Re: [PATCH v2 4/8] vpci: add a wait operation to the vpci vcpu pending actions

[Mykyta Poturai <Mykyta_Poturai@epam.com>]

On 3/31/26 17:55, Jan Beulich wrote:
> On 09.03.2026 12:08, Mykyta Poturai wrote:
>> --- a/xen/drivers/vpci/header.c
>> +++ b/xen/drivers/vpci/header.c
>> @@ -175,76 +175,92 @@ static void modify_decoding(const struct pci_dev *pdev, uint16_t cmd,
>>   
>>   bool vpci_process_pending(struct vcpu *v)
>>   {
>> -    const struct pci_dev *pdev = v->vpci.pdev;
>> -    struct vpci_header *header = NULL;
>> -    unsigned int i;
>> -
>> -    if ( !pdev )
>> -        return false;
>> -
>> -    read_lock(&v->domain->pci_lock);
>> -
>> -    if ( !pdev->vpci || (v->domain != pdev->domain) )
>> +    switch ( v->vpci.task )
>>       {
>> -        v->vpci.pdev = NULL;
>> -        read_unlock(&v->domain->pci_lock);
>> -        return false;
>> -    }
>> -
>> -    header = &pdev->vpci->header;
>> -    for ( i = 0; i < ARRAY_SIZE(header->bars); i++ )
>> +    case MODIFY_MEMORY:
>>       {
>> -        struct vpci_bar *bar = &header->bars[i];
>> -        struct rangeset *mem = v->vpci.bar_mem[i];
>> -        struct map_data data = {
>> -            .d = v->domain,
>> -            .map = v->vpci.cmd & PCI_COMMAND_MEMORY,
>> -            .bar = bar,
>> -        };
>> -        int rc;
>> +        const struct pci_dev *pdev = v->vpci.memory.pdev;
>> +        struct vpci_header *header = NULL;
>> +        unsigned int i;
>>   
>> -        if ( rangeset_is_empty(mem) )
>> -            continue;
>> +        if ( !pdev )
>> +            break;
>>   
>> -        rc = rangeset_consume_ranges(mem, map_range, &data);
>> +        read_lock(&v->domain->pci_lock);
>>   
>> -        if ( rc == -ERESTART )
>> +        if ( !pdev->vpci || (v->domain != pdev->domain) )
>>           {
>> +            v->vpci.memory.pdev = NULL;
>>               read_unlock(&v->domain->pci_lock);
>> -            return true;
>> +            break;
>>           }
>>   
>> -        if ( rc )
>> +        header = &pdev->vpci->header;
>> +        for ( i = 0; i < ARRAY_SIZE(header->bars); i++ )
>>           {
>> -            spin_lock(&pdev->vpci->lock);
>> -            /* Disable memory decoding unconditionally on failure. */
>> -            modify_decoding(pdev, v->vpci.cmd & ~PCI_COMMAND_MEMORY,
>> -                            false);
>> -            spin_unlock(&pdev->vpci->lock);
>> +            struct vpci_bar *bar = &header->bars[i];
>> +            struct rangeset *mem = v->vpci.bar_mem[i];
>> +            struct map_data data = {
>> +                .d = v->domain,
>> +                .map = v->vpci.memory.cmd & PCI_COMMAND_MEMORY,
>> +                .bar = bar,
>> +            };
>> +            int rc;
>> +
>> +            if ( rangeset_is_empty(mem) )
>> +                continue;
>>   
>> -            /* Clean all the rangesets */
>> -            for ( i = 0; i < ARRAY_SIZE(header->bars); i++ )
>> -                if ( !rangeset_is_empty(v->vpci.bar_mem[i]) )
>> -                     rangeset_purge(v->vpci.bar_mem[i]);
>> +            rc = rangeset_consume_ranges(mem, map_range, &data);
>>   
>> -            v->vpci.pdev = NULL;
>> +            if ( rc == -ERESTART )
>> +            {
>> +                read_unlock(&v->domain->pci_lock);
>> +                return true;
>> +            }
>>   
>> -            read_unlock(&v->domain->pci_lock);
>> +            if ( rc )
>> +            {
>> +                spin_lock(&pdev->vpci->lock);
>> +                /* Disable memory decoding unconditionally on failure. */
