On 19.05.23 12:10, Marek Marczykowski-Górecki wrote:
> On Fri, May 19, 2023 at 06:04:05AM +0200, Christoph Hellwig wrote:
>> On Thu, May 18, 2023 at 08:18:39PM +0200, Marek Marczykowski-Górecki wrote:
>>> On Thu, May 18, 2023 at 03:42:51PM +0200, Christoph Hellwig wrote:
>>>> Remove the dangerous late initialization of xen-swiotlb in
>>>> pci_xen_swiotlb_init_late and instead just always initialize
>>>> xen-swiotlb in the boot code if CONFIG_XEN_PCIDEV_FRONTEND is enabled.
>>>>
>>>> Signed-off-by: Christoph Hellwig <hch@lst.de>
>>>
>>> Doesn't it mean all the PV guests will basically waste 64MB of RAM
>>> by default each if they don't really have PCI devices?
>>
>> If CONFIG_XEN_PCIDEV_FRONTEND is enabled, and the kernel's isn't booted
>> with swiotlb=noforce, yes.
> 
> That's "a bit" unfortunate, since that might be significant part of the
> VM memory, or if you have a lot of VMs, a significant part of the host
> memory - it quickly adds up.
> While I would say PCI passthrough is not very common for PV guests, can
> the decision about xen-swiotlb be delayed until you can enumerate
> xenstore to check if there are any PCI devices connected (and not
> allocate xen-swiotlb by default if there are none)? This would
> still not cover the hotplug case (in which case, you'd need to force it
> with a cmdline), but at least you wouldn't loose much memory just
> because one of your VMs may use PCI passthrough (so, you have it enabled
> in your kernel).
> Please remember that guest kernel is not always under full control of
> the host admin, so making guests loose 64MB of RAM always, in default
> setup isn't good for customers of such VMs...
> 

In normal cases PCI passthrough in PV guests requires to start the guest
with e820_host=1. So it should be rather easy to limit allocating the
64MB in PV guests to the cases where the memory map has non-RAM regions
especially in the first 1MB of the memory.

This will cover even hotplug cases. The only case not covered would be a
guest started with e820_host=1 even if no PCI passthrough was planned.
But this should be rather rare (at least I hope so).


Juergen