Proposed subject:
PCI/IOV: Add dynamic MSI-X vector assignment sysfs interface
On Tue, Feb 16, 2021 at 09:33:44AM +0200, Leon Romanovsky wrote:
> On Mon, Feb 15, 2021 at 03:01:06PM -0600, Bjorn Helgaas wrote:
> > On Tue, Feb 09, 2021 at 03:34:42PM +0200, Leon Romanovsky wrote:
> > > From: Leon Romanovsky <leonro@xxxxxxxxxx>
Here's a draft of the sort of thing I'm looking for here:
A typical cloud provider SR-IOV use case is to create many VFs for
use by guest VMs. The VFs may not be assigned to a VM until a
customer requests a VM of a certain size, e.g., number of CPUs. A
VF may need MSI-X vectors proportional to the number of CPUs in the
VM, but there is no standard way to change the number of MSI-X
vectors supported by a VF.
Some Mellanox ConnectX devices support dynamic assignment of MSI-X
vectors to SR-IOV VFs. This can be done by the PF driver after VFs
are enabled, and it can be done without affecting VFs that are
already in use. The hardware supports a limited pool of MSI-X
vectors that can be assigned to the PF or to individual VFs. This
is device-specific behavior that requires support in the PF driver.
Add a read-only "sriov_vf_total_msix" sysfs file for the PF and a
writable "sriov_vf_msix_count" file for each VF. Management
software may use these to learn how many MSI-X vectors are available
and to dynamically assign them to VFs before the VFs are passed
through to a VM.
If the PF driver implements the ->sriov_get_vf_total_msix()
callback, "sriov_vf_total_msix" contains the total number of MSI-X
vectors available for distribution among VFs.
If no driver is bound to the VF, writing "N" to
"sriov_vf_msix_count" uses the PF driver ->sriov_set_msix_vec_count()
callback to assign "N" MSI-X vectors to the VF. When a VF driver
subsequently reads the MSI-X Message Control register, it will see
the new Table Size "N".
> > > Extend PCI sysfs interface with a new callback that allows configuration
> > > of the number of MSI-X vectors for specific SR-IOV VF. This is needed
> > > to optimize the performance of VFs devices by allocating the number of
> > > vectors based on the administrator knowledge of the intended use of the VF.
> > >
> > > This function is applicable for SR-IOV VF because such devices allocate
> > > their MSI-X table before they will run on the VMs and HW can't guess the
> > > right number of vectors, so some devices allocate them statically and equally.
> >
> > This commit log should be clear that this functionality is motivated
> > by *mlx5* behavior. The description above makes it sound like this is
> > generic PCI spec behavior, and it is not.
> >
> > It may be a reasonable design that conforms to the spec, and we hope
> > the model will be usable by other designs, but it is not required by
> > the spec and AFAIK there is nothing in the spec you can point to as
> > background for this.
> >
> > So don't *remove* the text you have above, but please *add* some
> > preceding background information about how mlx5 works.
> >
> > > 1) The newly added /sys/bus/pci/devices/.../sriov_vf_msix_count
> > > file will be seen for the VFs and it is writable as long as a driver is not
> > > bound to the VF.
> >
> > This adds /sys/bus/pci/devices/.../sriov_vf_msix_count for VF
> > devices and is writable ...
> >
> > > The values accepted are:
> > > * > 0 - this will be number reported by the Table Size in the VF's MSI-X Message
> > > Control register
> > > * < 0 - not valid
> > > * = 0 - will reset to the device default value
> >
> > = 0 - will reset to a device-specific default value
> >
> > > 2) In order to make management easy, provide new read-only sysfs file that
> > > returns a total number of possible to configure MSI-X vectors.
> >
> > For PF devices, this adds a read-only
> > /sys/bus/pci/devices/.../sriov_vf_total_msix file that contains the
> > total number of MSI-X vectors available for distribution among VFs.
> >
> > Just as in sysfs-bus-pci, this file should be listed first, because
> > you must read it before you can use vf_msix_count.
