Re: [PATCH 2/2] xen/swiotlb: don't destroy contiguous region in all cases

[Jan Beulich <jbeulich@suse.com>]

On 12.02.2025 12:11, Jürgen Groß wrote:
> On 12.02.25 08:38, Jan Beulich wrote:
>> On 11.02.2025 13:04, Juergen Gross wrote:
>>> In case xen_swiotlb_alloc_coherent() needed to create a contiguous
>>> region only for other reason than the memory not being compliant with
>>> the device's DMA mask, there is no reason why this contiguous region
>>> should be destroyed by xen_swiotlb_free_coherent() later. Destroying
>>> this region should be done only, if the memory of the region was
>>> allocated with more stringent placement requirements than the memory
>>> it did replace.
>>
>> I'm not convinced of this: Even the mere property of being contiguous
>> may already be enough to warrant freeing when possible. The hypervisor
>> may not have that many contiguous areas available. The bigger the
>> chunk, the more important to give it back once no longer needed in
>> this shape.
> 
> Really? When creating a domain Xen tries to use GB pages and 2MB pages as
> much as possible. Why would this special case here have more restrictions?

There aren't many Gb pages to be had from the space below 4Gb; frequently
there'll be at most one (covering the range from 1 to 2 Gb). For that as
well as 2Mb ones I think it is a mistake that Xen may hand them out, when
the caller could fall back to 4k allocations. Thing is that without extra
information it's hard to come up with a good heuristic to decide whether
the caller is capable of falling back. Perhaps e.g. populate_physmap()
should add MEMF_no_dma to higher order allocation requests targeting other
than the current domain, or when !d->creation_finished.

Jan