference on the page.
- At this point we know we are the only one know about that
memory and we can safely copy it to device memory.
- Once everything is copied and fine on device side we free
the system ram pages.

Migration from device memory back to system memory is simpler:
- We get exclusive access for each pte we want to migrate back
(special swap pte value).
- We allocate system memory (memcg and anon_vma handled here).
- We copy back device memory content into system memory and
update device page table to point to system memory.
- We update CPU page table to point to new system memory.

Cheers,
JÃrÃme

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/