[PATCH v4 5/7] PCI/IOV: Check that VF BAR fits within the reservation

Wed Oct 30 16:55:01 UTC 2024

On Wed, Oct 30, 2024 at 12:43:19PM +0100, Michał Winiarski wrote:
> On Mon, Oct 28, 2024 at 11:56:04AM -0500, Bjorn Helgaas wrote:
> > On Fri, Oct 25, 2024 at 11:50:36PM +0200, Michał Winiarski wrote:
> > > VF MMIO resource reservation, either created by system firmware and
> > > inherited by Linux PCI subsystem or created by the subsystem itself,
> > > should contain enough space to fit the BAR of all SR-IOV Virtual
> > > Functions that can potentially be created (total VFs supported by the
> > > device).
> > 
> > I don't think "VF resource reservation ... should contain enough
> > space" is really accurate or actionable.  It would be *nice* if the PF
> > BAR is large enough to accommodate the largest supported VF BARs for
> > all possible VFs, but if it doesn't, it's not really an error.  It's
> > just a reflection of the fact that resource space is limited.
> 
> From PCI perspective, you're right, IOV resources are optional, and it's
> not really an error for PF device itself.
> From IOV perspective - we do need those resources to be able to create
> VFs.
> 
> All I'm trying to say here, is that the context of the change is the
> "success" case, where the VF BAR reservation was successfully assigned,
> and the PF will be able to create VFs.
> The case where there were not enough resources for VF BAR (and PF won't
> be able to create VFs) remains unchanged.
> 
> > > However, that assumption only holds in an environment where VF BAR size
> > > can't be modified.
> > 
> > There's no reason to assume anything about how many VF BARs fit.  The
> > existing code should avoid enabling the requested nr_virtfn VFs if the
> > PF doesn't have enough space -- I think that's what the "if
> > (res->parent)" is supposed to be checking.
> > 
> > The fact that you need a change here makes me suspect that we're
> > missing some resource claim (and corresponding res->parent update)
> > elsewhere when resizing the VF BAR.
> 
> My understanding is that res->parent is only expressing that the
> resource is assigned.
> We don't really want to change that, the resource is still there and is
> assigned - we just want to make sure that VF enabling fails if the
> caller wants to enable more VFs than possible for current resource size.
> 
> Let's use an example. A device with a single BAR.
> initial_vf_bar_size = X
> total_vfs = 4
> supported_vf_resizable_bar_sizes = X, 2X, 4X

In addition, IIUC we're assuming the PF BAR size is 4X, since the
conclusion is that 4 VF BARs of size X fill it completely.

> With that - the initial underlying resource looks like this:
>             +----------------------+
>             |+--------------------+|
>             ||                    ||
>             |+--------------------+|
>             |+--------------------+|
>             ||                    ||
>             |+--------------------+|
>             |+--------------------+|
>             ||                    ||
>             |+--------------------+|
>             |+--------------------+|
>             ||                    ||
>             |+--------------------+|
>             +----------------------+
> Its size is 4X, and it contains BAR for 4 VFs.
> "resource_size >= vf_bar_size * num_vfs" is true for any num_vfs
> Let's assume that there are enough resources to assign it.
> 
> Patch 4/7 allows to resize the entire resource (in addition to changing
> the VF BAR size), which means that after calling:
> pci_resize_resource() with size = 2X, the underlying resource will look
> like this:
>             +----------------------+ 
>             |+--------------------+| 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             |+--------------------+| 
>             |+--------------------+| 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             |+--------------------+| 
>             |+--------------------+| 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             |+--------------------+| 
>             |+--------------------+| 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             ||                    || 
>             |+--------------------+| 
>             +----------------------+ 
> Its size is 8X, and it contains BAR for 4 VFs.
> "resource_size >= vf_bar_size * num_vfs" is true for any num_vfs

With the assumption that the PF BAR size is 4X, these VFs would no
longer fit.  I guess that's basically what you say here:

> It does require an extra 4X of MMIO resources, so this can fail in
> resource constrained environment, even though the original 4X resource
> was able to be assigned.
> 
> The following patch 6/7 allows to change VF BAR size without touching
> the underlying reservation size.
> After calling pci_iov_vf_bar_set_size() to 4X and enabling a single VF,
> the underlying resource will look like this:
>             +----------------------+ 
>             |+--------------------+| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             ||░░░░░░░░░░░░░░░░░░░░|| 
>             |+--------------------+| 
>             +----------------------+ 
> Its size is 4X, but since pci_iov_vf_bar_set_size() was called, it is no
> longer able to accomodate 4 VFs.
> "resource_size >= vf_bar_size * num_vfs" is only true for num_vfs = 1
> and any attempts to create more than 1 VF should fail.
> We don't need to worry about being MMIO resource constrained, no extra
> MMIO resources are needed.

IIUC this series only resizes VF BARs.  Those VF BARs are carved out
of a PF BAR, and this series doesn't touch the PF BAR resizing path.
I guess the driver might be able to increase the PF BAR size if
necessary, and then increase the VF BAR size.

It sounds like this patch is really a bug fix independent of VF BAR
resizing.  If we currently allow enabling more VFs than will fit in a
PF BAR, that sounds like a bug.

So if we try to enable too many VFs, sriov_enable() should fail.  I
still don't see why this check should change the res->parent test,
though.

> > > Add an additional check that verifies that VF BAR for all enabled VFs
> > > fits within the underlying reservation resource.
> > > 
> > > Signed-off-by: Michał Winiarski <michal.winiarski at intel.com>
> > > ---
> > >  drivers/pci/iov.c | 8 ++++++--
> > >  1 file changed, 6 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/drivers/pci/iov.c b/drivers/pci/iov.c
> > > index 79143c1bc7bb4..5de828e5a26ea 100644
> > > --- a/drivers/pci/iov.c
> > > +++ b/drivers/pci/iov.c
> > > @@ -645,10 +645,14 @@ static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
> > >  
> > >  	nres = 0;
> > >  	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
> > > +		int vf_bar_sz = pci_iov_resource_size(dev,
> > > +						      pci_resource_to_iov(i));
> > >  		bars |= (1 << pci_resource_to_iov(i));
> > >  		res = &dev->resource[pci_resource_to_iov(i)];
> > > -		if (res->parent)
> > > -			nres++;
> > > +		if (!res->parent || vf_bar_sz * nr_virtfn > resource_size(res))
> > > +			continue;
> > > +
> > > +		nres++;
> > >  	}
> > >  	if (nres != iov->nres) {
> > >  		pci_err(dev, "not enough MMIO resources for SR-IOV\n");
> > > -- 
> > > 2.47.0
> > > 