Re: [PATCH] [RFC] vpci: allow BAR write while mapped

[Jan Beulich <jbeulich@xxxxxxxx>]

On 14.03.2025 09:21, Roger Pau Monné wrote:
> On Fri, Mar 14, 2025 at 09:04:55AM +0100, Jan Beulich wrote:
>> On 13.03.2025 18:48, Roger Pau Monné wrote:
>>> On Wed, Mar 12, 2025 at 03:50:17PM -0400, Stewart Hildebrand wrote:
>>>> @@ -585,10 +601,31 @@ static void cf_check bar_write(
>>>>      {
>>>>          /* If the value written is the current one avoid printing a 
>>>> warning. */
>>>>          if ( val != (uint32_t)(bar->addr >> (hi ? 32 : 0)) )
>>>> +        {
>>>>              gprintk(XENLOG_WARNING,
>>>> -                    "%pp: ignored BAR %zu write while mapped\n",
>>>> +                    "%pp: allowing BAR %zu write while mapped\n",
>>>>                      &pdev->sbdf, bar - pdev->vpci->header.bars + hi);
>>>
>>> If Xen now handles BARs writes with memory decoding enabled the
>>> message can be removed.  It's only purpose was to signal this missing
>>> handling.
>>>
>>>> -        return;
>>>> +            ASSERT(rw_is_write_locked(&pdev->domain->pci_lock));
>>>> +            ASSERT(spin_is_locked(&pdev->vpci->lock));
>>>> +            reenable = true;
>>>> +            cmd = pci_conf_read16(pdev->sbdf, PCI_COMMAND);
>>>> +            /*
>>>> +             * Write-while-mapped: unmap the old BAR in p2m. We want this 
>>>> to
>>>> +             * finish right away since the deferral machinery only 
>>>> supports
>>>> +             * unmap OR map, not unmap-then-remap. Ultimately, it 
>>>> probably would
>>>> +             * be better to defer the write-while-mapped case just like 
>>>> regular
>>>> +             * BAR writes (but still only allow it for 32-bit BAR writes).
>>>> +             */
>>>> +            /* Disable memory decoding */
>>>> +            modify_bars(pdev, cmd & ~PCI_COMMAND_MEMORY, false);
>>>
>>> I think if the guest explicitly avoids toggling memory decoding Xen
>>> should also to the same, and not touch the bit.
>>
>> For Dom0 I'm inclined to agree, but for DomU-s it may be unsafe to do so.
>> (You may have meant it like this, but you said "guest".)
> 
> Sorry, I'm not sure I'm following.  For domUs the BAR register write
> is not propagated to the hardware, it's just the p2m mappings that
> change, and hence it's even safer to not toggle the memory decoding
> bit for that case? (as there's no write to the device BAR registers
> for domUs).

Oh, right. I'm sorry for the noise.

Jan