[kernel] PCI: Disable IOV before pcibios_sriov_disable()

Message ID 20170811081933.17474-1-aik@ozlabs.ru
State Accepted
Headers show

Commit Message

Alexey Kardashevskiy Aug. 11, 2017, 8:19 a.m.
From: Gavin Shan <gwshan@linux.vnet.ibm.com>

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@gmail.com
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Reported-by: Carol L Soto <clsoto@us.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Tested-by: Carol L Soto <clsoto@us.ibm.com>
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
---

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.

>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(-)

Comments

Alexey Kardashevskiy Aug. 17, 2017, 10:05 p.m. | #1
On 11/08/17 18:19, Alexey Kardashevskiy wrote:
> From: Gavin Shan <gwshan@linux.vnet.ibm.com>
> 
> 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@gmail.com
> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> Cc: Paul Mackerras <paulus@samba.org>
> Reported-by: Carol L Soto <clsoto@us.ibm.com>
> Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
> Tested-by: Carol L Soto <clsoto@us.ibm.com>
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> ---
> 
> 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.

> 
>> 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");
>  
>
Bjorn Helgaas Aug. 18, 2017, 3:27 p.m. | #2
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@linux.vnet.ibm.com>
> > 
> > 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@gmail.com
> > Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> > Cc: Paul Mackerras <paulus@samba.org>
> > Reported-by: Carol L Soto <clsoto@us.ibm.com>
> > Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
> > Tested-by: Carol L Soto <clsoto@us.ibm.com>
> > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> > ---
> > 
> > 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.

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

> >> 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 Kardashevskiy Aug. 21, 2017, 4 a.m. | #3
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@linux.vnet.ibm.com>
>>>
>>> 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@gmail.com
>>> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
>>> Cc: Paul Mackerras <paulus@samba.org>
>>> Reported-by: Carol L Soto <clsoto@us.ibm.com>
>>> Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
>>> Tested-by: Carol L Soto <clsoto@us.ibm.com>
>>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>>> ---
>>>
>>> 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
Bjorn Helgaas Aug. 30, 2017, 7:02 p.m. | #4
On Fri, Aug 11, 2017 at 06:19:33PM +1000, Alexey Kardashevskiy wrote:
> From: Gavin Shan <gwshan@linux.vnet.ibm.com>
> 
> 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@gmail.com
> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> Cc: Paul Mackerras <paulus@samba.org>
> Reported-by: Carol L Soto <clsoto@us.ibm.com>
> Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
> Tested-by: Carol L Soto <clsoto@us.ibm.com>
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> ---
> 
> 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/

I gave up on the previous issue.  I think this patch makes sense as-is
at least as far as the fact that we can't update a struct resource
while the device is still consuming it.  I reworked the changelog to
emphasize that.

I assume the fact that pci_iov_update_resource() dropped the resource
update caused some user-visible issue later on, and I might mention
that, too, if I knew what it was.

Here's what I would consider putting on pci/virtualization (the diff
is unchanged from your post):


commit 08132e7759b3929bea0ccdf8afe81ebf05351389
Author: Gavin Shan <gwshan@linux.vnet.ibm.com>
Date:   Fri Aug 11 18:19:33 2017 +1000

    PCI: Disable VF decoding before updating resources in pcibios_sriov_disable()
    
    A struct resource represents the address space consumed by a device.  We
    should not modify that resource while the device is actively using the
    address space.  For VFs, pci_iov_update_resource() enforces this by
    printing a warning and doing nothing if the VFE (VF Enable) and MSE (VF
    Memory Space Enable) bits are set.
    
    Previously, both sriov_enable() and sriov_disable() called the
    pcibios_sriov_disable() arch hook, which may update the struct resource,
    while VFE and MSE were enabled.  This effectively dropped the resource
    update pcibios_sriov_disable() intended to do.
    
    Disable VF memory decoding before calling pcibios_sriov_disable().
    
    Reported-by: Carol L Soto <clsoto@us.ibm.com>
    Tested-by: Carol L Soto <clsoto@us.ibm.com>
    Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
    Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
    [bhelgaas: changelog]
    Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
    Cc: shan.gavin@gmail.com
    Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
    Cc: Paul Mackerras <paulus@samba.org>

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 Kardashevskiy Aug. 31, 2017, 3:08 a.m. | #5
On 31/08/17 05:02, Bjorn Helgaas wrote:
> On Fri, Aug 11, 2017 at 06:19:33PM +1000, Alexey Kardashevskiy wrote:
>> From: Gavin Shan <gwshan@linux.vnet.ibm.com>
>>
>> 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@gmail.com
>> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
>> Cc: Paul Mackerras <paulus@samba.org>
>> Reported-by: Carol L Soto <clsoto@us.ibm.com>
>> Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
>> Tested-by: Carol L Soto <clsoto@us.ibm.com>
>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>> ---
>>
>> 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/
> 
> I gave up on the previous issue.  I think this patch makes sense as-is
> at least as far as the fact that we can't update a struct resource
> while the device is still consuming it.  I reworked the changelog to
> emphasize that.
> 
> I assume the fact that pci_iov_update_resource() dropped the resource
> update caused some user-visible issue later on, and I might mention
> that, too, if I knew what it was.

