Hi Leon,
On 2021/1/6 14:06, Leon Romanovsky wrote:
On Wed, Jan 06, 2021 at 10:27:49AM +0800, Lu Baolu wrote:
The pci_subdevice_msi_create_irq_domain() should fail if the underlying
platform is not able to support IMS (Interrupt Message Storage). Otherwise,
the isolation of interrupt is not guaranteed.
For x86, IMS is only supported on bare metal for now. We could enable it
in the virtualization environments in the future if interrupt HYPERCALL
domain is supported or the hardware has the capability of interrupt
isolation for subdevices.
Suggested-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Link: https://lore.kernel.org/linux-pci/87pn4nk7nn.fsf@xxxxxxxxxxxxxxxxxxxxxxx/
Link: https://lore.kernel.org/linux-pci/877dqrnzr3.fsf@xxxxxxxxxxxxxxxxxxxxxxx/
Link: https://lore.kernel.org/linux-pci/877dqqmc2h.fsf@xxxxxxxxxxxxxxxxxxxxxxx/
Signed-off-by: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx>
---
arch/x86/pci/common.c | 47 +++++++++++++++++++++++++++++++++++++
drivers/base/platform-msi.c | 8 +++++++
include/linux/msi.h | 1 +
3 files changed, 56 insertions(+)
Background:
Learnt from the discussions in this thread:
https://lore.kernel.org/linux-pci/160408357912.912050.17005584526266191420.stgit@xxxxxxxxxxxxxxxxxxxxxxxxxx/
The device IMS (Interrupt Message Storage) should not be enabled in any
virtualization environments unless there is a HYPERCALL domain which
makes the changes in the message store managed by the hypervisor.
As the initial step, we allow the IMS to be enabled only if we are
running on the bare metal. It's easy to enable IMS in the virtualization
environments if above preconditions are met in the future.
We ever thought about moving on_bare_metal() to a generic file so that
it could be well maintained and used. But we need some suggestions about
where to put it. Your comments are very appreciated.
This patch is only for comments purpose. Please don't merge it. We will
include it in the Intel IMS implementation later once we reach a
consensus.
Change log:
v1->v2:
- v1:
https://lore.kernel.org/linux-pci/20201210004624.345282-1-baolu.lu@xxxxxxxxxxxxxxx/
- Rename probably_on_bare_metal() with on_bare_metal();
- Some vendors might use the same name for both bare metal and virtual
environment. Before we add vendor specific code to distinguish
between them, let's return false in on_bare_metal(). This won't
introduce any regression. The only impact is that the coming new
platform msi feature won't be supported until the vendor specific code
is provided.
Best regards,
baolu
diff --git a/arch/x86/pci/common.c b/arch/x86/pci/common.c
index 3507f456fcd0..963e0401f2b2 100644
--- a/arch/x86/pci/common.c
+++ b/arch/x86/pci/common.c
@@ -724,3 +724,50 @@ struct pci_dev *pci_real_dma_dev(struct pci_dev *dev)
return dev;
}
#endif
+
+/*
+ * We want to figure out which context we are running in. But the hardware
+ * does not introduce a reliable way (instruction, CPUID leaf, MSR, whatever)
+ * which can be manipulated by the VMM to let the OS figure out where it runs.
+ * So we go with the below probably on_bare_metal() function as a replacement
+ * for definitely on_bare_metal() to go forward only for the very simple reason
+ * that this is the only option we have.
+ *
+ * People might use the same vendor name for both bare metal and virtual
+ * environment. We can remove those names once we have vendor specific code to
+ * distinguish between them.
+ */
+static const char * const vmm_vendor_name[] = {
+ "QEMU", "Bochs", "KVM", "Xen", "VMware", "VMW", "VMware Inc.",
+ "innotek GmbH", "Oracle Corporation", "Parallels", "BHYVE",
+ "Microsoft Corporation", "Amazon EC2"
+};
Maybe it is not concern at all, but this approach will make
forward/backward compatibility without kernel upgrade impossible.
Once QEMU (example) will have needed support, someone will need to remove
the QEMU from this array, rewrite on_bare_metal() because it is not bare
vs. virtual anymore and require kernel upgrade/downgrade every time QEMU
version is switched.
Plus need to update stable@ and distros.
I'm already feeling pain from the fields while they debug such code.
Am I missing it completely?
The basic need here is that we want to disallow a brand new feature
(device ims) to be enabled in any VMM environment.
The cpuid (X86_FEATURE_HYPERVISOR) is a good choice, but it's optional
and even not documented. So besides it, we maintain a block list
(vmm_vendor_name) which lists all possible VMM vendor names. If
dmi_match(DMI_SYS_VENDOR) hits, the new feature is not allowed to be
enabled.
This block list is a bit overkill since some vendor names could also be
used on bare metal. We will delay enabling the new feature for those
cases until we have a vendor-specific way to distinguish between bare
metal and VMM environments.
Honestly speaking, I can't see any compatible issue as it's common that
a new feature is supported in a new kernel but not in an old one.
Best regards,
baolu
