Re: [Qemu-devel] [RFC] virtio-mem: paravirtualized memory

[Stefan Hajnoczi]

On Wed, Jun 21, 2017 at 02:32:48PM +0200, David Hildenbrand wrote:
> On 21.06.2017 13:08, Stefan Hajnoczi wrote:
> > On Mon, Jun 19, 2017 at 12:26:52PM +0200, David Hildenbrand wrote:
> >> On 19.06.2017 12:08, Stefan Hajnoczi wrote:
> >>> On Fri, Jun 16, 2017 at 04:20:02PM +0200, David Hildenbrand wrote:
> >>>> Important restrictions of this concept:
> >>>> - Guests without a virtio-mem guest driver can't see that memory.
> >>>> - We will always require some boot memory that cannot get unplugged.
> >>>>   Also, virtio-mem memory (as all other hotplugged memory) cannot become
> >>>>   DMA memory under Linux. So the boot memory also defines the amount of
> >>>>   DMA memory.
> >>>
> >>> I didn't know that hotplug memory cannot become DMA memory.
> >>>
> >>> Ouch.  Zero-copy disk I/O with O_DIRECT and network I/O with virtio-net
> >>> won't be possible.
> >>>
> >>> When running an application that uses O_DIRECT file I/O this probably
> >>> means we now have 2 copies of pages in memory: 1. in the application and
> >>> 2. in the kernel page cache.
> >>>
> >>> So this increases pressure on the page cache and reduces performance :(.
> >>>
> >>> Stefan
> >>>
> >>
> >> arch/x86/mm/init_64.c:
> >>
> >> /*
> >>  * Memory is added always to NORMAL zone. This means you will never get
> >>  * additional DMA/DMA32 memory.
> >>  */
> >> int arch_add_memory(int nid, u64 start, u64 size, bool for_device)
> >> {
> >>
> >> The is for sure something to work on in the future. Until then, base
> >> memory of 3.X GB should be sufficient, right?
> > 
> > I'm not sure that helps because applications typically don't control
> > where their buffers are located?
> 
> Okay, let me try to explain what is going on here (no expert, please
> someone correct me if I am wrong).
> 
> There is a difference between DMA and DMA memory in Linux. DMA memory is
> simply memory with special addresses. DMA is the general technique of a
> device directly copying data to ram, bypassing the CPU.
> 
> ZONE_DMA contains all* memory < 16MB
> ZONE_DMA32 contains all* memory < 4G
> * meaning available on boot via a820 map, not hotplugged.
> 
> So memory from these zones can be used by devices that can only deal
> with 24bit/32bit addresses.
> 
> Hotplugged memory is never added to the ZONE_DMA/DMA32, but to
> ZONE_NORMAL. That means, kmalloc(.., GFP_DMA will) not be able to use
> hotplugged memory. Say you have 1GB of main storage and hotplug 1G (on
> address 1G). This memory will not be available in the ZONE_DMA, although
> below 4g.
> 
> Memory in ZONE_NORMAL is used for ordinary kmalloc(), so all these
> memory can be used to do DMA, but you are not guaranteed to get 32bit
> capable addresses. I pretty much assume that virtio-net can deal with
> 64bit addresses.
> 
> 
> My understanding of O_DIRECT:
> 
> The user space buffers (O_DIRECT) is directly used to do DMA. This will
> work just fine as long as the device can deal with 64bit addresses. I
> guess this is the case for virtio-net, otherwise there would be the
> exact same problem already without virtio-mem.
> 
> Summary:
> 
> virtio-mem memory can be used for DMA, it will simply not be added to
> ZONE_DMA/DMA32 and therefore won't be available for kmalloc(...,
> GFP_DMA). This should work just fine with O_DIRECT as before.
> 
> If necessary, we could try to add memory to the ZONE_DMA later on,
> however for now I would rate this a minor problem. By simply using 3.X
> GB of base memory, basically all memory that could go to ZONE_DMA/DMA32
> already is in these zones without virtio-mem.

Nice, thanks for clearing this up!

Stefan

signature.asc
Description: PGP signature