Skip to contentSkip to conversation

nvme-pci: disable SR-IOV VFs on driver unbind

PATCH In review Qinyun Tan 6 participants 10 messages Jan 27 – Feb 6

Subsystems (2)
lore
Share & tools
lore

Qinyun Tan

author
#0 lore raw

PATCH 1/1 nvme-pci: disable SR-IOV VFs on driver unbind

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

This causes the PCI core to emit a warning in pci_iov_remove():
"driver left SR-IOV enabled after remove"

According to Documentation/PCI/pci-iov-howto.rst, PCI drivers that support SR-IOV should call pci_disable_sriov() in their remove callback to properly clean up VFs before the driver is unloaded.

Fix this by disabling SR-IOV in nvme_remove(). If VFs are not assigned to a guest, disable SR-IOV. If VFs are still assigned, emit a warning since forcibly disabling would disrupt the guest.

---
 drivers/nvme/host/pci.c | 9 +++++++++
 1 file changed, 9 insertions(+)
drivers/nvme/host/pci.c +9 -0
index 58f3097888a7..4f2dc13de48b 100644
@@ -3666,6 +3666,15 @@ static void nvme_remove(struct pci_dev *pdev)
 	nvme_stop_ctrl(&dev->ctrl);
 	nvme_remove_namespaces(&dev->ctrl);
 	nvme_dev_disable(dev, true);
+
+	if (pci_num_vf(pdev)) {
+		if (pci_vfs_assigned(pdev))
+			dev_warn(&pdev->dev,
+				 "WARNING: Removing PF while VFs are assigned - VFs will not be deallocated!\n");
+		else
+			pci_disable_sriov(pdev);
+	}
+
 	nvme_free_host_mem(dev);
 	nvme_dev_remove_admin(dev);
 	nvme_dbbuf_dma_free(dev);
Signature · 1 line
2.43.5
#1 lore raw
Quoted 4 lines · Qinyun Tan

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

That sounds dangerous.

Quoted 3 lines

According to Documentation/PCI/pci-iov-howto.rst, PCI drivers that support SR-IOV should call pci_disable_sriov() in their remove callback to properly clean up VFs before the driver is unloaded.

Bjorn and other PCI folks: is there any reason to not do this in the PCI code and leave a landmine for the drivers?

Quoted 3 lines

Fix this by disabling SR-IOV in nvme_remove(). If VFs are not assigned to a guest, disable SR-IOV. If VFs are still assigned, emit a warning since forcibly disabling would disrupt the guest.

Well, I think we have to distrupt it, at least for hot unplug. This sounds like we need some better handling in the core code as well.

Quoted 22 lines
drivers/nvme/host/pci.c +9 -0
index 58f3097888a7..4f2dc13de48b 100644
@@ -3666,6 +3666,15 @@ static void nvme_remove(struct pci_dev *pdev)
 	nvme_stop_ctrl(&dev->ctrl);
 	nvme_remove_namespaces(&dev->ctrl);
 	nvme_dev_disable(dev, true);
+
+	if (pci_num_vf(pdev)) {
+		if (pci_vfs_assigned(pdev))
+			dev_warn(&pdev->dev,
+				 "WARNING: Removing PF while VFs are assigned - VFs will not be deallocated!\n");
+		else
+			pci_disable_sriov(pdev);
+	}
+
 	nvme_free_host_mem(dev);
 	nvme_dev_remove_admin(dev);
 	nvme_dbbuf_dma_free(dev);
Signature · 1 line
2.43.5

---end quoted text---

#2 lore raw
Quoted 7 lines · Christoph Hellwig
Quoted 4 lines · Qinyun Tan

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

That sounds dangerous.

It is not. In a real SR-IOV device, VFs are created by the hardware and are independent of their PF. There are several use cases where an operator unbinds the PF and reuses it to improve overall device utilization.

We have already discussed this in the context of Rust.
https://lore.kernel.org/all/20251122185701.GZ18335@unreal/

Quoted 4 lines
Quoted 4 lines

According to Documentation/PCI/pci-iov-howto.rst, PCI drivers that support SR-IOV should call pci_disable_sriov() in their remove callback to properly clean up VFs before the driver is unloaded.

I could not find that claim in Documentation/PCI/pci-iov-howto.rst. Can you point to the specific sentence that supports it?

Quoted 3 lines

Bjorn and other PCI folks: is there any reason to not do this in the PCI code and leave a landmine for the drivers?

