Messages in this thread | | | Date | Tue, 25 Feb 2014 10:45:51 -0800 | Subject | Re: [PATCH v2] mm: per-thread vma caching | From | Linus Torvalds <> |
| |
On Tue, Feb 25, 2014 at 10:37 AM, Linus Torvalds <torvalds@linux-foundation.org> wrote: > > - clear all the cache entries (of the new 'struct task_struct'! - so > not in dup_mmap, but make sure it's zeroed when allocating!)( > > - set vmcache_seqnum to 0 in dup_mmap (since any sequence number is > fine when it got invalidated, and 0 is best for "avoid overflow").
Btw, as far as I can tell, that also makes the per-thread vmacache automatically do the right thing for the non-MMU case, so that you could just remove the difference between CONFIG_MMU and NOMMU.
Basically, dup_mmap() should no longer have anything to do with the vmacache, since it is now per-thread, not per-mm.
So :
- allocating a new "struct mm_struct" should clear the vmacache_seqnum for that new mm, to try to minimize unnecessary future overflow.
- thread allocation should just zero the cache entries, and set "tsk->vmacache_seqnum = mm->vmacache_seqnum" (after dup_mm()) to avoid future unnecessary flushes.
and as far as I can tell, the logic would be exactly the same on NOMMU (the dup_mm just doesn't happen, since all forks are basically sharing mm).
And maybe you'd want to make VMACACHE_SIZE be 1 on NOMMU (and make sure to change the "& 3" to "& (VMACACHE_SIZE-1)". Just to keep the size down on small systems that really don't need it.
Linus
| |