I could not identify any issue so far in my test setup - I recreated VFs
several times, run some traffic through them on one of mellanox'es so the
message+backtrace seems to be the only issue for now.


> Here's what I would consider putting on pci/virtualization (the diff
> is unchanged from your post):


This sounds good to me, thanks for updating the commit log.

I'll still try and finish my homework with updating that comment about the
hole in arch/powerpc/platforms/powernv/pci-ioda.c.

Cheers.

> 
> 
> commit 08132e7759b3929bea0ccdf8afe81ebf05351389
> Author: Gavin Shan <gwshan@linux.vnet.ibm.com>
> Date:   Fri Aug 11 18:19:33 2017 +1000
> 
>     PCI: Disable VF decoding before updating resources in pcibios_sriov_disable()
>     
>     A struct resource represents the address space consumed by a device.  We
>     should not modify that resource while the device is actively using the
>     address space.  For VFs, pci_iov_update_resource() enforces this by
>     printing a warning and doing nothing if the VFE (VF Enable) and MSE (VF
>     Memory Space Enable) bits are set.
>     
>     Previously, both sriov_enable() and sriov_disable() called the
>     pcibios_sriov_disable() arch hook, which may update the struct resource,
>     while VFE and MSE were enabled.  This effectively dropped the resource
>     update pcibios_sriov_disable() intended to do.
>     
>     Disable VF memory decoding before calling pcibios_sriov_disable().
>     
>     Reported-by: Carol L Soto <clsoto@us.ibm.com>
>     Tested-by: Carol L Soto <clsoto@us.ibm.com>
>     Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
>     Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>     [bhelgaas: changelog]
>     Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
>     Cc: shan.gavin@gmail.com
>     Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
>     Cc: Paul Mackerras <paulus@samba.org>

> 
> 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");
>  
>
Bjorn Helgaas Aug. 31, 2017, 5:04 p.m. | #6
On Wed, Aug 30, 2017 at 02:02:06PM -0500, Bjorn Helgaas wrote:
> On Fri, Aug 11, 2017 at 06:19:33PM +1000, Alexey Kardashevskiy wrote:
> > From: Gavin Shan <gwshan@linux.vnet.ibm.com>
> > 
> > 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@gmail.com
> > Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> > Cc: Paul Mackerras <paulus@samba.org>
> > Reported-by: Carol L Soto <clsoto@us.ibm.com>
> > Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
> > Tested-by: Carol L Soto <clsoto@us.ibm.com>
> > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> > ---
> > 
> > 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/
> 
> I gave up on the previous issue.  I think this patch makes sense as-is
> at least as far as the fact that we can't update a struct resource
> while the device is still consuming it.  I reworked the changelog to
> emphasize that.
> 
> I assume the fact that pci_iov_update_resource() dropped the resource
> update caused some user-visible issue later on, and I might mention
> that, too, if I knew what it was.
> 
> Here's what I would consider putting on pci/virtualization (the diff
> is unchanged from your post):

I applied the patch below on pci/virtualization for v4.14.

> commit 08132e7759b3929bea0ccdf8afe81ebf05351389
> Author: Gavin Shan <gwshan@linux.vnet.ibm.com>
> Date:   Fri Aug 11 18:19:33 2017 +1000
> 
>     PCI: Disable VF decoding before updating resources in pcibios_sriov_disable()
>     
>     A struct resource represents the address space consumed by a device.  We
>     should not modify that resource while the device is actively using the
>     address space.  For VFs, pci_iov_update_resource() enforces this by
>     printing a warning and doing nothing if the VFE (VF Enable) and MSE (VF
>     Memory Space Enable) bits are set.
>     
>     Previously, both sriov_enable() and sriov_disable() called the
>     pcibios_sriov_disable() arch hook, which may update the struct resource,
>     while VFE and MSE were enabled.  This effectively dropped the resource
>     update pcibios_sriov_disable() intended to do.
>     
>     Disable VF memory decoding before calling pcibios_sriov_disable().
>     
>     Reported-by: Carol L Soto <clsoto@us.ibm.com>
>     Tested-by: Carol L Soto <clsoto@us.ibm.com>
>     Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
>     Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>     [bhelgaas: changelog]
>     Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
>     Cc: shan.gavin@gmail.com
>     Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
>     Cc: Paul Mackerras <paulus@samba.org>
> 
> 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");
>

Patch

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");