Messages in this thread | | | From | Mel Gorman <> | Subject | [PATCH 0/2] Fix migration races in rmap_walk() | Date | Mon, 26 Apr 2010 23:37:56 +0100 |
| |
After digging around a lot, I believe the following two patches are the best way to close the race that allows a migration PTE to be left behind triggering a BUG check in migration_entry_to_page().
Patch one alters has fork() wait for migration to complete. Patch two has vma_adjust() acquire the anon_vma lock it is aware of and makes rmap_walk() aware that different VMAs can be encountered during the walk.
I dropped the use of the seq counter because there were still races in place. For example, while the seq counter would catch when vma_adjust() and rmap_walk() were looking at the same VMA, there was still insufficient protection on the VMA list being modified.
The reproduction case was as follows;
1. Run kernel compilation in a loop 2. Start two processes that repeatedly fork()ed and manipulated mappings 3. Constantly compact memory using /proc/sys/vm/compact_memory 4. Optionally add/remove swap
With these two patches applied, I was unable to trigger the bug check in migration_entry_to_page() but it would be really helpful if Rik could comment on the anon_vma locking requirements and whether patch 2 is 100% safe or not. The tests have only been running 8 hours but I'm posting now anyway and will see how it survives running for a few days.
The other issues raised about expand_downwards will need to be re-examined to see if they still exist and transparent hugepage support will need further thinking to see if split_huge_page() can deal with these situations.
mm/ksm.c | 13 +++++++++++++ mm/memory.c | 25 ++++++++++++++++--------- mm/mmap.c | 6 ++++++ mm/rmap.c | 23 ++++++++++++++++++++--- 4 files changed, 55 insertions(+), 12 deletions(-)
| |