Re: [RFC PATCH] vfio/pci: Allow force needs_pm_restore as specified by device:vendor

From: Colin Xu
Date: Mon Dec 14 2020 - 20:17:42 EST



On 12/10/20 1:09 AM, Alex Williamson wrote:
On Wed, 9 Dec 2020 13:14:00 +0800
Colin Xu <Colin.Xu@xxxxxxxxx> wrote:

On 11/27/20 11:35 PM, Alex Williamson wrote:
On Fri, 27 Nov 2020 11:53:39 +0800
Colin Xu <Colin.Xu@xxxxxxxxx> wrote:
On 11/25/20 11:53 PM, Alex Williamson wrote:
On Wed, 25 Nov 2020 10:18:24 +0800
Colin Xu <colin.xu@xxxxxxxxx> wrote:
Force specific device listed in params pm_restore_ids to follow
device state save/restore as needs_pm_restore.
Some device has NoSoftRst so will skip current state save/restore enabled
by needs_pm_restore. However once the device experienced power state
D3<->D0 transition, either by idle_d3 or the guest driver changes PM_CTL,
the guest driver won't get correct devie state although the configure
space doesn't change.
It sounds like you're describing a device that incorrectly exposes
NoSoftRst when there is in fact some sort of internal reset that
requires reprogramming config space. What device requires this? How
is a user to know when this option is required? It seems like this
would be better handled via a quirk in PCI core that sets a device flag
that the NoSoftRst value is incorrect for the specific affected
devices. Thanks,

Alex
Thanks for the feedback.

The device found are: Comet Lake PCH Serial IO I2C Controller
[8086:06e8]
[8086:06e9]

Yes you're right, there is no straight way for user to know the device.
The above device I found is during pass through them to VM. Although
adding such param may help in certain scenario, it still too
device-specific but not common in most cases.
The chipset i2c controller seems like a pretty suspicious device for
Intel to advocate assigning to a VM. Are you assigning this to satisfy
the isolation issue that we often see where a device like a NIC is
grouped together with platform management devices due to lack of
multifunction ACS? If that's the case, I would think it would make
more sense to investigate from the perspective of whether there is
actually DMA isolation between those integrated, multifunction devices
and if so, implement ACS quirks to expose that isolation. Thanks,

Alex
Hi Alex,

Sorry for late reply. E-mail incorrectly filtered so didn't see this one
until manual search.

The mentioned two I2C controller are in same iommu group and there is NO
other device in the same group. The I2C controller is integrated in PCH
chipset and there are other devices integrated too, but in different
group. When assigning them to a VM, both are assigned, and function 0 is
set with multifunction=on. If iommu driver group no other device in the
same group, could we assume there is no DMA isolation issue?

Yes, we should always have DMA isolation between IOMMU groups. My
concern is that I understand these i2c devices to generally provide
access to system/chipset management features, therefore by assigning
them to a VM you're granting that VM not only access to a single
device, but potentially management features which can affect the host
system. This would generally not be advised for security reasons, so
these configurations are usually only created due to IOMMU grouping
restrictions. It seems here that you're intentionally assigning these
devices to a VM. Why? If this is a development exercise to support
creation of drivers for these devices in a VM, great! I'm just trying
to figure out if Intel is endorsing this configuration for some
generally useful purpose. Thanks,

Alex

No, passing through such device is not for some common purpose. The using scenario is a single VM case. Like iGFX via VT-d, only one VM will use that device and host won't use it. As you mentioned if there are multiple VMs, pass through chipset device to one VM could affect other system resource. If above scenario there should be no such concern.

Back to the original question there are no other VMs nor host system could affect the I2C device state. In native environment the i2c controller driver (intel-lpss) doesn't have to save/restore the device. However in VT-d case, once vfio-pci or intel-lpss in guest did the power state transition, the state is lost unless save/restore in vfio-pci. It may be related to iommu, also may related to some ACPI methods of the I2C. When passing through it, I didn't duplicate all ACPI methods of the I2C to guest, only expose necessary resource so that the I2C device in guest could find it. Still doesn't have clue yet. I think I'll check if some missing _PS0/_PS3 method implementation in guest ACPI could cause this.

