Re: [Qemu-devel] [RFC PATCH 0/8] MemoryRegion and FlatView refcounting, replace hostmem with memory_region_find

[Paolo Bonzini <pbonzini@redhat.com>]

Il 09/05/2013 02:53, liu ping fan ha scritto:
> On Wed, May 8, 2013 at 11:44 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:
>> Il 08/05/2013 08:20, liu ping fan ha scritto:
>>> On Mon, May 6, 2013 at 10:25 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:
>>>> Hi,
>>>>
>>>> this is an alternative approach to refactoring of dataplane's HostMem
>>>> code.  Here, I take Ping Fan's idea of RCU-style updating of the
>>>> region list and apply it to the AddressSpace's FlatView.  With this
>>>
>>> In fact, I am worrying about the priority of MemoryListener, if it is
>>> true, then we should drop RCU-style idea.
>>
>> You mean in hostmem, or in general as in this patch?  Note that this
>> patch releases the old FlatView at the end of all MemoryListener operations.
>>
> Both in hostmem and this patch, they all broke the original design of
> the  MemoryListener, see notes for priority in code.

I think both hostmem and this patch are fine.  The hypervisor is never
involved, all accesses go through the "old" FlatView and regions cannot
disappear thanks to ref/unref.

In fact, we need _more_ RCU-style updates, not less.  For BQL-less
dispatch, address space mapping/translation can race against the
MemoryListeners in exec.c.  To fix this, phys_sections and
AddressSpaceDispatch need to be reference counted and RCU-ified as well.

Paolo

> I have set out 2 patches to highlight this issue, and have CC you and Stefanha.
> 
> Regards,
> Pingfan
> 
>> Paolo
>>
>>> Also if it is true, there is
>>> already a bug with hostmem listener. It should use region_del, not
>>> region_nop to reconstruct the local view. But just let me have a deep
>>> thinking.
>>>
>>> Regards,
>>> Pingfan
>>>> change, dataplane can simply use memory_region_find instead of
>>>> hostmem.
>>>>
>>>> This is a somewhat larger change, but I prefer it for two reasons.
>>>>
>>>> 1) it splits the task of adding BQL-less memory dispatch in two parts,
>>>>    tacking memory_region_find first (which is simpler because locking
>>>>    is left to the caller).
>>>>
>>>> 2) HostMem duplicates a lot of the FlatView logic, and adding the
>>>>    RCU-style update in FlatView benefits everyone.
>>>>
>>>> The missing ingredients here are:
>>>>
>>>> 1) remember and unreference the MemoryRegions that are used in
>>>>    a vring entry.  In order to implement this, it is probably simpler
>>>>    to change vring.c to use virtio.c's VirtQueueElement data structure.
>>>>    We want something like that anyway in order to support migration.
>>>>
>>>> 2) add an owner field to MemoryRegion, and set it for all MemoryRegions
>>>>    for hot-unpluggable devices.  In this series, ref/unref are stubs.
>>>>
>>>> For simplicity I based the patches on my IOMMU rebase.  I placed the
>>>> tree at git://github.com/bonzini/qemu.git, branch iommu.
>>>>
>>>> Paolo
>>>>
>>>> Paolo Bonzini (8):
>>>>   memory: add ref/unref calls
>>>>   exec: check MRU in qemu_ram_addr_from_host
>>>>   memory: return MemoryRegion from qemu_ram_addr_from_host
>>>>   memory: ref/unref memory across address_space_map/unmap
>>>>   memory: access FlatView from a local variable
>>>>   memory: use a new FlatView pointer on every topology update
>>>>   memory: add reference counting to FlatView
>>>>   dataplane: replace hostmem with memory_region_find
>>>>
>>>>  exec.c                                |   63 +++++++++---
>>>>  hw/core/loader.c                      |    1 +
>>>>  hw/display/exynos4210_fimd.c          |    6 +
>>>>  hw/display/framebuffer.c              |   10 +-
>>>>  hw/i386/kvm/ioapic.c                  |    2 +
>>>>  hw/i386/kvmvapic.c                    |    1 +
>>>>  hw/misc/vfio.c                        |    2 +
>>>>  hw/virtio/dataplane/Makefile.objs     |    2 +-
>>>>  hw/virtio/dataplane/hostmem.c         |  176 ---------------------------------
>>>>  hw/virtio/dataplane/vring.c           |   56 +++++++++--
>>>>  hw/virtio/vhost.c                     |    2 +
>>>>  hw/virtio/virtio-balloon.c            |    1 +
>>>>  hw/xen/xen_pt.c                       |    4 +
>>>>  include/exec/cpu-common.h             |    2 +-
>>>>  include/exec/memory.h                 |    9 ++
>>>>  include/hw/virtio/dataplane/hostmem.h |   57 -----------
>>>>  include/hw/virtio/dataplane/vring.h   |    3 +-
>>>>  kvm-all.c                             |    2 +
>>>>  memory.c                              |  142 +++++++++++++++++++++-----
>>>>  target-arm/kvm.c                      |    2 +
>>>>  target-i386/kvm.c                     |    4 +-
>>>>  target-sparc/mmu_helper.c             |    1 +
>>>>  xen-all.c                             |    2 +
>>>>  23 files changed, 253 insertions(+), 297 deletions(-)
>>>>  delete mode 100644 hw/virtio/dataplane/hostmem.c
>>>>  delete mode 100644 include/hw/virtio/dataplane/hostmem.h
>>>>
>>