Skip to contentSkip to conversation

vfio/pci: Don't export DMABUFs for unmappable BARs

PATCH In review Matt Evans 3 participants 11 messages Apr 15 – Apr 21

Subsystems (2)
lore
Share & tools
lore

Matt Evans

author
#0 lore raw

PATCH 1/1 vfio/pci: Don't export DMABUFs for unmappable BARs

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
 drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
index f87fd32e4a01..4ccaf3531e02 100644
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
 	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
 		return -ENODEV;
 
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;
+
 	dma_ranges = memdup_array_user(&arg->dma_ranges, get_dma_buf.nr_ranges,
 				       sizeof(*dma_ranges));
 	if (IS_ERR(dma_ranges))
Signature · 1 line
2.47.3
#1 lore raw
Quoted 19 lines · Matt Evans

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
 drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
index f87fd32e4a01..4ccaf3531e02 100644

drivers/vfio/pci/vfio_pci_dmabuf.c int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags, · line 253

Quoted 5 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +2 -0
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
 	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
 		return -ENODEV;
 
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;

I noticed this check in vfio_pci_core_mmap(). Isn't that sufficient?

Thanks

Quoted 7 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +1 -0
+
 	dma_ranges = memdup_array_user(&arg->dma_ranges, get_dma_buf.nr_ranges,
 				       sizeof(*dma_ranges));
 	if (IS_ERR(dma_ranges))
Signature · 1 line
2.47.3
#2 lore raw
In reply to Matt Evans
Quoted 19 lines · Matt Evans

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
 drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
index f87fd32e4a01..4ccaf3531e02 100644

drivers/vfio/pci/vfio_pci_dmabuf.c int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags, · line 254

Quoted 6 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
 	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
 		return -ENODEV;
 
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;
+

And it looks like AI has valid concern about this line too.
https://sashiko.dev/#/patchset/[email protected]

Thanks

Quoted 6 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
 	dma_ranges = memdup_array_user(&arg->dma_ranges, get_dma_buf.nr_ranges,
 				       sizeof(*dma_ranges));
 	if (IS_ERR(dma_ranges))
Signature · 1 line
2.47.3
author
#3 lore raw

Hi Leon,

Quoted 24 lines · Leon Romanovsky
Quoted 1 line
Quoted 22 lines · Matt Evans

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
  drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
  1 file changed, 3 insertions(+)
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
index f87fd32e4a01..4ccaf3531e02 100644
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
  	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
  		return -ENODEV;
 
Quoted 6 lines

drivers/vfio/pci/vfio_pci_dmabuf.c int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags, · line 254

Quoted 3 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;
+

And it looks like AI has valid concern about this line too.
https://urldefense.com/v3/__https://sashiko.dev/*/patchset/[email protected]__;Iw!!Bt8RZUm9aw!5DxsN8cDUviPIZqEjG0pZ_VYYbl_RdmWucTGdTZ3ZzlVP_Ysb0n7ykr0eXwFXdpuqvZH2FK3$

Ah, Sashiko has a point, and I think its suggestion of checking lower down in the default .get_dmabuf_phys (vfio_pci_core_get_dmabuf_phys()) and preserving driver overrides is decent. Will revisit.

To your other question:

Quoted 1 line

I noticed this check in vfio_pci_core_mmap(). Isn't that sufficient?

The scenario in mind is doing a DMABUF-export for BARs that you haven't necessarily noticed can't be mmap()ed, and both paths should be checking.