It will break a lot of real users.

Quoted 7 lines
Quoted 4 lines

Fix this by disabling SR-IOV in nvme_remove(). If VFs are not assigned to a guest, disable SR-IOV. If VFs are still assigned, emit a warning since forcibly disabling would disrupt the guest.

Well, I think we have to distrupt it, at least for hot unplug. This sounds like we need some better handling in the core code as well.

As mentioned earlier, there are valid users of this functionality relying on legitimate devices that operate correctly regardless of whether the PF is bound.

Quoted 25 lines
Quoted 23 lines
drivers/nvme/host/pci.c +9 -0
index 58f3097888a7..4f2dc13de48b 100644
@@ -3666,6 +3666,15 @@ static void nvme_remove(struct pci_dev *pdev)
 	nvme_stop_ctrl(&dev->ctrl);
 	nvme_remove_namespaces(&dev->ctrl);
 	nvme_dev_disable(dev, true);
+
+	if (pci_num_vf(pdev)) {
+		if (pci_vfs_assigned(pdev))
+			dev_warn(&pdev->dev,
+				 "WARNING: Removing PF while VFs are assigned - VFs will not be deallocated!\n");
+		else
+			pci_disable_sriov(pdev);
+	}
+
 	nvme_free_host_mem(dev);
 	nvme_dev_remove_admin(dev);
 	nvme_dbbuf_dma_free(dev);
Signature · 1 line
2.43.5

---end quoted text---

maintainer
#3 lore raw
Quoted 7 lines · Leon Romanovsky
Quoted 6 lines · Christoph Hellwig
Quoted 4 lines · Qinyun Tan

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

Quoted 6 lines
Quoted 1 line

That sounds dangerous.

It is not. In a real SR-IOV device, VFs are created by the hardware and are independent of their PF. There are several use cases where an operator unbinds the PF and reuses it to improve overall device utilization.

If this is expected, should the warn message "driver left SR-IOV enabled after remove" be downgraded to 'info' level?

#4 lore raw
Quoted 10 lines · Keith Busch
Quoted 9 lines · Leon Romanovsky
Quoted 7 lines · Christoph Hellwig
Quoted 4 lines · Qinyun Tan

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

That sounds dangerous.

Quoted 7 lines
Quoted 4 lines

It is not. In a real SR-IOV device, VFs are created by the hardware and are independent of their PF. There are several use cases where an operator unbinds the PF and reuses it to improve overall device utilization.

If this is expected, should the warn message "driver left SR-IOV enabled after remove" be downgraded to 'info' level?

It is not important, no one complained about it. People who unbind PF, simply ignore this warning.

BTW, the use case which I presented is for SR-IOV handled by drivers. Maybe VFs created by NVMe are different here and they must be destroyed.

Thanks

#5 lore raw

[+cc Jakub, author of 38972375ef7b ("PCI/IOV: Reset total_VFs limit after detaching PF driver"), which added the message]

Quoted 19 lines · Leon Romanovsky
Quoted 10 lines · Keith Busch
Quoted 9 lines · Leon Romanovsky
Quoted 7 lines · Christoph Hellwig
Quoted 4 lines · Qinyun Tan

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

That sounds dangerous.

Quoted 7 lines
Quoted 4 lines

It is not. In a real SR-IOV device, VFs are created by the hardware and are independent of their PF. There are several use cases where an operator unbinds the PF and reuses it to improve overall device utilization.

If this is expected, should the warn message "driver left SR-IOV enabled after remove" be downgraded to 'info' level?

Quoted 6 lines

It is not important, no one complained about it. People who unbind PF, simply ignore this warning.

BTW, the use case which I presented is for SR-IOV handled by drivers. Maybe VFs created by NVMe are different here and they must be destroyed.

If it's to be expected, I do think 'info' would be more appropriate, if nothing else as an indication to code readers that nothing is wrong. Or maybe even no message at all. Or maybe the fact that we reset the driver_max_VFs value is of more interest.

Bjorn

#6 lore raw
Quoted 6 lines · On Tue, 27 Jan 2026 17:09:12 -0600 Bjorn Helgaas wrote:
Quoted 6 lines
Quoted 2 lines

If this is expected, should the warn message "driver left SR-IOV enabled after remove" be downgraded to 'info' level?

It is not important, no one complained about it. People who unbind PF, simply ignore this warning.

