iommu/vt-d: Clear Present bit before tearing down scalable-mode context entry
Subsystems (2)
PATCH 1/1 iommu/vt-d: Clear Present bit before tearing down scalable-mode context entry
device_pasid_table_teardown() zeroes the 128-bit scalable-mode context entry with context_clear_entry() while the Present bit is still set. This creates a window where the hardware can fetch a torn entry, with some fields already zeroed while Present is still set, leading to unpredictable behavior or spurious faults. The context-cache invalidation is issued only after the entry has been zeroed, and intel_pasid_free_table() then frees the PASID directory pages, so the IOMMU can keep walking a stale Present=1 entry that points at freed memory.
While x86 provides strong write ordering, the compiler may reorder the two 64-bit writes to the entry, and the hardware fetch is not guaranteed to be atomic with respect to multiple CPU writes.
Commit c1e4f1dccbe9d ("iommu/vt-d: Clear Present bit before tearing down context entry") fixed this exact pattern in domain_context_clear_one() and the copied-context path, but device_pasid_table_teardown() was not converted.
Align it with the "Guidance to Software for Invalidations" in the VT-d spec, Section 6.5.3.3, using the same ownership handshake as the sibling
fix: clear only the Present bit, flush it to the IOMMU, perform the context-cache invalidation, and only then zero the rest of the entry.
---
Found by static analysis while auditing the callers of context_clear_entry() for the same teardown ordering that c1e4f1dccbe9d addressed. This site is reachable only in scalable mode, so it does not manifest on the legacy-mode hardware available to me; I could not trigger a runtime fault and the change is verified by code inspection only, on the same basis as the sibling fix. Compile-tested on x86_64 with CONFIG_INTEL_IOMMU; no new warnings.
drivers/iommu/intel/pasid.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/iommu/intel/pasid.c +3 -1
@@ -748,10 +748,12 @@ static void device_pasid_table_teardown(struct device *dev, u8 bus, u8 devfn) } did =context_domain_id (context); -context_clear_entry (context); +context_clear_present (context);__iommu_flush_cache (iommu, context,sizeof (*context));spin_unlock (&iommu->lock);intel_context_flush_no_pasid (info, context, did); +context_clear_entry (context); +__iommu_flush_cache (iommu, context,sizeof (*context)); }static int pci_pasid_table_teardown (struct pci_dev *pdev,u16 alias,void *data)
base-commit: 5200f5f493f79f14bbdc349e402a40dfb32f23c8
Signature · 1 line
2.53.0
Quoted 34 lines · Michael Bommarito
device_pasid_table_teardown()zeroes the 128-bit scalable-mode context entry withcontext_clear_entry()while the Present bit is still set. This creates a window where the hardware can fetch a torn entry, with some fields already zeroed while Present is still set, leading to unpredictable behavior or spurious faults. The context-cache invalidation is issued only after the entry has been zeroed, andintel_pasid_free_table()then frees the PASID directory pages, so the IOMMU can keep walking a stale Present=1 entry that points at freed memory.While x86 provides strong write ordering, the compiler may reorder the two 64-bit writes to the entry, and the hardware fetch is not guaranteed to be atomic with respect to multiple CPU writes.
Commit
c1e4f1dccbe9d("iommu/vt-d: Clear Present bit before tearing down context entry") fixed this exact pattern indomain_context_clear_one()and the copied-context path, butdevice_pasid_table_teardown()was not converted.Align it with the "Guidance to Software for Invalidations" in the VT-d spec, Section 6.5.3.3, using the same ownership handshake as the sibling
fix: clear only the Present bit, flush it to the IOMMU, perform the context-cache invalidation, and only then zero the rest of the entry.Fixes 81e921fd32161 ("iommu/vt-d: Fix NULL domain on device release")Signed-off-by Michael Bommarito<[email protected]>Assisted-by Claude:claude-opus-4-7---
Found by static analysis while auditing the callers ofcontext_clear_entry()for the same teardown ordering thatc1e4f1dccbe9daddressed. This site is reachable only in scalable mode, so it does not manifest on the legacy-mode hardware available to me; I could not trigger a runtime fault and the change is verified by code inspection only, on the same basis as the sibling fix. Compile-tested onx86_64withCONFIG_INTEL_IOMMU; no new warnings.
Quoted 2 lines
drivers/iommu/intel/pasid.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
Queued for linux-next. Thank you!
Quoted 35 lines · Baolu Lu
Quoted 34 lines · Michael Bommarito
device_pasid_table_teardown()zeroes the 128-bit scalable-mode context entry withcontext_clear_entry()while the Present bit is still set. This creates a window where the hardware can fetch a torn entry, with some fields already zeroed while Present is still set, leading to unpredictable behavior or spurious faults. The context-cache invalidation is issued only after the entry has been zeroed, andintel_pasid_free_table()then frees the PASID directory pages, so the IOMMU can keep walking a stale Present=1 entry that points at freed memory.While x86 provides strong write ordering, the compiler may reorder the two 64-bit writes to the entry, and the hardware fetch is not guaranteed to be atomic with respect to multiple CPU writes.
Commit
c1e4f1dccbe9d("iommu/vt-d: Clear Present bit before tearing down context entry") fixed this exact pattern indomain_context_clear_one()and the copied-context path, butdevice_pasid_table_teardown()was not converted.Align it with the "Guidance to Software for Invalidations" in the VT-d spec, Section 6.5.3.3, using the same ownership handshake as the sibling
fix: clear only the Present bit, flush it to the IOMMU, perform the context-cache invalidation, and only then zero the rest of the entry.Fixes 81e921fd32161 ("iommu/vt-d: Fix NULL domain on device release")Signed-off-by Michael Bommarito<[email protected]>Assisted-by Claude:claude-opus-4-7---
Found by static analysis while auditing the callers ofcontext_clear_entry()for the same teardown ordering thatc1e4f1dccbe9daddressed. This site is reachable only in scalable mode, so it does not manifest on the legacy-mode hardware available to me; I could not trigger a runtime fault and the change is verified by code inspection only, on the same basis as the sibling fix. Compile-tested onx86_64withCONFIG_INTEL_IOMMU; no new warnings.
