Re: [Qemu-devel] [PATCH 0/5] PCI hotplug fixes/cleanup

[Michael S. Tsirkin]

On Wed, Apr 04, 2012 at 11:51:00PM -0600, Alex Williamson wrote:
> We've been batting this one back and forth.  This series includes
> several of the cleanups and specification clarifications from my
> series awhile back.  Patch 5 is my proposed alternative to
> Michael's PCI hotplug race fix.  Since that version I added slot
> present tracking so we can be a little more strategic about which
> slots we ask the guest to check.  The approach for that path is
> described in the commit log.

So I would have organized it differently,
with a bugfix first and the rework on top
(e.g. make life easier if we want to cherry-pick this change
to the stable branch) and there's a bit of
discussion about documentation but I'm prepared to apply
this as is too if you prefer.

For the series:
Acked-by: Michael S. Tsirkin <address@hidden>

>  I tested hotplug on both Linux an
> Windows guests (XP + 2k8), intentionally trying to do back to back
> device_add and device_del to get a race, but couldn't (I did
> however get a glibc double free that seems unrelated to this
> series).

Reproducible?


> Long term I'd like to deprecate the up/down PCI hotplug interface
> and move to a new model.  I think perhaps we should define 3 new
> registers:
> 
> 1) Device present in slot bitmap (foundation already in 5/5 here)
> 2) Virtual slot power state bitmap
> 3) Requested slot date bitmap
> 
> With these we should be able to do proper _STA, _PS0, and _PS3.
> We'd maintain the eject register for _EJ0, but deprecate up/down.
> 
> For a device_add, the 'present' & 'request' bitmaps gets set for
> the slot by qemu, gpe.sts set and SCI interrupt sent.  The OSPM
> can simply compare requested state vs power state and send the
> appropriate notify.  Same general idea for removal.  _STA can be
> formed by looking at the 'present' and 'power' bitmaps, and
> _PS0/3 perform writes to the power register to indicate the virtual
> state.  I think this provides a better register ownership model
> (qemu owns present/request, OSPM owns power) and also allows us
> to differentiate between a guest initiated remove and a platform
> (vm manager) initiated remove so we can avoid changing the vm
> state in the former case.  I'll try to send a patch for this out
> soon, but let me know if there are any concerns or suggestions for
> a redesign.  Thanks,
> 
> Alex
> 
> ---
> 
> Alex Williamson (5):
>       acpi_piix4: Fix PCI hotplug race
>       acpi_piix4: Remove PCI_RMV_BASE write code
>       acpi_piix4: Use pci_get/set_byte
>       acpi_piix4: Only allow writes to PCI hotplug eject register
>       acpi_piix4: Disallow write to up/down PCI hotplug registers
> 
> 
>  docs/specs/acpi_pci_hotplug.txt |    8 +-
>  hw/acpi_piix4.c                 |  178 
> ++++++++++++++++++++-------------------
>  2 files changed, 95 insertions(+), 91 deletions(-)