Cheers,

Matt

Quoted 9 lines

Thanks

Quoted 6 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
  	dma_ranges = memdup_array_user(&arg->dma_ranges, get_dma_buf.nr_ranges,
  				       sizeof(*dma_ranges));
  	if (IS_ERR(dma_ranges))
Signature · 1 line
2.47.3
#4 lore raw
Quoted 36 lines · Matt Evans

Hi Leon,

Quoted 15 lines · Leon Romanovsky
Quoted 15 lines · Matt Evans

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
  drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
  1 file changed, 3 insertions(+)
Quoted 13 lines

drivers/vfio/pci/vfio_pci_dmabuf.c int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags, · line 253

Quoted 10 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
index f87fd32e4a01..4ccaf3531e02 100644
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
  	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
  		return -ENODEV;
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;
+

And it looks like AI has valid concern about this line too.
https://urldefense.com/v3/__https://sashiko.dev/*/patchset/[email protected]__;Iw!!Bt8RZUm9aw!5DxsN8cDUviPIZqEjG0pZ_VYYbl_RdmWucTGdTZ3ZzlVP_Ysb0n7ykr0eXwFXdpuqvZH2FK3$

Ah, Sashiko has a point, and I think its suggestion of checking lower down in the default .get_dmabuf_phys (vfio_pci_core_get_dmabuf_phys()) and preserving driver overrides is decent. Will revisit.

Quoted 5 lines

To your other question:

Quoted 1 line

I noticed this check in vfio_pci_core_mmap(). Isn't that sufficient?

The scenario in mind is doing a DMABUF-export for BARs that you haven't necessarily noticed can't be mmap()ed, and both paths should be checking.

I added the validation checks that matter on the kernel side, but mmap is primarily important for callers. What I am missing is an explanation of why the kernel should impose this restriction on itself.

Thanks

Quoted 17 lines

Cheers,

Matt

Quoted 9 lines

Thanks

Quoted 6 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
  	dma_ranges = memdup_array_user(&arg->dma_ranges, get_dma_buf.nr_ranges,
  				       sizeof(*dma_ranges));
  	if (IS_ERR(dma_ranges))
Signature · 1 line
2.47.3
author
#5 lore raw

Hi Leon,

Quoted 39 lines · Leon Romanovsky
Quoted 37 lines · Matt Evans

Hi Leon,

Quoted 15 lines · Leon Romanovsky
Quoted 15 lines · Matt Evans

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
   drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
   1 file changed, 3 insertions(+)
Quoted 13 lines

drivers/vfio/pci/vfio_pci_dmabuf.c int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags, · line 253

Quoted 10 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
index f87fd32e4a01..4ccaf3531e02 100644
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
   	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
   		return -ENODEV;
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;
+

And it looks like AI has valid concern about this line too.
https://urldefense.com/v3/__https://sashiko.dev/*/patchset/[email protected]__;Iw!!Bt8RZUm9aw!5DxsN8cDUviPIZqEjG0pZ_VYYbl_RdmWucTGdTZ3ZzlVP_Ysb0n7ykr0eXwFXdpuqvZH2FK3$

Ah, Sashiko has a point, and I think its suggestion of checking lower down in the default .get_dmabuf_phys (vfio_pci_core_get_dmabuf_phys()) and preserving driver overrides is decent. Will revisit.

To your other question:

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

I noticed this check in vfio_pci_core_mmap(). Isn't that sufficient?

The scenario in mind is doing a DMABUF-export for BARs that you haven't necessarily noticed can't be mmap()ed, and both paths should be checking.

I added the validation checks that matter on the kernel side, but mmap is primarily important for callers. What I am missing is an explanation of why the kernel should impose this restriction on itself.

I don't understand your question, really sorry! Can you rephrase it please? I want to make sure I answer it fully.

Although mmap() fails for BARs that are unmappable (for whatever reason), a DMABUF export for the same ones could in some slim cases succeed -- because the checks aren't identical. If export succeeds, it could potentially allow P2P (or CPU via a future DMABUF mmap()) access to something possibly unmappable, no?

For the checks that vfio_pci_probe_mmaps() does (leading to bar_mmap_supported[] = false), most have corresponding-but-different checks reachable from DMABUF export:

If a BAR is: Then DMABUF export...:

size < pagesize vfio_pci_core_fill_phys_vec() catches it
Not IORESOURCE_MEM pcim_p2pdma_provider() rejects it
non_mappable_bars ... nothing? Export allowed

As a quick test, if I hack in non_mappable_bars=1 for my function, it appears exporting a DMABUF from it works.

