Re: [Qemu-devel] [PATCH] mmap-alloc: use same backend for all mappings

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Nov 30, 2015 at 02:46:31PM +0100, Greg Kurz wrote:
> On Mon, 30 Nov 2015 15:06:33 +0200
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
> > On Mon, Nov 30, 2015 at 11:51:57AM +0100, Greg Kurz wrote:
> > > Since commit 8561c9244ddf1122d "exec: allocate PROT_NONE pages on top of RAM",
> > > it is no longer possible to back guest RAM with hugepages on ppc64 hosts:
> > > 
> > > mmap(NULL, 285212672, PROT_NONE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x3fff57000000
> > > mmap(0x3fff57000000, 268435456, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, 19, 0) = -1 EBUSY (Device or resource busy)
> > > 
> > > This is due to a limitation on ppc64 that requires MAP_FIXED mappings to have
> > > the same page size as other mappings already present in the same "slice" of
> > > virtual address space (Cc'ing Ben for details).
> > 
> > I'd like some details please.
> > What do you mean when you say "same page size" and "slice"?
> > 
> 
> On ppc64, the address space is divided in 256MB-sized segments where all pages
> have the same size. This is a hw limitation IIUC. I don't know if it can be
> fixed and I'll let Ben comment on it.

But it's anonymous memory with PROT_NONE.  There should be no pages there:
just a chunk of virtual memory reserved.

> Hugepage support is implemented using an abstraction of segments called
> "slices". Here's a quote from the related commit changelog in the kernel
> tree:
> 
> commit d0f13e3c20b6fb73ccb467bdca97fa7cf5a574cd
> Author: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> Date:   Tue May 8 16:27:27 2007 +1000
> 
>     [POWERPC] Introduce address space "slices"
> 
> ...
> 
>     The main issues are:
>     
>      - To maintain/keep track of the page size per "segment" (as we can
>     only have one page size per segment on powerpc, which are 256MB
>     divisions of the address space).
>     
>      - To make sure special mappings stay within their allotted
>     "segments" (including MAP_FIXED crap)
>     
>      - To make sure everybody else doesn't mmap/brk/grow_stack into a
>     "segment" that is used for a special mapping
> ...
> 
> > > This is exactly what happens
> > > when calling mmap() above: first one uses native host page size (64k) and
> > > second one uses huge page size (16M).
> > > 
> > > To be sure we always have the same page size, let's use the same backend for
> > > both calls to mmap(): this is enough to fix the ppc64 issue.
> > > 
> > > This has no effect on RAM based mappings.
> > > 
> > > Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
> > > ---
> > > 
> > > This is a bug fix for 2.5
> > > 
> > >  util/mmap-alloc.c |    3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/util/mmap-alloc.c b/util/mmap-alloc.c
> > > index c37acbe58ede..0ff221dd94f4 100644
> > > --- a/util/mmap-alloc.c
> > > +++ b/util/mmap-alloc.c
> > > @@ -21,7 +21,8 @@ void *qemu_ram_mmap(int fd, size_t size, size_t align, bool shared)
> > >       * space, even if size is already aligned.
> > >       */
> > >      size_t total = size + align;
> > > -    void *ptr = mmap(0, total, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
> > > +    void *ptr = mmap(0, total, PROT_NONE,
> > > +                     (fd == -1 ? MAP_ANONYMOUS : 0) | MAP_PRIVATE, fd, 0);
> > >      size_t offset = QEMU_ALIGN_UP((uintptr_t)ptr, align) - (uintptr_t)ptr;
> > >      void *ptr1;
> > >  
> >