Messages in this thread | | | Date | Wed, 9 Oct 2013 12:39:32 +0200 | From | Peter Zijlstra <> | Subject | Re: [RFC] perf: mmap2 not covering VM_CLONE regions |
| |
On Wed, Oct 09, 2013 at 11:59:06AM +0200, Ingo Molnar wrote: > PeterZ didn't like exposing the physical RAM address, right?
More than not like; its useless and broken, see below.
> Peter: could we expose the page frame number instead? (Maybe mix it with a > random seed through the hash-mixer and expose that.) User-space will thus > be able to discover whether two pages are shared or not, but not extract > other information.
No, all physical stuff is useless, a page can get moved about through swap/ksm/thp/compaction/etc.
> The global MM ID thing looked rather ugly as well.
I'm still having the argument with Stephane -- although it fell off-list -- on why this wouldn't work.
AFAICT the mm_id should work for CLONE_VM, as CLONE_VM inherits the entire address space so all should be identical when mm_id match.
> That way we could drop the inode info as well? That feels a bit hacky too.
No, the inode stuff is required for shared pages, its the only way to track them; and barring a filename (which we've so far used for this) the dev:ino:gen thing is the only way to uniquely identify files -- it even works where filenames don't, eg. filesystem namespaces and weird things like shm that don't have stable filenames.
| |