[PATCH V2] accel/amdxdna: Set HMM_PFN_REQ_WRITE for writable BO faults

[PATCH V2] accel/amdxdna: Set HMM_PFN_REQ_WRITE for writable BO faults

[Lizhi Hou]

When aie2_populate_range() faults in pages, it only sets
HMM_PFN_REQ_FAULT. Without HMM_PFN_REQ_WRITE, handle_mm_fault() does not
request write permissions and may populate the CPU page tables with
read-only PTEs.

Check the VMA flags and set HMM_PFN_REQ_WRITE when the VMA has VM_WRITE.

Fixes: e486147c912f ("accel/amdxdna: Add BO import and export")
Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
---
V2:
  Fix sashiko comment.
 drivers/accel/amdxdna/amdxdna_gem.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)

diff --git a/drivers/accel/amdxdna/amdxdna_gem.c b/drivers/accel/amdxdna/amdxdna_gem.c
index e861db6f9369..42d5f856a839 100644
--- a/drivers/accel/amdxdna/amdxdna_gem.c
+++ b/drivers/accel/amdxdna/amdxdna_gem.c
@@ -249,6 +249,20 @@ u64 amdxdna_gem_dev_addr(struct amdxdna_gem_obj *abo)
 	return amdxdna_obj_dma_addr(abo);
 }
 
+static void amdxdna_umap_set_fault_flags(struct amdxdna_umap *mapp,
+					 struct vm_area_struct *vma)
+{
+	mapp->range.default_flags = HMM_PFN_REQ_FAULT;
+
+	if (!vma)
+		vma = vma_lookup(mapp->notifier.mm, mapp->range.start);
+
+	if (vma && vma->vm_start <= mapp->range.start &&
+	    vma->vm_end >= mapp->range.end &&
+	    (vma->vm_flags & VM_WRITE))
+		mapp->range.default_flags |= HMM_PFN_REQ_WRITE;
+}
+
 static bool amdxdna_hmm_invalidate(struct mmu_interval_notifier *mni,
 				   const struct mmu_notifier_range *range,
 				   unsigned long cur_seq)
@@ -270,6 +284,7 @@ static bool amdxdna_hmm_invalidate(struct mmu_interval_notifier *mni,
 	down_write(&xdna->notifier_lock);
 	abo->mem.map_invalid = true;
 	mapp->invalid = true;
+	amdxdna_umap_set_fault_flags(mapp, NULL);
 	mmu_interval_set_seq(&mapp->notifier, cur_seq);
 	up_write(&xdna->notifier_lock);
 
@@ -418,7 +433,7 @@ static int amdxdna_hmm_register(struct amdxdna_gem_obj *abo,
 	mapp->range.notifier = &mapp->notifier;
 	mapp->range.start = vma->vm_start;
 	mapp->range.end = vma->vm_end;
-	mapp->range.default_flags = HMM_PFN_REQ_FAULT;
+	amdxdna_umap_set_fault_flags(mapp, vma);
 	mapp->abo = abo;
 	kref_init(&mapp->refcnt);
 
-- 
2.34.1

Re: [PATCH V2] accel/amdxdna: Set HMM_PFN_REQ_WRITE for writable BO faults

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] accel/amdxdna: data race on default_flags and illegal vma_lookup in invalidation
--

commit a61f87b4e6515640e66ee7d41700b39d243ec56e
Author: Lizhi Hou <lizhi.hou@amd.com>

accel/amdxdna: Set HMM_PFN_REQ_WRITE for writable BO faults

This commit ensures HMM_PFN_REQ_WRITE is set when handling faults for VMAs
with VM_WRITE. It introduces amdxdna_umap_set_fault_flags() to check VMA
permissions and apply the flag, calling it from the notifier registration and
invalidation paths.

> diff --git a/drivers/accel/amdxdna/amdxdna_gem.c b/drivers/accel/amdxdna/amdxdna_gem.c
> index 5a1a2ee261b76..4f6fe982583fd 100644
> --- a/drivers/accel/amdxdna/amdxdna_gem.c
> +++ b/drivers/accel/amdxdna/amdxdna_gem.c
[ ... ]
> @@ -249,6 +249,20 @@ u64 amdxdna_gem_dev_addr(struct amdxdna_gem_obj *abo)
>  	return amdxdna_obj_dma_addr(abo);
>  }
>  
> +static void amdxdna_umap_set_fault_flags(struct amdxdna_umap *mapp,
> +					 struct vm_area_struct *vma)
> +{
> +	mapp->range.default_flags = HMM_PFN_REQ_FAULT;

[Severity: High]
Does this code introduce a data race? amdxdna_hmm_invalidate is an MMU
interval notifier invalidation callback, which can run concurrently with the
driver's page fault handler (hmm_range_fault_unlocked_timeout). Modifying
mapp->range.default_flags here without synchronization can race with
hmm_range_fault, which reads range->default_flags locklessly.

> +
> +	if (!vma)
> +		vma = vma_lookup(mapp->notifier.mm, mapp->range.start);

[Severity: High]
Can calling vma_lookup here trigger a lockdep warning or use-after-free?
When called from amdxdna_hmm_invalidate, MMU notifiers (such as those
triggered by page reclaim or migration) do not guarantee that mmap_lock is
held, and this code does not take rcu_read_lock(). Calling vma_lookup
without either lock is unsafe, and may cause a use-after-free when
dereferencing vma->vm_flags on a potentially freed VMA.

> +
> +	if (vma && vma->vm_start <= mapp->range.start &&
> +	    vma->vm_end >= mapp->range.end &&
> +	    (vma->vm_flags & VM_WRITE))
> +		mapp->range.default_flags |= HMM_PFN_REQ_WRITE;
> +}
> +
>  static bool amdxdna_hmm_invalidate(struct mmu_interval_notifier *mni,
>  				   const struct mmu_notifier_range *range,
>  				   unsigned long cur_seq)
> @@ -270,6 +284,7 @@ static bool amdxdna_hmm_invalidate(struct mmu_interval_notifier *mni,
>  	down_write(&xdna->notifier_lock);
>  	abo->mem.map_invalid = true;
>  	mapp->invalid = true;
> +	amdxdna_umap_set_fault_flags(mapp, NULL);
>  	mmu_interval_set_seq(&mapp->notifier, cur_seq);
>  	up_write(&xdna->notifier_lock);

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260917165406.1881899-1-lizhi.hou@amd.com?part=1