On Tue, Oct 08, 2013 at 09:34:30AM +0900, Minchan Kim wrote: > On Mon, Oct 07, 2013 at 05:18:40PM -0700, John Stultz wrote: > > On 10/07/2013 05:13 PM, Minchan Kim wrote: > > > Hello Peter, > > > > > > On Mon, Oct 07, 2013 at 04:59:40PM -0700, H. Peter Anvin wrote: > > >> On 10/07/2013 04:54 PM, John Stultz wrote: > > >>>> And wouldn't this apply to MADV_DONTNEED just as well? Perhaps what we > > >>>> should do is an enhanced madvise() call? > > >>> Well, I think MADV_DONTNEED doesn't *have* do to anything at all. Its > > >>> advisory after all. So it may immediately wipe out any data, but it may not. > > >>> > > >>> Those advisory semantics work fine w/ VRANGE_VOLATILE. However, > > >>> VRANGE_NONVOLATILE is not quite advisory, its telling the system that it > > >>> requires the memory at the specified range to not be volatile, and we > > >>> need to correctly inform userland how much was changed and if any of the > > >>> memory we did change to non-volatile was purged since being set volatile. > > >>> > > >>> In that way it is sort of different from madvise. Some sort of an > > >>> madvise2 could be done, but then the extra purge state argument would be > > >>> oddly defined for any other mode. > > >>> > > >>> Is your main concern here just wanting to have a zero-fill mode with > > >>> volatile ranges? Or do you really want to squeeze this in to the madvise > > >>> call interface? > > >> The point is that MADV_DONTNEED is very similar in that sense, > > >> especially if allowed to be lazy. It makes a lot of sense to permit > > >> both scrubbing modes orthogonally. > > >> > > >> The point you're making has to do with withdrawal of permission to flush > > >> on demand, which is a result of having the lazy mode (ongoing > > >> permission) and having to be able to withdraw such permission. > > > I'm sorry I could not understand what you wanted to say. > > > Could you elaborate a bit? > > My understanding of his point is that VRANGE_VOLATILE is like a lazy > > MADV_DONTNEED (with sigbus, rather then zero fill on fault), suggests > > that we should find a way to have VRANGE_VOLATILE be something like > > MADV_DONTNEED|MADV_LAZY|MADV_SIGBUS_FAULT, instead of adding a new > > syscall. This would provide more options, since one could instead just > > do MADV_DONTNEED|MADV_LAZY if they wanted zero-fill faults. > > Hmm, actually, I have thought VRANGE_SIGBUS option because Address/Thread > sanitizer people wanted it as you know and someone might want it, too. > > I agree it's orthogonal but not sure MADV_LAZY and MADV_SIGBUS_FAULT can be > used for other combination of advise except MADV_DONTNEED so it might > confuse userland without benefit. > > > > > And indeed, for the VRANGE_VOLATILE case, we could do something like > > that, but the unresolved problem I see is that that we still need to > > handle the VRANGE_NONVOLATILE case, and the madvise() interface doesn't > > seem to accomodate the needed semantics well. > > VRANGE_VOLATILE case could be a problem. In my mind, I had an idea to > return purged state when we call vrange(VRANGE_VOLATILE) because kernel > could purge them as soon as vrange(VRANGE_VOLATILE) called if memory is > really tight so userland can notice "purging" earlier and kernel can > discard them more efficiently. > And we should return the number of bytes marked but madvise returns error. -- Kind regards, Minchan Kim -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>