RE: [PATCH 2/3] PCI: Enable host bridge emulation for PCI_DOMAINS_GENERIC platforms

[dan.j.williams]

Michael Kelley wrote:
> From: Dan Williams <dan.j.williams@xxxxxxxxx> Sent: Wednesday, July 16, 2025 9:09 AM

Thanks for taking a look Michael!

[..]
> > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> > index e9448d55113b..833ebf2d5213 100644
> > --- a/drivers/pci/pci.c
> > +++ b/drivers/pci/pci.c
> > @@ -6692,9 +6692,50 @@ static void pci_no_domains(void)
> >  #endif
> >  }
> > 
> > +#ifdef CONFIG_PCI_DOMAINS
> > +static DEFINE_IDA(pci_domain_nr_dynamic_ida);
> > +
> > +/*
> > + * Find a free domain_nr either allocated by pci_domain_nr_dynamic_ida or
> > + * fallback to the first free domain number above the last ACPI segment number.
> > + * Caller may have a specific domain number in mind, in which case try to
> > + * reserve it.
> > + *
> > + * Note that this allocation is freed by pci_release_host_bridge_dev().
> > + */
> > +int pci_bus_find_emul_domain_nr(int hint)
> > +{
> > +	if (hint >= 0) {
> > +		hint = ida_alloc_range(&pci_domain_nr_dynamic_ida, hint, hint,
> > +				       GFP_KERNEL);
> 
> This almost preserves the existing functionality in pci-hyperv.c. But if the
> "hint" passed in is zero, current code in pci-hyperv.c treats that as a
> collision and allocates some other value. The special treatment of zero is
> necessary per the comment with the definition of HVPCI_DOM_INVALID.
> 
> I don't have an opinion on whether the code here should treat a "hint"
> of zero as invalid, or whether that should be handled in pci-hyperv.c.

Oh, I see what you are saying. I made the "hint == 0" case start working
where previously it should have failed. I feel like that's probably best
handled in pci-hyperv.c with something like the following which also
fixes up a regression I caused with @dom being unsigned:

diff --git a/drivers/pci/controller/pci-hyperv.c b/drivers/pci/controller/pci-hyperv.c
index cfe9806bdbe4..813757db98d1 100644
--- a/drivers/pci/controller/pci-hyperv.c
+++ b/drivers/pci/controller/pci-hyperv.c
@@ -3642,9 +3642,9 @@ static int hv_pci_probe(struct hv_device *hdev,
 {
 	struct pci_host_bridge *bridge;
 	struct hv_pcibus_device *hbus;
-	u16 dom_req, dom;
+	int ret, dom = -EINVAL;
+	u16 dom_req;
 	char *name;
-	int ret;
 
 	bridge = devm_pci_alloc_host_bridge(&hdev->device, 0);
 	if (!bridge)
@@ -3673,7 +3673,8 @@ static int hv_pci_probe(struct hv_device *hdev,
 	 * collisions) in the same VM.
 	 */
 	dom_req = hdev->dev_instance.b[5] << 8 | hdev->dev_instance.b[4];
-	dom = pci_bus_find_emul_domain_nr(dom_req);
+	if (dom_req)
+		dom = pci_bus_find_emul_domain_nr(dom_req);
 
 	if (dom < 0) {
 		dev_err(&hdev->device,

> > +
> > +		if (hint >= 0)
> > +			return hint;
> > +	}
> > +
> > +	if (acpi_disabled)
> > +		return ida_alloc(&pci_domain_nr_dynamic_ida, GFP_KERNEL);
> > +
> > +	/*
> > +	 * Emulated domains start at 0x10000 to not clash with ACPI _SEG
> > +	 * domains.  Per ACPI r6.0, sec 6.5.6,  _SEG returns an integer, of
> > +	 * which the lower 16 bits are the PCI Segment Group (domain) number.
> > +	 * Other bits are currently reserved.
> > +	 */
> 
> Back in 2018 and 2019, the Microsoft Linux team encountered problems with
> PCI domain IDs that exceeded 0xFFFF. User space code, such as the Xorg X server,
> assumed PCI domain IDs were at most 16 bits, and retained only the low 16 bits
> if the value was larger. My memory of the details is vague, but I believe some
> or all of this behavior was tied to libpciaccess. As a result of these user space
> limitations, the pci-hyperv.c code made sure to not create any domain IDs
> larger than 0xFFFF. The problem was not just theoretical -- Microsoft had
> customers reporting issues due to the "32-bit domain ID problem" and the
> pci-hyperv.c code was updated to avoid it.
> 
> I don't have information on whether user space code has been fixed, or
> the extent to which such a fix has propagated into distro versions. At the
> least, a VM with old user space code might break if the kernel is upgraded
> to one with this patch. How do people see the risks now that it is 6 years
> later? I don't have enough data to make an assessment.

A couple observations:

- I think it would be reasonable to not fallback in the hint case with
  something like this:

diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index 833ebf2d5213..0bd2053dbe8a 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -6705,14 +6705,10 @@ static DEFINE_IDA(pci_domain_nr_dynamic_ida);
  */
 int pci_bus_find_emul_domain_nr(int hint)
 {
-	if (hint >= 0) {
-		hint = ida_alloc_range(&pci_domain_nr_dynamic_ida, hint, hint,
+	if (hint >= 0)
+		return ida_alloc_range(&pci_domain_nr_dynamic_ida, hint, hint,
 				       GFP_KERNEL);
 
-		if (hint >= 0)
-			return hint;
-	}
-
 	if (acpi_disabled)
 		return ida_alloc(&pci_domain_nr_dynamic_ida, GFP_KERNEL);
 
- The VMD driver has been allocating 32-bit PCI domain numbers since
  v4.5 185a383ada2e ("x86/PCI: Add driver for Intel Volume Management
  Device (VMD)"). At a minimum if it is still a problem, it is a shared
  problem, but the significant deployment of VMD in the time likely
  indicates it is ok. If not, the above change at least makes the
  hyper-v case avoid 32-bit domain numbers.