1. Patchew
  2. linux
  3. View series

[PATCH] PCI: hide mysterious 8MB 64-bit pref BAR on Intel Arc PCIe Switch

Icenowy Zheng posted 1 patch 2 months, 2 weeks ago 
drivers/pci/quirks.c | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
[PATCH] PCI: hide mysterious 8MB 64-bit pref BAR on Intel Arc PCIe Switch
Posted by Icenowy Zheng 2 months, 2 weeks ago 
The upstream port device of Intel Arc series dGPUs' internal PCIe switch
contains a mysterious 8MB 64-bit prefetchable BAR. All reads to memory
mapped to that BAR returns 0xFFFFFFFF and writes have no effect.

When the PCI bus isn't configured by any firmware (e.g. a PCIe
controller solely initialized by Linux kernel), the PCI space allocation
algorithm of Linux will allocate the main VRAM BAR of Arc dGPU device at
base+0, and then the 8MB BAR at base+256M, which prevents the main VRAM
BAR gets resized. As the functionality and performance of Arc dGPU will
get severely restricted with small BAR, this makes a problem.

Hide the mysterious 8MB BAR to Linux PCI subsystem, to allow resizing
the VRAM BAR to VRAM size with the Linux PCI space allocation algorithm.

Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
---
 drivers/pci/quirks.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
index d7f4ee634263..df304bfec6e9 100644
--- a/drivers/pci/quirks.c
+++ b/drivers/pci/quirks.c
@@ -3650,6 +3650,22 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d0, quirk_broken_intx_masking);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d1, quirk_broken_intx_masking);
 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d2, quirk_broken_intx_masking);
 
+/*
+ * Intel Arc dGPUs' internal switch upstream port contains a mysterious 8MB
+ * 64-bit prefetchable BAR that blocks resize of main dGPU VRAM BAR with
+ * Linux's PCI space allocation algorithm.
+ */
+static void quirk_intel_xe_upstream(struct pci_dev *pdev)
+{
+	memset(&pdev->resource[0], 0, sizeof(pdev->resource[0]));
+}
+/* Intel Arc A380 PCI Express Switch Upstream Port */
+DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa1, quirk_intel_xe_upstream);
+/* Intel Arc A770 PCI Express Switch Upstream Port */
+DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa0, quirk_intel_xe_upstream);
+/* Intel Arc B580 PCI Express Switch Upstream Port */
+DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0xe2ff, quirk_intel_xe_upstream);
+
 static u16 mellanox_broken_intx_devs[] = {
 	PCI_DEVICE_ID_MELLANOX_HERMON_SDR,
 	PCI_DEVICE_ID_MELLANOX_HERMON_DDR,
-- 
2.50.1
Re: [PATCH] PCI: hide mysterious 8MB 64-bit pref BAR on Intel Arc PCIe Switch
Posted by Vivian Wang 2 months, 2 weeks ago 
On 7/22/25 01:30, Icenowy Zheng wrote:

> The upstream port device of Intel Arc series dGPUs' internal PCIe switch
> contains a mysterious 8MB 64-bit prefetchable BAR. All reads to memory
> mapped to that BAR returns 0xFFFFFFFF and writes have no effect.

Reading all-ones is certainly not a good indication that this BAR does
nothing. My guess would be that maybe it's some write-only registers
like MSI doorbells.

AMD has docs on how their GPUs have a tiny doorbell default-2M 64bit
pref BAR [1], which I find interesting, even though very indirect as
evidence.

[1]: https://rocm.docs.amd.com/en/latest/how-to/Bar-Memory.html#bar-configuration-for-amd-gpus

> When the PCI bus isn't configured by any firmware (e.g. a PCIe
> controller solely initialized by Linux kernel), the PCI space allocation
> algorithm of Linux will allocate the main VRAM BAR of Arc dGPU device at
> base+0, and then the 8MB BAR at base+256M, which prevents the main VRAM
> BAR gets resized. As the functionality and performance of Arc dGPU will
> get severely restricted with small BAR, this makes a problem.

While trying to look around for if any documentation is available on the
"mysterious BAR", I found that this problem has indeed been discovered
before with no satisfactory resolution. Linking for posterity:

[2]: https://issues.redhat.com/browse/RHEL-3672
[3]: https://github.com/raspberrypi/linux/issues/6621

If my doorbell theory was correct, and that this is analogous for
amdgpu, there is this interesting comment in
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c since commit d6895ad39f3b
("drm/amdgpu: resize VRAM BAR for CPU access v6"):

  /* Free the VRAM and doorbell BAR, we most likely need to move both. */

That seems to be the right way *if* it makes sense that the "mysterious
BAR" is supposed to be managed by the GPU driver. Of course, only
someone from Intel can confirm for sure what's going on...