Quoted 4 lines
Quoted 2 lines
drivers/iommu/intel/pasid.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)Queued for linux-next. Thank you!
What happened to your work to move over to the ARM updator that doesn't have any of these bugs? :)
Jason
Baolu Lu
maintainerQuoted 36 lines · Jason Gunthorpe
Quoted 35 lines · Baolu Lu
Quoted 34 lines · Michael Bommarito
device_pasid_table_teardown()zeroes the 128-bit scalable-mode context entry withcontext_clear_entry()while the Present bit is still set. This creates a window where the hardware can fetch a torn entry, with some fields already zeroed while Present is still set, leading to unpredictable behavior or spurious faults. The context-cache invalidation is issued only after the entry has been zeroed, andintel_pasid_free_table()then frees the PASID directory pages, so the IOMMU can keep walking a stale Present=1 entry that points at freed memory.While x86 provides strong write ordering, the compiler may reorder the two 64-bit writes to the entry, and the hardware fetch is not guaranteed to be atomic with respect to multiple CPU writes.
Commit
c1e4f1dccbe9d("iommu/vt-d: Clear Present bit before tearing down context entry") fixed this exact pattern indomain_context_clear_one()and the copied-context path, butdevice_pasid_table_teardown()was not converted.Align it with the "Guidance to Software for Invalidations" in the VT-d spec, Section 6.5.3.3, using the same ownership handshake as the sibling
fix: clear only the Present bit, flush it to the IOMMU, perform the context-cache invalidation, and only then zero the rest of the entry.Fixes 81e921fd32161 ("iommu/vt-d: Fix NULL domain on device release")Signed-off-by Michael Bommarito<[email protected]>Assisted-by Claude:claude-opus-4-7---
Found by static analysis while auditing the callers ofcontext_clear_entry()for the same teardown ordering thatc1e4f1dccbe9daddressed. This site is reachable only in scalable mode, so it does not manifest on the legacy-mode hardware available to me; I could not trigger a runtime fault and the change is verified by code inspection only, on the same basis as the sibling fix. Compile-tested onx86_64withCONFIG_INTEL_IOMMU; no new warnings.
Quoted 5 lines
Quoted 3 lines
Quoted 2 lines
drivers/iommu/intel/pasid.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)Queued for linux-next. Thank you!
What happened to your work to move over to the ARM updator that doesn't have any of these bugs? 🙂
I am working on that series, but since this is a fix that should be backported, I queued it for this merge cycle.
Thanks,
baolu
Baolu Lu
maintainerQuoted 34 lines · Michael Bommarito
device_pasid_table_teardown()zeroes the 128-bit scalable-mode context entry withcontext_clear_entry()while the Present bit is still set. This creates a window where the hardware can fetch a torn entry, with some fields already zeroed while Present is still set, leading to unpredictable behavior or spurious faults. The context-cache invalidation is issued only after the entry has been zeroed, andintel_pasid_free_table()then frees the PASID directory pages, so the IOMMU can keep walking a stale Present=1 entry that points at freed memory.While x86 provides strong write ordering, the compiler may reorder the two 64-bit writes to the entry, and the hardware fetch is not guaranteed to be atomic with respect to multiple CPU writes.
Commit
c1e4f1dccbe9d("iommu/vt-d: Clear Present bit before tearing down context entry") fixed this exact pattern indomain_context_clear_one()and the copied-context path, butdevice_pasid_table_teardown()was not converted.Align it with the "Guidance to Software for Invalidations" in the VT-d spec, Section 6.5.3.3, using the same ownership handshake as the sibling
fix: clear only the Present bit, flush it to the IOMMU, perform the context-cache invalidation, and only then zero the rest of the entry.Fixes 81e921fd32161 ("iommu/vt-d: Fix NULL domain on device release")Signed-off-by Michael Bommarito<[email protected]>Assisted-by Claude:claude-opus-4-7---
Found by static analysis while auditing the callers ofcontext_clear_entry()for the same teardown ordering thatc1e4f1dccbe9daddressed. This site is reachable only in scalable mode, so it does not manifest on the legacy-mode hardware available to me; I could not trigger a runtime fault and the change is verified by code inspection only, on the same basis as the sibling fix. Compile-tested onx86_64withCONFIG_INTEL_IOMMU; no new warnings.
Quoted 2 lines
drivers/iommu/intel/pasid.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
Sashiko reported a pre-existing issue:
https://sashiko.dev/#/patchset/20260528025557.3209367-1-michael.bommarito%40gmail.com
I have added it to my task list and will follow up with a fix in a separate patch.
[Severity: Critical]
This is a pre-existing issue, but does this context cache flush actually invalidate the alias BDF?
It looks like intel_context_flush_no_pasid() issues a device-selective invalidation that is hardcoded to use the primary device's BDF from info (PCI_DEVID(info->bus, info->devfn)), completely ignoring the alias's bus and devfn parameters passed into device_pasid_table_teardown().
If device_pasid_table_teardown() is called for a DMA alias, the IOMMU context cache for the alias BDF might never be invalidated.
When the domain's page tables or scalable-mode PASID directories are subsequently freed, could any DMA arriving under the alias BDF cause the IOMMU hardware to walk the stale cache entry into freed system memory?
Thanks,
baolu