Re: [PATCH] dom0/pvh: fix processing softirqs during memory map population

[Jan Beulich <jbeulich@xxxxxxxx>]

On 07.02.2022 10:51, Roger Pau Monné wrote:
> On Mon, Feb 07, 2022 at 09:21:11AM +0100, Jan Beulich wrote:
>> On 05.02.2022 11:18, Roger Pau Monne wrote:
>>> --- a/xen/arch/x86/hvm/dom0_build.c
>>> +++ b/xen/arch/x86/hvm/dom0_build.c
>>> @@ -186,7 +186,7 @@ static int __init pvh_populate_memory_range(struct 
>>> domain *d,
>>>          start += 1UL << order;
>>>          nr_pages -= 1UL << order;
>>>          order_stats[order]++;
>>> -        if ( (++i % MAP_MAX_ITER) == 0 )
>>> +        if ( (i++ % MAP_MAX_ITER) == 0 )
>>>              process_pending_softirqs();
>>>      }
>>
>> This way is perhaps easiest, so
>>
>> Acked-by: Jan Beulich <jbeulich@xxxxxxxx>
>>
>> but I'd like you to consider to avoid doing this on the first
>> iteration. How about keeping the code here as is, but instead
>> insert an invocation in the sole caller (and there unconditionally
>> at the end of every successful loop iteration)?
> 
> In fact I was thinking that we should call process_pending_softirqs on
> every iteration: the calls to guest_physmap_add_page could use a 1G
> page order, so if not using sync-pt (at least until your series for
> IOMMU super-page support is committed) mapping a whole 1G page using
> 4K chunks on the IOMMU page-tables could be quite time consuming, and
> hence we would likely need to process softirqs on every iteration.

Good point; please do so.

Jan