We could add another check for non_mappable_bars, but my thinking was that we don't want to keep adding to an independent set of DMABUF checks, especially if a future quirk/etc. could create another scenario where BARs aren't mappable. I.e. we should reject DMABUF export in exactly the same scenarios as mmap() would be rejected, symmetrically, by testing bar_mmap_supported[].

Hope that goes some way to answering the Q, hopefully I haven't missed something!

Thanks,

Matt

Quoted 20 lines

Thanks

Quoted 17 lines

Cheers,

Matt

Quoted 9 lines

Thanks

Quoted 6 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
   	dma_ranges = memdup_array_user(&arg->dma_ranges, get_dma_buf.nr_ranges,
   				       sizeof(*dma_ranges));
   	if (IS_ERR(dma_ranges))
Signature · 1 line
2.47.3
maintainer
#6 lore raw
Quoted 80 lines · On Thu, 16 Apr 2026 19:03:40 +0100 Matt Evans <[email protected]> wrote:

Hi Leon,

Quoted 39 lines · Leon Romanovsky
Quoted 37 lines · Matt Evans

Hi Leon,

Quoted 27 lines · Leon Romanovsky
Quoted 20 lines · Matt Evans

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
   drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
   1 file changed, 3 insertions(+)
drivers/vfio/pci/vfio_pci_dmabuf.c +0 -0
index f87fd32e4a01..4ccaf3531e02 100644

drivers/vfio/pci/vfio_pci_dmabuf.c int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags, · line 253

Quoted 5 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
   	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
   		return -ENODEV;
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;
+  

And it looks like AI has valid concern about this line too.

Quoted 1 line
https://urldefense.com/v3/__https://sashiko.dev/*/patchset/[email protected]__;Iw!!Bt8RZUm9aw!5DxsN8cDUviPIZqEjG0pZ_VYYbl_RdmWucTGdTZ3ZzlVP_Ysb0n7ykr0eXwFXdpuqvZH2FK3$  

Ah, Sashiko has a point, and I think its suggestion of checking lower down in the default .get_dmabuf_phys (vfio_pci_core_get_dmabuf_phys()) and preserving driver overrides is decent. Will revisit.

To your other question:

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

I noticed this check in vfio_pci_core_mmap(). Isn't that sufficient?

The scenario in mind is doing a DMABUF-export for BARs that you haven't necessarily noticed can't be mmap()ed, and both paths should be checking.

I added the validation checks that matter on the kernel side, but mmap is primarily important for callers. What I am missing is an explanation of why the kernel should impose this restriction on itself.

I don't understand your question, really sorry! Can you rephrase it please? I want to make sure I answer it fully.

Although mmap() fails for BARs that are unmappable (for whatever reason), a DMABUF export for the same ones could in some slim cases succeed -- because the checks aren't identical. If export succeeds, it could potentially allow P2P (or CPU via a future DMABUF mmap()) access to something possibly unmappable, no?

For the checks that vfio_pci_probe_mmaps() does (leading to bar_mmap_supported[] = false), most have corresponding-but-different checks reachable from DMABUF export:

If a BAR is: Then DMABUF export...:

size < pagesize vfio_pci_core_fill_phys_vec() catches it
Not IORESOURCE_MEM pcim_p2pdma_provider() rejects it
non_mappable_bars ... nothing? Export allowed

As a quick test, if I hack in non_mappable_bars=1 for my function, it appears exporting a DMABUF from it works.

We could add another check for non_mappable_bars, but my thinking was that we don't want to keep adding to an independent set of DMABUF checks, especially if a future quirk/etc. could create another scenario where BARs aren't mappable. I.e. we should reject DMABUF export in exactly the same scenarios as mmap() would be rejected, symmetrically, by testing bar_mmap_supported[].

Quoted 2 lines

Hope that goes some way to answering the Q, hopefully I haven't missed something!