>> +                modify_decoding(pdev, v->vpci.memory.cmd & ~PCI_COMMAND_MEMORY,
>> +                                false);
>> +                spin_unlock(&pdev->vpci->lock);
>> +
>> +                /* Clean all the rangesets */
>> +                for ( i = 0; i < ARRAY_SIZE(header->bars); i++ )
>> +                    if ( !rangeset_is_empty(v->vpci.bar_mem[i]) )
>> +                        rangeset_purge(v->vpci.bar_mem[i]);
>> +
>> +                v->vpci.memory.pdev = NULL;
>> +
>> +                read_unlock(&v->domain->pci_lock);
>>   
>> -            if ( !is_hardware_domain(v->domain) )
>> -                domain_crash(v->domain);
>> +                if ( !is_hardware_domain(v->domain) )
>> +                    domain_crash(v->domain);
>>   
>> -            return false;
>> +                break;
>> +            }
>>           }
>> -    }
>> -    v->vpci.pdev = NULL;
>> +        v->vpci.memory.pdev = NULL;
>>   
>> -    spin_lock(&pdev->vpci->lock);
>> -    modify_decoding(pdev, v->vpci.cmd, v->vpci.rom_only);
>> -    spin_unlock(&pdev->vpci->lock);
>> +        spin_lock(&pdev->vpci->lock);
>> +        modify_decoding(pdev, v->vpci.memory.cmd, v->vpci.memory.rom_only);
>> +        spin_unlock(&pdev->vpci->lock);
>>   
>> -    read_unlock(&v->domain->pci_lock);
>> +        read_unlock(&v->domain->pci_lock);
>> +
>> +        break;
>> +    }
>> +    case WAIT:
>> +        if ( NOW() < v->vpci.wait.end )
>> +            return true;
>> +        v->vpci.wait.callback(v->vpci.wait.data);
>> +        break;
> 
> As just indicated in reply to patch 6, busy waiting isn't really acceptable.
> This is even more so when the waiting exceeds the typical length of a
> scheduling timeslice.
> 
> In that other reply I said to put the vCPU to sleep, but you need to be careful
> there too: The domain may not expect its vCPU to not make any progress for such
> an extended period of time. This may need doing entirely differently: Once the
> command register was written, you may want to record the time after which
> accesses to the VF registers are permitted. Earlier accesses would simply be
> terminated. You may still additionally need a timer, in order to kick off BAR
> mapping after that time. (Yet better would  be if the BAR mapping could be
> done during those 100ms. After all that may be a reason why this long a delay
> is specified: Firmware on the device may also require some time to set up the
> BARs accordingly.)
> 
> Jan

I am not sure it would work that way. If we look at how linux 
initialized sriov, it writes VFE and MSE bits, waits 100ms and then 
expects VFs to be operational. If they are not operational at that 
moment, then it considers the operation failed and removes all VFs. If 
we also wait 100ms before enabling access, the probability of a guest 
trying to access something before we allow it would be very high.

So I think there is no way to add VFs in Xen without blocking the 
guest’s vCPU in some way. We can revert back to the old variant and rely 
on physdev op to add VFs one by one as they are discovered by Dom0, then 
we will not need to explicitly wait.
@Roger are you okay with that?


Snippet from Linux:

static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
{
	...
	pci_iov_set_numvfs(dev, nr_virtfn);
	iov->ctrl |= PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE;
	pci_cfg_access_lock(dev);
	pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
	msleep(100);
	pci_cfg_access_unlock(dev);

	rc = sriov_add_vfs(dev, initial);
	if (rc)
		goto err_pcibios;
	...
}





-- 
Mykyta