Vivian "dramforever" Wang

> Hide the mysterious 8MB BAR to Linux PCI subsystem, to allow resizing
> the VRAM BAR to VRAM size with the Linux PCI space allocation algorithm.
>
> Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
> ---
>  drivers/pci/quirks.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
>
> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> index d7f4ee634263..df304bfec6e9 100644
> --- a/drivers/pci/quirks.c
> +++ b/drivers/pci/quirks.c
> @@ -3650,6 +3650,22 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d0, quirk_broken_intx_masking);
>  DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d1, quirk_broken_intx_masking);
>  DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d2, quirk_broken_intx_masking);
>  
> +/*
> + * Intel Arc dGPUs' internal switch upstream port contains a mysterious 8MB
> + * 64-bit prefetchable BAR that blocks resize of main dGPU VRAM BAR with
> + * Linux's PCI space allocation algorithm.
> + */
> +static void quirk_intel_xe_upstream(struct pci_dev *pdev)
> +{
> +	memset(&pdev->resource[0], 0, sizeof(pdev->resource[0]));
> +}
> +/* Intel Arc A380 PCI Express Switch Upstream Port */
> +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa1, quirk_intel_xe_upstream);
> +/* Intel Arc A770 PCI Express Switch Upstream Port */
> +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa0, quirk_intel_xe_upstream);
> +/* Intel Arc B580 PCI Express Switch Upstream Port */
> +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0xe2ff, quirk_intel_xe_upstream);
> +
>  static u16 mellanox_broken_intx_devs[] = {
>  	PCI_DEVICE_ID_MELLANOX_HERMON_SDR,
>  	PCI_DEVICE_ID_MELLANOX_HERMON_DDR,
Re: [PATCH] PCI: hide mysterious 8MB 64-bit pref BAR on Intel Arc PCIe Switch
Posted by Bjorn Helgaas 2 months, 2 weeks ago 
[+cc Ilpo]

On Tue, Jul 22, 2025 at 01:30:57AM +0800, Icenowy Zheng wrote:
> The upstream port device of Intel Arc series dGPUs' internal PCIe switch
> contains a mysterious 8MB 64-bit prefetchable BAR. All reads to memory
> mapped to that BAR returns 0xFFFFFFFF and writes have no effect.
> 
> When the PCI bus isn't configured by any firmware (e.g. a PCIe
> controller solely initialized by Linux kernel), the PCI space allocation
> algorithm of Linux will allocate the main VRAM BAR of Arc dGPU device at
> base+0, and then the 8MB BAR at base+256M, which prevents the main VRAM
> BAR gets resized. As the functionality and performance of Arc dGPU will
> get severely restricted with small BAR, this makes a problem.
> 
> Hide the mysterious 8MB BAR to Linux PCI subsystem, to allow resizing
> the VRAM BAR to VRAM size with the Linux PCI space allocation algorithm.

There's no reason a switch upstream port should not have a BAR.  I
suspect this BAR probably does have a legitimate purpose, and it's
only "mysterious" because we don't know how to use it.

This sounds like it may be a deficiency in the Linux BAR assignment
code.  Any other device could have a similar problem.  

> Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
> ---
>  drivers/pci/quirks.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
> 
> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> index d7f4ee634263..df304bfec6e9 100644
> --- a/drivers/pci/quirks.c
> +++ b/drivers/pci/quirks.c
> @@ -3650,6 +3650,22 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d0, quirk_broken_intx_masking);
>  DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d1, quirk_broken_intx_masking);
>  DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d2, quirk_broken_intx_masking);
>  
> +/*
> + * Intel Arc dGPUs' internal switch upstream port contains a mysterious 8MB
> + * 64-bit prefetchable BAR that blocks resize of main dGPU VRAM BAR with
> + * Linux's PCI space allocation algorithm.
> + */
> +static void quirk_intel_xe_upstream(struct pci_dev *pdev)
> +{
> +	memset(&pdev->resource[0], 0, sizeof(pdev->resource[0]));

This doesn't touch the BAR itself, so we may be leaving the BAR
decoding accesses, which could lead to an address conflict.  It also
prevents a driver for the upstream port from using the BAR.