Quoted 8 lines
Quoted 3 lines

BTW, the use case which I presented is for SR-IOV handled by drivers. Maybe VFs created by NVMe are different here and they must be destroyed.

If it's to be expected, I do think 'info' would be more appropriate, if nothing else as an indication to code readers that nothing is wrong. Or maybe even no message at all. Or maybe the fact that we reset the driver_max_VFs value is of more interest.

FWIW I probably just added that message because most drivers end up printing something along those lines. Tho, I strongly suspect it's a mindless copy/paste in majority of the cases. No preference here.

#7 lore raw
In reply to Bjorn Helgaas
Quoted 26 lines · Bjorn Helgaas

[+cc Jakub, author of 38972375ef7b ("PCI/IOV: Reset total_VFs limit after detaching PF driver"), which added the message]

Quoted 22 lines · Leon Romanovsky
Quoted 10 lines · Keith Busch
Quoted 9 lines · Leon Romanovsky
Quoted 7 lines · Christoph Hellwig
Quoted 4 lines · Qinyun Tan

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

That sounds dangerous.

Quoted 7 lines
Quoted 4 lines

It is not. In a real SR-IOV device, VFs are created by the hardware and are independent of their PF. There are several use cases where an operator unbinds the PF and reuses it to improve overall device utilization.

If this is expected, should the warn message "driver left SR-IOV enabled after remove" be downgraded to 'info' level?

It is not important, no one complained about it. People who unbind PF, simply ignore this warning.

Quoted 7 lines
Quoted 3 lines

BTW, the use case which I presented is for SR-IOV handled by drivers. Maybe VFs created by NVMe are different here and they must be destroyed.

If it's to be expected, I do think 'info' would be more appropriate, if nothing else as an indication to code readers that nothing is wrong. Or maybe even no message at all.

The issue is that not all devices are created equal. Some devices from major vendor rely on messages sent from the VF to the PF, and expect the PF driver to perform certain configuration steps in response. These devices must disable SR-IOV when the PF is unbound.

From one side, for users of such hardware, this message is important. From another, for many other devices, this message is not important.

Like Jakub, I don't have any preference here.

Thanks

author
#8 lore raw
In reply to Christoph Hellwig

Hi All,

Thank you all for the insightful discussion!

I agree with Leon's point that not all devices are created equal when it comes to SR-IOV handling during driver unbind.

Looking at existing driver implementations, I found two different approaches:

1) mlx5 - unconditionally disables SR-IOV in remove:

   drivers/net/ethernet/mellanox/mlx5/core/main.c:
   static void remove_one(struct pci_dev *pdev)
   {
       ...
       mlx5_sriov_disable(pdev, false);
       ...
   }

   drivers/net/ethernet/mellanox/mlx5/core/sriov.c:
   void mlx5_sriov_disable(struct pci_dev *pdev, bool num_vf_change)
   {
       struct mlx5_core_dev *dev  = pci_get_drvdata(pdev);
       struct devlink *devlink = priv_to_devlink(dev);
       int num_vfs = pci_num_vf(dev->pdev);

pci_disable_sriov(pdev); /* Always disable, no
pci_vfs_assigned() check */

       devl_lock(devlink);
       mlx5_device_disable_sriov(dev, num_vfs, true, num_vf_change);
       devl_unlock(devlink);

}

2) ixgbe - checks pci_vfs_assigned() and skips disable if VFs are in use:

   drivers/net/ethernet/intel/ixgbe/ixgbe_main.c:
   static void ixgbe_remove(struct pci_dev *pdev)
   {
       ...
   #ifdef CONFIG_PCI_IOV
       ixgbe_disable_sriov(adapter);
   #endif
       ...
   }

drivers/net/ethernet/intel/ixgbe/ixgbe_sriov.c:
#ifdef CONFIG_PCI_IOV
if (pci_vfs_assigned(adapter->pdev)) {
e_dev_warn("Unloading driver while VFs are assigned - VFs

will not be deallocated\n");
           return -EPERM;
       }
       pci_disable_sriov(adapter->pdev);

#endif

Regarding the warning level discussion: I would prefer keeping it as dev_warn() rather than downgrading to dev_info(). As Leon mentioned, some devices do require SR-IOV to be disabled when the PF is unbound, and for those cases, this warning is important for operators to notice and take action. A warning level helps ensure it doesn't get lost in normal system logs.

Please let me know how you'd like to proceed.

