Re: [PATCH kernel] PCI: Disable IOV before pcibios_sriov_disable()

From: Alexey Kardashevskiy
Date: Mon Aug 21 2017 - 00:01:10 EST


On 19/08/17 01:27, Bjorn Helgaas wrote:
> On Fri, Aug 18, 2017 at 08:05:42AM +1000, Alexey Kardashevskiy wrote:
>> On 11/08/17 18:19, Alexey Kardashevskiy wrote:
>>> From: Gavin Shan <gwshan@xxxxxxxxxxxxxxxxxx>
>>>
>>> The PowerNV platform is the only user of pcibios_sriov_disable().
>>> The IOV BAR could be shifted by pci_iov_update_resource(). The
>>> warning message in the function is printed if the IOV capability
>>> is in enabled (PCI_SRIOV_CTRL_VFE && PCI_SRIOV_CTRL_MSE) state.
>>>
>>> This is the backtrace of what is happening:
>>> pci_disable_sriov
>>> sriov_disable
>>> pnv_pci_sriov_disable
>>> pnv_pci_vf_resource_shift
>>> pci_update_resource
>>> pci_iov_update_resource
>>>
>>> This fixes the issue by disabling IOV capability before calling
>>> pcibios_sriov_disable(). With it, the disabling path matches
>>> the enabling path: pcibios_sriov_enable() is called before the
>>> IOV capability is enabled.
>>>
>>> Cc: shan.gavin@xxxxxxxxx
>>> Cc: Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx>
>>> Cc: Paul Mackerras <paulus@xxxxxxxxx>
>>> Reported-by: Carol L Soto <clsoto@xxxxxxxxxx>
>>> Signed-off-by: Gavin Shan <gwshan@xxxxxxxxxxxxxxxxxx>
>>> Tested-by: Carol L Soto <clsoto@xxxxxxxxxx>
>>> Signed-off-by: Alexey Kardashevskiy <aik@xxxxxxxxx>
>>> ---
>>>
>>> This is repost. Since Gavin left the team, I am trying to push it out.
>>> The previos converstion is here: https://patchwork.ozlabs.org/patch/732653/
>>>
>>> Two questions were raised then. I'll try to comment on this below.
>>
>> Bjorn, ping? Thanks.
>
> Thanks for the reminder. This is in patchwork, so it's on my to-do
> list.
>
> My last response in the thread above was:
>
> I'm not going to merge this without a comment in
> pnv_pci_vf_resource_shift() that addresses the two questions I
> raised in my very first response. I don't think the existing
> comment about "After doing so, there would be a 'hole'" is
> sufficient. If it were sufficient, I wouldn't have raised the
> questions in the first place.
>
> The problem here is that I'm looking for a comment *in the code*, and
> you and Gavin are giving responses and clarifications in email.

I understand.

> What we need to do is transfer this email information into something
> useful when reading the code, i.e., a comment in the code.

I agree. Here are few problems though.

1. I am not sure if there is nothing to fix (like adding dummy regions to
prevent these "holes" to be used for something else)

2. I do not really know how exactly to update the comment in the code to
make it clearer for the future readers, and before doing this, I'd like to
understand about 1).


An example of the comment which would make sense to me:

@@ -1002,9 +1016,13 @@ static int pnv_pci_vf_resource_shift(struct pci_dev
*dev, int offset)
}

/*
- * After doing so, there would be a "hole" in the /proc/iomem when
- * offset is a positive value. It looks like the device return some
- * mmio back to the system, which actually no one could use it.
+ * Since M64 BAR shares segments among all possible 256 PEs,
+ * we have to shift the beginning of PF IOV BAR to make it start from
+ * the segment which belongs to the PE number assigned to the first VF.
+ * This creates a "hole" in the /proc/iomem which could be used for
+ * allocating other resources, however this is not expected to happen
+ * on IODA as the only possibility would be a PCI hotplug and IODA
+ * hardware only allows it on a slot with dedicated PHB.
*/


Does this make sense? Thanks.



>
>>>> 1) "res" is already in the resource tree, so we shouldn't be changing
>>>> its start address, because that may make the tree inconsistent,
>>>> e.g., the resource may no longer be completely contained in its
>>>> parent, it may conflict with a sibling, etc.
>>>
>>> We should not, yes. But...
>>>
>>> At the boot time IOV BAR gets as much MMIO space as it can possibly use.
>>> (Embarassingly I cannot trace where this is coming from, 8GB is selected
>>> via pci_assign_unassigned_root_bus_resources() path somehow).
>>> For example, it is 256*32MB=8GB where 256 is maximum PEs number and 32MB
>>> is a PF/VF BAR size. Whatever shifting we do afterwards, the boudaries of
>>> that 8GB area do not change and we test it in pnv_pci_vf_resource_shift():
>>>
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/platforms/powernv/pci-ioda.c#n987
>>>
>>>> 2) If we update "res->start", shouldn't we update "res->end"
>>>> correspondingly?
>>>
>>> We have to update the PF's IOV BAR address as we allocate PEs dynamically
>>> and we do not know in advance where our VF numbers start in that
>>> 8GB window. So we change IOV BASR start. Changing the end may make it
>>> look more like there is a free area to use but in reality it won't be
>>> usable as well as the area we "release" by shifting the start address.
>>>
>>> We could probably move that M64 MMIO window by the same delta in
>>> opposite direction so the IOV BAR start address would remain the same
>>> but its VF#0 would be mapped to let's say PF#5. I am just afraid there
>>> is an alignment requirement for these M64 window start address; and this
>>> would be even more tricky to manage.
>>>
>>> We could also create reserved areas for the amount of space "release" by
>>> moving the start address, not sure how though.
>>>
>>> So how do we proceed with this particular patch now? Thanks.
>>> ---
>>> drivers/pci/iov.c | 7 ++++---
>>> 1 file changed, 4 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/pci/iov.c b/drivers/pci/iov.c
>>> index 120485d6f352..ac41c8be9200 100644
>>> --- a/drivers/pci/iov.c
>>> +++ b/drivers/pci/iov.c
>>> @@ -331,7 +331,6 @@ static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
>>> while (i--)
>>> pci_iov_remove_virtfn(dev, i, 0);
>>>
>>> - pcibios_sriov_disable(dev);
>>> err_pcibios:
>>> iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
>>> pci_cfg_access_lock(dev);
>>> @@ -339,6 +338,8 @@ static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
>>> ssleep(1);
>>> pci_cfg_access_unlock(dev);
>>>
>>> + pcibios_sriov_disable(dev);
>>> +
>>> if (iov->link != dev->devfn)
>>> sysfs_remove_link(&dev->dev.kobj, "dep_link");
>>>
>>> @@ -357,14 +358,14 @@ static void sriov_disable(struct pci_dev *dev)
>>> for (i = 0; i < iov->num_VFs; i++)
>>> pci_iov_remove_virtfn(dev, i, 0);
>>>
>>> - pcibios_sriov_disable(dev);
>>> -
>>> 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);
>>> ssleep(1);
>>> pci_cfg_access_unlock(dev);
>>>
>>> + pcibios_sriov_disable(dev);
>>> +
>>> if (iov->link != dev->devfn)
>>> sysfs_remove_link(&dev->dev.kobj, "dep_link");
>>>
>>>
>>
>>
>> --
>> Alexey


--
Alexey