That's the concern as I see it as well, it's a choice whether to attempt to support sub-PAGE_SIZE mappings, but if a device is reporting non_mappable_bars are we're exporting those BARs through dma-buf for mmap, that's a problem. Should pcim_p2pdma_provider() test this flag rather than vfio_pci_dmabuf though? Thanks,

Alex

author
#7 lore raw

Hi Alex,

Quoted 84 lines · Alex Williamson
Quoted 80 lines · On Thu, 16 Apr 2026 19:03:40 +0100 Matt Evans <[email protected]> wrote:

Hi Leon,

Quoted 39 lines · Leon Romanovsky
Quoted 37 lines · Matt Evans

Hi Leon,

Quoted 15 lines · Leon Romanovsky
Quoted 15 lines · Matt Evans

Although vfio_pci_core_feature_dma_buf() validates that both requested DMABUF ranges and the PCI resources being referenced are page-aligned, there may be reasons other than alignment that cause a BAR to be unmappable.

Add a check for vdev->bar_mmap_supported[index], similar to the VFIO mmap path.

Fixes 5d74781ebc86c ("vfio/pci: Add dma-buf export support for MMIO regions")
---
    drivers/vfio/pci/vfio_pci_dmabuf.c | 3 +++
    1 file changed, 3 insertions(+)
Quoted 13 lines

drivers/vfio/pci/vfio_pci_dmabuf.c int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags, · line 253

Quoted 10 lines
drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
index f87fd32e4a01..4ccaf3531e02 100644
@@ -249,6 +249,9 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
    	if (get_dma_buf.region_index >= VFIO_PCI_ROM_REGION_INDEX)
    		return -ENODEV;
+	if (!vdev->bar_mmap_supported[get_dma_buf.region_index])
+		return -EINVAL;
+

And it looks like AI has valid concern about this line too.
https://urldefense.com/v3/__https://sashiko.dev/*/patchset/[email protected]__;Iw!!Bt8RZUm9aw!5DxsN8cDUviPIZqEjG0pZ_VYYbl_RdmWucTGdTZ3ZzlVP_Ysb0n7ykr0eXwFXdpuqvZH2FK3$

Ah, Sashiko has a point, and I think its suggestion of checking lower down in the default .get_dmabuf_phys (vfio_pci_core_get_dmabuf_phys()) and preserving driver overrides is decent. Will revisit.

To your other question:

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

I noticed this check in vfio_pci_core_mmap(). Isn't that sufficient?

The scenario in mind is doing a DMABUF-export for BARs that you haven't necessarily noticed can't be mmap()ed, and both paths should be checking.

I added the validation checks that matter on the kernel side, but mmap is primarily important for callers. What I am missing is an explanation of why the kernel should impose this restriction on itself.

I don't understand your question, really sorry! Can you rephrase it please? I want to make sure I answer it fully.

Although mmap() fails for BARs that are unmappable (for whatever reason), a DMABUF export for the same ones could in some slim cases succeed -- because the checks aren't identical. If export succeeds, it could potentially allow P2P (or CPU via a future DMABUF mmap()) access to something possibly unmappable, no?

For the checks that vfio_pci_probe_mmaps() does (leading to bar_mmap_supported[] = false), most have corresponding-but-different checks reachable from DMABUF export:

If a BAR is: Then DMABUF export...:

size < pagesize vfio_pci_core_fill_phys_vec() catches it
Not IORESOURCE_MEM pcim_p2pdma_provider() rejects it
non_mappable_bars ... nothing? Export allowed

As a quick test, if I hack in non_mappable_bars=1 for my function, it appears exporting a DMABUF from it works.

We could add another check for non_mappable_bars, but my thinking was that we don't want to keep adding to an independent set of DMABUF checks, especially if a future quirk/etc. could create another scenario where BARs aren't mappable. I.e. we should reject DMABUF export in exactly the same scenarios as mmap() would be rejected, symmetrically, by testing bar_mmap_supported[].

Quoted 7 lines
Quoted 2 lines

Hope that goes some way to answering the Q, hopefully I haven't missed something!