> +}
> +/* Intel Arc A380 PCI Express Switch Upstream Port */
> +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa1, quirk_intel_xe_upstream);
> +/* Intel Arc A770 PCI Express Switch Upstream Port */
> +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa0, quirk_intel_xe_upstream);
> +/* Intel Arc B580 PCI Express Switch Upstream Port */
> +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0xe2ff, quirk_intel_xe_upstream);
> +
>  static u16 mellanox_broken_intx_devs[] = {
>  	PCI_DEVICE_ID_MELLANOX_HERMON_SDR,
>  	PCI_DEVICE_ID_MELLANOX_HERMON_DDR,
> -- 
> 2.50.1
>
Re: [PATCH] PCI: hide mysterious 8MB 64-bit pref BAR on Intel Arc PCIe Switch
Posted by Ilpo Järvinen 2 months, 2 weeks ago 
On Mon, 21 Jul 2025, Bjorn Helgaas wrote:

> [+cc Ilpo]
> 
> On Tue, Jul 22, 2025 at 01:30:57AM +0800, Icenowy Zheng wrote:
> > The upstream port device of Intel Arc series dGPUs' internal PCIe switch
> > contains a mysterious 8MB 64-bit prefetchable BAR. All reads to memory
> > mapped to that BAR returns 0xFFFFFFFF and writes have no effect.
> > 
> > When the PCI bus isn't configured by any firmware (e.g. a PCIe
> > controller solely initialized by Linux kernel), the PCI space allocation
> > algorithm of Linux will allocate the main VRAM BAR of Arc dGPU device at
> > base+0, and then the 8MB BAR at base+256M, which prevents the main VRAM
> > BAR gets resized.

__resource_resize_store() tries to release all resoures with the same 
flags as the resource to be resized. But it seems the release doesn't work 
across devices.

I don't like that flags check anyway, I'd want to replace all such black 
magic with a function that consistently determines the bridge window a 
resouce is assigned to. I've a series to that effect but it doesn't cover 
resize cases yet and it requires more testing anyway to confirm it doesn't 
change any parent windows resources get assigned to.

So IMO, the correct logic on resize would be to:

1) Get the relevant upstream bridge window
2) Release all child resource of that bridge window. But that will 
require further checks whether all those resources (from foreign PCI devs) 
can be released which might run a foul with dev lock ordering.

...So it might turn out hard to implement in practice.

> > As the functionality and performance of Arc dGPU will
> > get severely restricted with small BAR, this makes a problem.
> > 
> > Hide the mysterious 8MB BAR to Linux PCI subsystem, to allow resizing
> > the VRAM BAR to VRAM size with the Linux PCI space allocation algorithm.
> 
> There's no reason a switch upstream port should not have a BAR.  I
> suspect this BAR probably does have a legitimate purpose, and it's
> only "mysterious" because we don't know how to use it.
> 
> This sounds like it may be a deficiency in the Linux BAR assignment
> code.  Any other device could have a similar problem.  

I'm still working also with the resource fitting logic to make it consider 
resizable BARs when sizing the resource which would address this problem 
another way.

> > Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
> > ---
> >  drivers/pci/quirks.c | 16 ++++++++++++++++
> >  1 file changed, 16 insertions(+)
> > 
> > diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> > index d7f4ee634263..df304bfec6e9 100644
> > --- a/drivers/pci/quirks.c
> > +++ b/drivers/pci/quirks.c
> > @@ -3650,6 +3650,22 @@ DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d0, quirk_broken_intx_masking);
> >  DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d1, quirk_broken_intx_masking);
> >  DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_INTEL, 0x37d2, quirk_broken_intx_masking);
> >  
> > +/*
> > + * Intel Arc dGPUs' internal switch upstream port contains a mysterious 8MB
> > + * 64-bit prefetchable BAR that blocks resize of main dGPU VRAM BAR with
> > + * Linux's PCI space allocation algorithm.
> > + */
> > +static void quirk_intel_xe_upstream(struct pci_dev *pdev)
> > +{
> > +	memset(&pdev->resource[0], 0, sizeof(pdev->resource[0]));
> 
> This doesn't touch the BAR itself, so we may be leaving the BAR
> decoding accesses, which could lead to an address conflict.  It also
> prevents a driver for the upstream port from using the BAR.
> 
> > +}
> > +/* Intel Arc A380 PCI Express Switch Upstream Port */
> > +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa1, quirk_intel_xe_upstream);
> > +/* Intel Arc A770 PCI Express Switch Upstream Port */
> > +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x4fa0, quirk_intel_xe_upstream);
> > +/* Intel Arc B580 PCI Express Switch Upstream Port */
> > +DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0xe2ff, quirk_intel_xe_upstream);
> > +
> >  static u16 mellanox_broken_intx_devs[] = {
> >  	PCI_DEVICE_ID_MELLANOX_HERMON_SDR,
> >  	PCI_DEVICE_ID_MELLANOX_HERMON_DDR,
> > -- 
> > 2.50.1
> > 
> 

-- 
 i.

Patchew 2.1-devel-b67e4c2

© 2016 - 2025 Red Hat, Inc.