On Wed, Nov 28, 2012 at 08:18:01PM -0800, John Stultz wrote:Yes, depending on the pattern that MADV_VOLATILE and MADV_NOVOLATILE is applied, we can get a result where data is purged, but we aren't notified of it. Also, since madvise returns early if it encounters an error, in the case where you have checkerboard volatile regions (say every other page is volatile), which you mark non-volatile with one large MADV_NOVOLATILE call, the first volatile vma will be marked non-volatile, but since it returns purged, the madvise loop will stop and the following volatile regions will be left volatile.On 11/21/2012 04:36 PM, John Stultz wrote:Hmm, my patch doesn't allow to merge volatile with another one by2) Being able to use this with tmpfs files. I'm currently tryingHey Minchan,
to better understand the rmap code, looking to see if there's a
way to have try_to_unmap_file() work similarly to
try_to_unmap_anon(), to allow allow users to madvise() on mmapped
tmpfs files. This would provide a very similar interface as to
what I've been proposing with fadvise/fallocate, but just using
process virtual addresses instead of (fd, offset) pairs. The
benefit with (fd,offset) pairs for Android is that its easier to
manage shared volatile ranges between two processes that are
sharing data via an mmapped tmpfs file (although this actual use
case may be fairly rare). I believe we should still be able to
rework the ashmem internals to use madvise (which would provide
legacy support for existing android apps), so then its just a
question of if we could then eventually convince Android apps to
use the madvise interface directly, rather then the ashmem unpin
I've been playing around with your patch trying to better
understand your approach and to extend it to support tmpfs files. In
doing so I've found a few bugs, and have some rough fixes I wanted
to share. There's still a few edge cases I need to deal with (the
vma-purged flag isn't being properly handled through vma merge/split
operations), but its starting to come along.
inserting VM_VOLATILE into VM_SPECIAL so I guess merge isn't problem.
In case of split, __split_vma copy old vma to new vma like this
*new = *vma;
So the problem shouldn't happen, I guess.
Did you see the real problem about that?
Anyway, take a look at the tree here and let me know what you think.