Cc: Swee Yee Fonn <swee.yee.fonn@xxxxxxxxx>
Signed-off-by: Colin Xu <colin.xu@xxxxxxxxx>
---
drivers/vfio/pci/vfio_pci.c | 66 ++++++++++++++++++++++++++++++++++++-
1 file changed, 65 insertions(+), 1 deletion(-)

diff --git a/drivers/vfio/pci/vfio_pci.c b/drivers/vfio/pci/vfio_pci.c
index e6190173482c..50a4141c9e1d 100644
--- a/drivers/vfio/pci/vfio_pci.c
+++ b/drivers/vfio/pci/vfio_pci.c
@@ -34,6 +34,15 @@
#define DRIVER_AUTHOR "Alex Williamson <alex.williamson@xxxxxxxxxx>"
#define DRIVER_DESC "VFIO PCI - User Level meta-driver"
+#define VFIO_MAX_PM_DEV 32
+struct vfio_pm_devs {
+ struct {
+ unsigned short vendor;
+ unsigned short device;
+ } ids[VFIO_MAX_PM_DEV];
+ u32 count;
+};
+
static char ids[1024] __initdata;
module_param_string(ids, ids, sizeof(ids), 0);
MODULE_PARM_DESC(ids, "Initial PCI IDs to add to the vfio driver, format is \"vendor:device[:subvendor[:subdevice[:class[:class_mask]]]]\" and multiple comma separated entries can be specified");
@@ -64,6 +73,10 @@ static bool disable_denylist;
module_param(disable_denylist, bool, 0444);
MODULE_PARM_DESC(disable_denylist, "Disable use of device denylist. Disabling the denylist allows binding to devices with known errata that may lead to exploitable stability or security issues when accessed by untrusted users.");
+static char pm_restore_ids[1024] __initdata;
+module_param_string(pm_restore_ids, pm_restore_ids, sizeof(pm_restore_ids), 0);
+MODULE_PARM_DESC(pm_restore_ids, "comma separated device in format of \"vendor:device\"");
+
static inline bool vfio_vga_disabled(void)
{
#ifdef CONFIG_VFIO_PCI_VGA
@@ -260,10 +273,50 @@ static bool vfio_pci_nointx(struct pci_dev *pdev)
return false;
}
+static struct vfio_pm_devs pm_devs = {0};
+static void __init vfio_pci_fill_pm_ids(void)
+{
+ char *p, *id;
+ int idx = 0;
+
+ /* no ids passed actually */
+ if (pm_restore_ids[0] == '\0')
+ return;
+
+ /* add ids specified in the module parameter */
+ p = pm_restore_ids;
+ while ((id = strsep(&p, ","))) {
+ unsigned int vendor, device = PCI_ANY_ID;
+ int fields;
+
+ if (!strlen(id))
+ continue;
+
+ fields = sscanf(id, "%x:%x", &vendor, &device);
+
+ if (fields != 2) {
+ pr_warn("invalid vendor:device string \"%s\"\n", id);
+ continue;
+ }
+
+ if (idx < VFIO_MAX_PM_DEV) {
+ pm_devs.ids[idx].vendor = vendor;
+ pm_devs.ids[idx].device = device;
+ pm_devs.count++;
+ idx++;
+ pr_info("add [%04x:%04x] for needs_pm_restore\n",
+ vendor, device);
+ } else {
+ pr_warn("Exceed maximum %d, skip adding [%04x:%04x] for needs_pm_restore\n",
+ VFIO_MAX_PM_DEV, vendor, device);
+ }
+ }
+}
+
static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
{
struct pci_dev *pdev = vdev->pdev;
- u16 pmcsr;
+ u16 pmcsr, idx;
if (!pdev->pm_cap)
return;
@@ -271,6 +324,16 @@ static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
+
+ for (idx = 0; idx < pm_devs.count; idx++) {
+ if (vdev->pdev->vendor == pm_devs.ids[idx].vendor &&
+ vdev->pdev->device == pm_devs.ids[idx].device) {
+ vdev->needs_pm_restore = true;
+ pr_info("force [%04x:%04x] to needs_pm_restore\n",
+ vdev->pdev->vendor, vdev->pdev->device);
+ break;
+ }
+ }
}
/*
@@ -2423,6 +2486,7 @@ static int __init vfio_pci_init(void)
goto out_driver;
vfio_pci_fill_ids();
+ vfio_pci_fill_pm_ids();
if (disable_denylist)
pr_warn("device denylist disabled.\n");

--
Best Regards,
Colin Xu