Thanks,
Qinyun

Quoted 45 lines · Christoph Hellwig
Quoted 4 lines · Qinyun Tan

The NVMe PCI driver exports the sriov_configure callback via pci_sriov_configure_simple(), which allows userspace to enable SR-IOV VFs through sysfs. However, when the PF driver is unbound, the driver does not disable SR-IOV, leaving VFs orphaned in the system.

That sounds dangerous.

Quoted 3 lines

According to Documentation/PCI/pci-iov-howto.rst, PCI drivers that support SR-IOV should call pci_disable_sriov() in their remove callback to properly clean up VFs before the driver is unloaded.

Bjorn and other PCI folks: is there any reason to not do this in the PCI code and leave a landmine for the drivers?

Quoted 3 lines

Fix this by disabling SR-IOV in nvme_remove(). If VFs are not assigned to a guest, disable SR-IOV. If VFs are still assigned, emit a warning since forcibly disabling would disrupt the guest.

Well, I think we have to distrupt it, at least for hot unplug. This sounds like we need some better handling in the core code as well.

Quoted 22 lines
drivers/nvme/host/pci.c +9 -0
index 58f3097888a7..4f2dc13de48b 100644
@@ -3666,6 +3666,15 @@ static void nvme_remove(struct pci_dev *pdev)
  	nvme_stop_ctrl(&dev->ctrl);
  	nvme_remove_namespaces(&dev->ctrl);
  	nvme_dev_disable(dev, true);
+
+	if (pci_num_vf(pdev)) {
+		if (pci_vfs_assigned(pdev))
+			dev_warn(&pdev->dev,
+				 "WARNING: Removing PF while VFs are assigned - VFs will not be deallocated!\n");
+		else
+			pci_disable_sriov(pdev);
+	}
+
  	nvme_free_host_mem(dev);
  	nvme_dev_remove_admin(dev);
  	nvme_dbbuf_dma_free(dev);
Signature · 1 line
2.43.5

---end quoted text---

#9 lore raw
Quoted 56 lines · qinyuntan

Hi All,

Thank you all for the insightful discussion!

I agree with Leon's point that not all devices are created equal when it comes to SR-IOV handling during driver unbind.

Looking at existing driver implementations, I found two different approaches:

1) mlx5 - unconditionally disables SR-IOV in remove:

   drivers/net/ethernet/mellanox/mlx5/core/main.c:
   static void remove_one(struct pci_dev *pdev)
   {
       ...
       mlx5_sriov_disable(pdev, false);
       ...
   }

   drivers/net/ethernet/mellanox/mlx5/core/sriov.c:
   void mlx5_sriov_disable(struct pci_dev *pdev, bool num_vf_change)
   {
       struct mlx5_core_dev *dev  = pci_get_drvdata(pdev);
       struct devlink *devlink = priv_to_devlink(dev);
       int num_vfs = pci_num_vf(dev->pdev);

pci_disable_sriov(pdev); /* Always disable, no pci_vfs_assigned()
check */

       devl_lock(devlink);
       mlx5_device_disable_sriov(dev, num_vfs, true, num_vf_change);
       devl_unlock(devlink);

}

2) ixgbe - checks pci_vfs_assigned() and skips disable if VFs are in use:

   drivers/net/ethernet/intel/ixgbe/ixgbe_main.c:
   static void ixgbe_remove(struct pci_dev *pdev)
   {
       ...
   #ifdef CONFIG_PCI_IOV
       ixgbe_disable_sriov(adapter);
   #endif
       ...
   }
   drivers/net/ethernet/intel/ixgbe/ixgbe_sriov.c:
   #ifdef CONFIG_PCI_IOV
       if (pci_vfs_assigned(adapter->pdev)) {
           e_dev_warn("Unloading driver while VFs are assigned - VFs will
not be deallocated\n");
           return -EPERM;
       }
       pci_disable_sriov(adapter->pdev);

#endif

Quoted 8 lines

Regarding the warning level discussion: I would prefer keeping it as dev_warn() rather than downgrading to dev_info(). As Leon mentioned, some devices do require SR-IOV to be disabled when the PF is unbound, and for those cases, this warning is important for operators to notice and take action. A warning level helps ensure it doesn't get lost in normal system logs.

Please let me know how you'd like to proceed.

"driver left SR-IOV enabled after remove\n" is KERN_WARNING today, and I'm OK with leaving it that way.