That's the concern as I see it as well, it's a choice whether to
attempt to support sub-PAGE_SIZE mappings, but if a device is reporting
non_mappable_bars are we're exporting those BARs through dma-buf for
mmap, that's a problem. Should pcim_p2pdma_provider() test this flag> rather than vfio_pci_dmabuf though? Thanks,

Do you mean "test this flag" to say 'non_mappable_bars' rather than the 'vdev->bar_mmap_supported[]' flag? (I think the former, as the latter isn't as easily available down there, sorry if that's not what you intended.)

Testing non_mappable_bars in pcim_p2pdma_provider() doesn't feel quite right:

- IIUC non_mappable_bars is about preventing mapping to userspace, not direct access/P2P/etc. Splitting hairs a bit, but P2P to such a device seems valid, so I think this check better stays within vfio-pci (specifically limiting only userspace access).

- But non_mappable_bars is just one kind of quirk, and couldn't there possibly be more? Good to avoid duplicating quirk checks in mmap() & export places (so any future maintenance updates just one place and no disparities arise).

All this to say: The existng logic in vfio_pci_probe_mmaps() is the right place to decide something is suitable for userspace/direct mapping (mappable, non-zero sized, not IORESOURCE_IO), IMHO we just need to be checking it consistently. (VFIO export is less lenient in terms of >= pagesize and imposes additional checks, and that's good as long as the checks are an overlay rather than parallel duplication.)

(Maybe you're also flagging that there could be other checks saying "Is P2P supported for this BAR?" and they could be done in a generic drivers/pci place? I think so; VFIO export criteria be the intersection of VFIO- and provider-based checks.)

I was thinking something like...

drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
index 00cedfe3a57d..9bb8bd153e12 100644
@@ -359,6 +359,9 @@ int vfio_pci_core_get_dmabuf_phys(struct vfio_pci_core_device *vdev,
        if (!*provider)
                return -EINVAL;
 
+       if (!vdev->bar_mmap_supported[region_index])
+               return -EINVAL;
+
        return vfio_pci_core_fill_phys_vec(
                phys_vec, dma_ranges, nr_ranges,
                pci_resource_start(pdev, region_index),

(I.e. leverage logic in vfio_pci_probe_mmaps(), catch bad DMABUF mappings this way, allows sub-drivers to override .get_dmabuf_phys.)

I'll repost that as v2 if this doesn't seem an outrageous start. :)

Matt

maintainer
#8 lore raw
Quoted 42 lines · On Fri, 17 Apr 2026 15:25:07 +0100 Matt Evans <[email protected]> wrote:
Quoted 34 lines · Alex Williamson
Quoted 32 lines · On Thu, 16 Apr 2026 19:03:40 +0100 Matt Evans <[email protected]> wrote:
Quoted 1 line · Leon Romanovsky

On Thu, Apr 16, 2026 at 02:05:30PM +0100, Matt Evans wrote:

I don't understand your question, really sorry! Can you rephrase it please? I want to make sure I answer it fully.

Although mmap() fails for BARs that are unmappable (for whatever reason), a DMABUF export for the same ones could in some slim cases succeed -- because the checks aren't identical. If export succeeds, it could potentially allow P2P (or CPU via a future DMABUF mmap()) access to something possibly unmappable, no?

For the checks that vfio_pci_probe_mmaps() does (leading to bar_mmap_supported[] = false), most have corresponding-but-different checks reachable from DMABUF export:

If a BAR is: Then DMABUF export...:

size < pagesize vfio_pci_core_fill_phys_vec() catches it
Not IORESOURCE_MEM pcim_p2pdma_provider() rejects it
non_mappable_bars ... nothing? Export allowed

As a quick test, if I hack in non_mappable_bars=1 for my function, it appears exporting a DMABUF from it works.

We could add another check for non_mappable_bars, but my thinking was that we don't want to keep adding to an independent set of DMABUF checks, especially if a future quirk/etc. could create another scenario where BARs aren't mappable. I.e. we should reject DMABUF export in exactly the same scenarios as mmap() would be rejected, symmetrically, by testing bar_mmap_supported[].

