Re: [PATCH] xen: allocate page for shared info page from low memory

[Boris Ostrovsky]

On 06/14/2017 01:11 PM, Juergen Gross wrote:
> On 14/06/17 18:58, Boris Ostrovsky wrote:
> > On 06/12/2017 07:53 AM, Juergen Gross wrote:
> > > In a HVM guest the kernel allocates the page for mapping the shared
> > > info structure via extend_brk() today. This will lead to a drop of
> > > performance as the underlying EPT entry will have to be split up into
> > > 4kB entries as the single shared info page is located in hypervisor
> > > memory.
> > >
> > > The issue has been detected by using the libmicro munmap test:
> > > unmapping 8kB of memory was faster by nearly a factor of two when no
> > > pv interfaces were active in the HVM guest.
> > >
> > > So instead of taking a page from memory which might be mapped via
> > > large EPT entries use a page which is already mapped via a 4kB EPT
> > > entry: we can take a page from the first 1MB of memory as the video
> > > memory at 640kB disallows using larger EPT entries.
> > >
> > > Signed-off-by: Juergen Gross <jgross@xxxxxxxx>
> > > ---
> > >   arch/x86/xen/enlighten_hvm.c | 31 ++++++++++++++++++++++++-------
> > >   arch/x86/xen/enlighten_pv.c  |  2 --
> > >   2 files changed, 24 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/arch/x86/xen/enlighten_hvm.c b/arch/x86/xen/enlighten_hvm.c
> > > index a6d014f47e52..c19477b6e43a 100644
> > > --- a/arch/x86/xen/enlighten_hvm.c
> > > +++ b/arch/x86/xen/enlighten_hvm.c
> > > @@ -1,5 +1,6 @@
> > >   #include <linux/cpu.h>
> > >   #include <linux/kexec.h>
> > > +#include <linux/memblock.h>
> > >   
> > >   #include <xen/features.h>
> > >   #include <xen/events.h>
> > > @@ -10,9 +11,11 @@
> > >   #include <asm/reboot.h>
> > >   #include <asm/setup.h>
> > >   #include <asm/hypervisor.h>
> > > +#include <asm/e820/api.h>
> > >   
> > >   #include <asm/xen/cpuid.h>
> > >   #include <asm/xen/hypervisor.h>
> > > +#include <asm/xen/page.h>
> > >   
> > >   #include "xen-ops.h"
> > >   #include "mmu.h"
> > > @@ -22,20 +25,34 @@ void __ref xen_hvm_init_shared_info(void)
> > >   {
> > >   	int cpu;
> > >   	struct xen_add_to_physmap xatp;
> > > -	static struct shared_info *shared_info_page;
> > > +	u64 pa;
> > > +
> > > +	if (HYPERVISOR_shared_info == &xen_dummy_shared_info) {
> > > +		/*
> > > +		 * Search for a free page starting at 4kB physical address.
> > > +		 * Low memory is preferred to avoid an EPT large page split up
> > > +		 * by the mapping.
> > > +		 * Starting below X86_RESERVE_LOW (usually 64kB) is fine as
> > > +		 * the BIOS used for HVM guests is well behaved and won't
> > > +		 * clobber memory other than the first 4kB.
> > > +		 */
> > > +		for (pa = PAGE_SIZE;
> > > +		     !e820__mapped_all(pa, pa + PAGE_SIZE, E820_TYPE_RAM) ||
> > > +		     memblock_is_reserved(pa);
> > > +		     pa += PAGE_SIZE)
> > > +			;
> >
> > Is it possible to never find a page here?
>
> Only if there is no memory available at all. :-)
>
> TBH: I expect this to _always_ succeed at the first loop iteration.

This patch seems to break (64-bit only) guests on dumpdata here. No 
problems on other machines.

So far all I know is that we did get the first page (0x1000) but not 
much more. I will poke at this more on Monday.

-boris