Re: [Qemu-devel] [PATCH] add MADV_DONTFORK to guest physical memory

[Michael Roth]

On 01/06/2011 11:49 AM, Andrea Arcangeli wrote:
> On Wed, Jan 05, 2011 at 03:27:37PM -0600, Michael Roth wrote:
> > On 01/05/2011 02:35 PM, Andrea Arcangeli wrote:
> > > On Wed, Jan 05, 2011 at 02:26:19PM -0600, Michael Roth wrote:
> > > > Yah you're right, but I've seen several discussions about using mempath
> > > > for tmpfs/ram-backed files for things like numa/zram/etc so tend to
> > > > think of it as something potentially more than just a hook for
> > > > hugetlbfs, which is becoming less and less useful. But the MADV_DONTFORK
> > > > stuff should still be immediately applicable.
> > >
> > > Yes, MADV_DONTFORK should be set all on all guest physical memory
> > > without options so I hope the new patch I just posted is fine to stop
> > > the spurious -ENOMEM failures in fork.
> >
> > The patch in this thread? A couple paths still aren't covered when using
> > -mem-path. Something like this should get them all:
>
> Well the reason of MADV_DONTFORK is to avoid accounting issues with
> anonymous memory, mem-path don't have that issue as hugetlbfs skips
> the accounting (it has to because hugetlbfs are not always taken from
> the regular page allocator). It could be however considered a minor
> performance optimization.

That's one case, but there's also a wonky fallback in that path that 
defaults to the normal qemu_vmalloc():

if (mem_path) {
#if defined (__linux__) && !defined(TARGET_S390X)
            new_block->host = file_ram_alloc(new_block, size, mem_path);
            if (!new_block->host) {
                new_block->host = qemu_vmalloc(size);
                qemu_madvise(new_block->host, size, QEMU_MADV_MERGEABLE);
            }

May make sense to only add coverage for that specific case though. If 
file_ram_alloc() is generalized we could deal with it then.

> Now you mention that you want to use -mem-path for other things too,
> so maybe that's why you need it there too. BTW, if you ever need it
> for more than hugetlbfs, I'm afraid this MAP_PRIVATE I see when
> mem_prealloc isn't set, is going to screw any other potential useful
> usage besides hugetlbfs, not exactly sure why it makes any sense to
> use MAP_PRIVATE there and not only MAP_SHARED.

Not sure either...but if it's another hugetlbfs-ism it shouldn't matter 
since using mem-path for something other than hugetlbfs would ideally be 
configurable with a -mem path=/dev/shm/vm0.mem,hugetlbfs=off or 
something along that line, which wouldn't necessarily set MAP_PRIVATE.