Quoted 7 lines
Quoted 2 lines

Hope that goes some way to answering the Q, hopefully I haven't missed something!

That's the concern as I see it as well, it's a choice whether to attempt to support sub-PAGE_SIZE mappings, but if a device is reporting non_mappable_bars are we're exporting those BARs through dma-buf for mmap, that's a problem. Should pcim_p2pdma_provider() test this flag

Quoted 6 lines

rather than vfio_pci_dmabuf though? Thanks,

Do you mean "test this flag" to say 'non_mappable_bars' rather than the 'vdev->bar_mmap_supported[]' flag? (I think the former, as the latter isn't as easily available down there, sorry if that's not what you intended.)

Yes

Quoted 7 lines

Testing non_mappable_bars in pcim_p2pdma_provider() doesn't feel quite right:

- IIUC non_mappable_bars is about preventing mapping to userspace, not direct access/P2P/etc. Splitting hairs a bit, but P2P to such a device seems valid, so I think this check better stays within vfio-pci (specifically limiting only userspace access).

Indeed the flag does specify userspace mapping in the comment, but I think it's more than that. The only device that currently uses it is a device unique to s390x, where AIUI, there is no P2P DMA anyway.

Also, for "legacy" mapping of device BARs through the vfio type1 backend, the mechanics of the mapping require the BAR can be mmap'd such that the user virtual address can then be used for the DMA mapping. So as far as vfio has traditionally been concerned, there's an inherent dependency.

I could definitely see that p2p folks could balk and we need a separate flag... or since there's exactly one device that reports this flag, maybe we can tweak the description.

Quoted 4 lines

- But non_mappable_bars is just one kind of quirk, and couldn't there possibly be more? Good to avoid duplicating quirk checks in mmap() & export places (so any future maintenance updates just one place and no disparities arise).

I'm not sure what this is getting at, I think non_mappable_bars is meant to be the flag that quirks might set if for any reason we can't directly map the BAR. I think "userspace" mapping is a bit of a red-herring, it's at least not mappable by the CPU, but I don't think the flag is actually meant to define a policy only for userspace. If it's not mappable by the CPU... is that also enough of a restriction to exclude it from P2P mappings?

Quoted 7 lines

All this to say: The existng logic in vfio_pci_probe_mmaps() is the
right place to decide something is suitable for userspace/direct mapping (mappable, non-zero sized, not IORESOURCE_IO), IMHO we just need to be checking it consistently. (VFIO export is less lenient in terms of >= pagesize and imposes additional checks, and that's good as long as the checks are an overlay rather than parallel duplication.)

I'd actually take the opposite stance and say that vfio is not the only consumer of pcim_p2pdma_provider() and if a device has a flag that it shouldn't be mapped, the p2p subsystem should also honor that, not just vfio.

Quoted 4 lines

(Maybe you're also flagging that there could be other checks saying "Is P2P supported for this BAR?" and they could be done in a generic drivers/pci place? I think so; VFIO export criteria be the intersection of VFIO- and provider-based checks.)

Yes (I think). If we determine that non_mappable_bars means both CPU and device mappings (which is compatible with its current use case), then fixing pcim_p2pdma_provider() to honor the flag fixes all consumers.

Quoted 19 lines

I was thinking something like...

drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
index 00cedfe3a57d..9bb8bd153e12 100644
@@ -359,6 +359,9 @@ int vfio_pci_core_get_dmabuf_phys(struct 
 vfio_pci_core_device *vdev,
         if (!*provider)
                 return -EINVAL;
 
+       if (!vdev->bar_mmap_supported[region_index])
+               return -EINVAL;
+
         return vfio_pci_core_fill_phys_vec(
                 phys_vec, dma_ranges, nr_ranges,

pci_resource_start(pdev, region_index),

Quoted 4 lines

(I.e. leverage logic in vfio_pci_probe_mmaps(), catch bad DMABUF mappings this way, allows sub-drivers to override .get_dmabuf_phys.)

