Re: [PATCH 11/11] xen/arm: Process pending vPCI map/unmap operations

[Oleksandr Andrushchenko <Oleksandr_Andrushchenko@xxxxxxxx>]

Hi, Julien!

On 03.09.21 12:04, Julien Grall wrote:
> Hi Oleksandr,
>
> On 03/09/2021 09:33, Oleksandr Andrushchenko wrote:
>> From: Oleksandr Andrushchenko <oleksandr_andrushchenko@xxxxxxxx>
>>
>> vPCI may map and unmap PCI device memory (BARs) being passed through which
>> may take a lot of time. For this those operations may be deferred to be
>> performed later, so that they can be safely preempted.
>> Run the corresponding vPCI code while switching a vCPU.
>
> IIUC, you are talking about the function map_range() in 
> xen/drivers/vpci/header. The function has the following todo for Arm:
>
>         /*
>          * ARM TODOs:
>          * - On ARM whether the memory is prefetchable or not should be passed
>          *   to map_mmio_regions in order to decide which memory attributes
>          *   should be used.
>          *
>          * - {un}map_mmio_regions doesn't support preemption.
>          */
>
> This doesn't seem to be addressed in the two series for PCI passthrough sent 
> so far. Do you have any plan to handle it?

No plan yet.

All the mappings are happening with p2m_mmio_direct_dev as of now.

>
>>
>> Signed-off-by: Oleksandr Andrushchenko <oleksandr_andrushchenko@xxxxxxxx>
>> ---
>>   xen/arch/arm/traps.c | 6 ++++++
>>   1 file changed, 6 insertions(+)
>>
>> diff --git a/xen/arch/arm/traps.c b/xen/arch/arm/traps.c
>> index 219ab3c3fbde..1571fb8afd03 100644
>> --- a/xen/arch/arm/traps.c
>> +++ b/xen/arch/arm/traps.c
>> @@ -34,6 +34,7 @@
>>   #include <xen/symbols.h>
>>   #include <xen/version.h>
>>   #include <xen/virtual_region.h>
>> +#include <xen/vpci.h>
>>     #include <public/sched.h>
>>   #include <public/xen.h>
>> @@ -2304,6 +2305,11 @@ static bool check_for_vcpu_work(void)
>>       }
>>   #endif
>>   +    local_irq_enable();
>> +    if ( has_vpci(v->domain) && vpci_process_pending(v) )
>
> Looking at the code of vpci_process_pending(), it looks like there are some 
> rework to do for guest. Do you plan to handle it as part of the vPCI series?
Yes, vPCI code is heavily touched to support guest non-identity mappings
>
>> +        raise_softirq(SCHEDULE_SOFTIRQ);
>> +    local_irq_disable();
>> +
>
> From my understanding of vcpi_process_pending(). The function will return 
> true if there are more work to schedule.
Yes
> However, if check_for_vcpu_for_work() return false, then we will return to 
> the guest before any work for vCPI has finished. This is because 
> check_for_vcpu_work() will not be called again.
Correct
>
> In this case, I think you want to return as soon as you know we need to 
> reschedule.
Not sure I understand this
>
> However, looking at the rest of the code, we already have a check for vpci in 
> the common IOREQ code.

Which may not be enabled as it depends on CONFIG_IOREQ_SERVER.

My understanding is that for x86 it is always enabled, but this might not be 
the case for Arm

> So we would end up to call twice vpci_process_pending().
So, if CONFIG_IOREQ_SERVER is not enabled then we end up with only calling it 
from traps.c on Arm
> Maybe we should move the call from the IOREQ to arch-code.

Hm. I would better think of moving it from IOREQ to some other common code: for 
x86 (if

my understanding correct about CONFIG_IOREQ_SERVER) it is by coincidence that 
we call vPCI

code from there and IOREQ is always enabled.

>
> Cheers,
>
Thank you,

Oleksandr