>
> No problem, I'll change, just remember that we are talking about commit
> message because in Documentation file, the order is exactly as you request.
Yes, I noticed that, thank you! It will be good to have them in the
same order in both the commit log and the Documentation file. I think
it will make more sense to readers.
> > > cat /sys/bus/pci/devices/.../sriov_vf_total_msix
> > > = 0 - feature is not supported
> > > > 0 - total number of MSI-X vectors available for distribution among the VFs
> > >
> > > Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxx>
> > > ---
> > > Documentation/ABI/testing/sysfs-bus-pci | 28 +++++
> > > drivers/pci/iov.c | 153 ++++++++++++++++++++++++
> > > include/linux/pci.h | 12 ++
> > > 3 files changed, 193 insertions(+)
>
> <...>
>
> > > + */
> > > +int pci_enable_vf_overlay(struct pci_dev *dev)
> > > +{
> > > + struct pci_dev *virtfn;
> > > + int id, ret;
> > > +
> > > + if (!dev->is_physfn || !dev->sriov->num_VFs)
> > > + return 0;
> > > +
> > > + ret = sysfs_create_files(&dev->dev.kobj, sriov_pf_dev_attrs);
> >
> > But I still don't like the fact that we're calling
> > sysfs_create_files() and sysfs_remove_files() directly. It makes
> > complication and opportunities for errors.
>
> It is not different from any other code that we have in the kernel.
It *is* different. There is a general rule that drivers should not
call sysfs_* [1]. The PCI core is arguably not a "driver," but it is
still true that callers of sysfs_create_files() are very special, and
I'd prefer not to add another one.
> Let's be concrete, can you point to the errors in this code that I
> should fix?
I'm not saying there are current errors; I'm saying the additional
code makes errors possible in future code. For example, we hope that
other drivers can use these sysfs interfaces, and it's possible they
may not call pci_enable_vf_overlay() or pci_disable_vfs_overlay()
correctly.
Or there may be races in device addition/removal. We have current
issues in this area, e.g., [2], and they're fairly subtle. I'm not
saying your patches have these issues; only that extra code makes more
chances for mistakes and it's more work to validate it.
> > I don't see the advantage of creating these files only when the PF
> > driver supports this. The management tools have to deal with
> > sriov_vf_total_msix == 0 and sriov_vf_msix_count == 0 anyway.
> > Having the sysfs files not be present at all might be slightly
> > prettier to the person running "ls", but I'm not sure the code
> > complication is worth that.
>
> It is more than "ls", right now sriov_numvfs is visible without relation
> to the driver, even if driver doesn't implement ".sriov_configure", which
> IMHO bad. We didn't want to repeat.
>
> Right now, we have many devices that supports SR-IOV, but small amount
> of them are capable to rewrite their VF MSI-X table siz. We don't want
> "to punish" and clatter their sysfs.
I agree, it's clutter, but at least it's just cosmetic clutter (but
I'm willing to hear discussion about why it's more than cosmetic; see
below).
>From the management software point of view, I don't think it matters.
That software already needs to deal with files that don't exist (on
old kernels) and files that contain zero (feature not supported or no
vectors are available).
>From my point of view, pci_enable_vf_overlay() or
pci_disable_vfs_overlay() are also clutter, at least compared to
static sysfs attributes.
> > I see a hint that Alex might have requested this "only visible when PF
> > driver supports it" functionality, but I don't see that email on
> > linux-pci, so I missed the background.
>
> First version of this patch had static files solution.
> https://lore.kernel.org/linux-pci/20210103082440.34994-2-leon@xxxxxxxxxx/#Z30drivers:pci:iov.c
Thanks for the pointer to the patch. Can you point me to the
discussion about why we should use the "only visible when PF driver
supports it" model?
Bjorn
[1] https://lore.kernel.org/linux-pci/YBmG7qgIDYIveDfX@xxxxxxxxx/
[2] https://lore.kernel.org/linux-pci/20200716110423.xtfyb3n6tn5ixedh@pali/