I'll repost that as v2 if this doesn't seem an outrageous start. :)

That's fixing the leaf driver rather than the subsystem, where pci/p2pdma really ought to honor its own flag indicating the BAR is not mappable. The precedent is already there in rejecting IO BARs. Thanks,

Alex

author
#9 lore raw

Hi Alex,

Quoted 71 lines · Alex Williamson
Quoted 42 lines · On Fri, 17 Apr 2026 15:25:07 +0100 Matt Evans <[email protected]> wrote:
Quoted 34 lines · Alex Williamson
Quoted 32 lines · On Thu, 16 Apr 2026 19:03:40 +0100 Matt Evans <[email protected]> wrote:
Quoted 1 line · Leon Romanovsky

On Thu, Apr 16, 2026 at 02:05:30PM +0100, Matt Evans wrote:

I don't understand your question, really sorry! Can you rephrase it please? I want to make sure I answer it fully.

Although mmap() fails for BARs that are unmappable (for whatever reason), a DMABUF export for the same ones could in some slim cases succeed -- because the checks aren't identical. If export succeeds, it could potentially allow P2P (or CPU via a future DMABUF mmap()) access to something possibly unmappable, no?

For the checks that vfio_pci_probe_mmaps() does (leading to bar_mmap_supported[] = false), most have corresponding-but-different checks reachable from DMABUF export:

If a BAR is: Then DMABUF export...:

size < pagesize vfio_pci_core_fill_phys_vec() catches it
Not IORESOURCE_MEM pcim_p2pdma_provider() rejects it
non_mappable_bars ... nothing? Export allowed

As a quick test, if I hack in non_mappable_bars=1 for my function, it appears exporting a DMABUF from it works.

We could add another check for non_mappable_bars, but my thinking was that we don't want to keep adding to an independent set of DMABUF checks, especially if a future quirk/etc. could create another scenario where BARs aren't mappable. I.e. we should reject DMABUF export in exactly the same scenarios as mmap() would be rejected, symmetrically, by testing bar_mmap_supported[].

Quoted 7 lines
Quoted 2 lines

Hope that goes some way to answering the Q, hopefully I haven't missed something!

That's the concern as I see it as well, it's a choice whether to attempt to support sub-PAGE_SIZE mappings, but if a device is reporting non_mappable_bars are we're exporting those BARs through dma-buf for mmap, that's a problem. Should pcim_p2pdma_provider() test this flag

Quoted 6 lines

rather than vfio_pci_dmabuf though? Thanks,

Do you mean "test this flag" to say 'non_mappable_bars' rather than the 'vdev->bar_mmap_supported[]' flag? (I think the former, as the latter isn't as easily available down there, sorry if that's not what you intended.)

Yes

Quoted 7 lines

Testing non_mappable_bars in pcim_p2pdma_provider() doesn't feel quite right:

- IIUC non_mappable_bars is about preventing mapping to userspace, not direct access/P2P/etc. Splitting hairs a bit, but P2P to such a device seems valid, so I think this check better stays within vfio-pci (specifically limiting only userspace access).

Indeed the flag does specify userspace mapping in the comment, but I think it's more than that. The only device that currently uses it is a device unique to s390x, where AIUI, there is no P2P DMA anyway.

Also, for "legacy" mapping of device BARs through the vfio type1 backend, the mechanics of the mapping require the BAR can be mmap'd such that the user virtual address can then be used for the DMA mapping. So as far as vfio has traditionally been concerned, there's an inherent dependency.

Quoted 3 lines

I could definitely see that p2p folks could balk and we need a separate flag... or since there's exactly one device that reports this flag, maybe we can tweak the description.

Thanks for explaining; with that context, ...

Quoted 5 lines
Quoted 4 lines

- But non_mappable_bars is just one kind of quirk, and couldn't there possibly be more? Good to avoid duplicating quirk checks in mmap() & export places (so any future maintenance updates just one place and no disparities arise).

Quoted 7 lines

I'm not sure what this is getting at, I think non_mappable_bars is meant to be the flag that quirks might set if for any reason we can't directly map the BAR. I think "userspace" mapping is a bit of a red-herring, it's at least not mappable by the CPU, but I don't think the flag is actually meant to define a policy only for userspace. If it's not mappable by the CPU... is that also enough of a restriction to exclude it from P2P mappings?

...and this, it makes sense then to just check non_mappable_bars, if it's not about userspace and has the semantic of "don't access this directly".

It'd be nice if it evolved to a per-BAR flag rather than per-function, BUT I acknowledge I was largely worrying about the shape of hypothetical future quirks and that isn't necessary just now.

Quoted 8 lines
Quoted 7 lines

All this to say: The existng logic in vfio_pci_probe_mmaps() is the
right place to decide something is suitable for userspace/direct mapping (mappable, non-zero sized, not IORESOURCE_IO), IMHO we just need to be checking it consistently. (VFIO export is less lenient in terms of >= pagesize and imposes additional checks, and that's good as long as the checks are an overlay rather than parallel duplication.)

Quoted 9 lines

I'd actually take the opposite stance and say that vfio is not the only consumer of pcim_p2pdma_provider() and if a device has a flag that it shouldn't be mapped, the p2p subsystem should also honor that, not just vfio.

Quoted 4 lines

(Maybe you're also flagging that there could be other checks saying "Is P2P supported for this BAR?" and they could be done in a generic drivers/pci place? I think so; VFIO export criteria be the intersection of VFIO- and provider-based checks.)

(Typo, inadvertent talk like a pirate.)

Quoted 4 lines

Yes (I think). If we determine that non_mappable_bars means both CPU and device mappings (which is compatible with its current use case), then fixing pcim_p2pdma_provider() to honor the flag fixes all consumers.

OK, makes sense. Seems OK to couple CPU mappings and P2P; I can imagine quirks (particularly for RCiEPs) where one works but the other doesn't but, until we've a concrete example, affecting both is safest.

For now, I'll do as you first suggested ( :) ), and if future more-elaborate quirks arise we can revisit.

Quoted 19 lines
Quoted 19 lines

I was thinking something like...

drivers/vfio/pci/vfio_pci_dmabuf.c +3 -0
index 00cedfe3a57d..9bb8bd153e12 100644
@@ -359,6 +359,9 @@ int vfio_pci_core_get_dmabuf_phys(struct
 vfio_pci_core_device *vdev,
          if (!*provider)
                  return -EINVAL;
 
+       if (!vdev->bar_mmap_supported[region_index])
+               return -EINVAL;
+
          return vfio_pci_core_fill_phys_vec(
                  phys_vec, dma_ranges, nr_ranges,

pci_resource_start(pdev, region_index),

Quoted 8 lines
Quoted 4 lines

(I.e. leverage logic in vfio_pci_probe_mmaps(), catch bad DMABUF mappings this way, allows sub-drivers to override .get_dmabuf_phys.)

I'll repost that as v2 if this doesn't seem an outrageous start. :)

That's fixing the leaf driver rather than the subsystem, where pci/p2pdma really ought to honor its own flag indicating the BAR is not mappable. The precedent is already there in rejecting IO BARs. Thanks,

All good; I appreciate the discussion, thanks, and will redo with checking non_mappable_bars in p2pdma.

Matt

author
#10 lore raw

Hi,

Quoted 10 lines · Matt Evans

Hi Alex,

Quoted 5 lines · Alex Williamson

[snip]

That's fixing the leaf driver rather than the subsystem, where pci/p2pdma really ought to honor its own flag indicating the BAR is not mappable.  The precedent is already there in rejecting IO BARs.  Thanks,

All good; I appreciate the discussion, thanks, and will redo with checking non_mappable_bars in p2pdma.

Since the patch has a radically different subject/lists, here's a courtesy link:

https://lore.kernel.org/linux-pci/[email protected]/

(Obviously we can drop this thread now.)

Matt