Re: [PATCH V2 2/3] accel/amdxdna: Fix amdxdna_client accessing race in amdxdna_drm_gem_mmap()

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Caching the `vm_area_struct` pointer in the MMU interval notifier leads to Use-After-Free on `munmap()`.
- [High] The patch incompletely fixes a Use-After-Free of `amdxdna_client` during device removal. The UAF remains in the standard GEM mmap path, IOCTLs, and GEM destruction callbacks.
- [High] Double put of DRM GEM object reference on the error path in `amdxdna_insert_pages`.
- [High] Premature destruction of `notifier_wq` leads to Use-After-Free when unmapping VMAs after device removal.
--

commit 4a66abc4067bb1f72f27f8b25d0d1de298787835
Author: Lizhi Hou <lizhi.hou@amd.com>

accel/amdxdna: Fix amdxdna_client accessing race in amdxdna_drm_gem_mmap()

amdxdna_drm_gem_mmap() can be called after drm_dev_unplug() during device
removal. In that case, the mmap path may access amdxdna_client after it
has been freed, leading to a use-after-free.

Guard the mmap path with drm_dev_enter()/drm_dev_exit() and return an
error if the device is already unplugged.

> diff --git a/drivers/accel/amdxdna/amdxdna_pci_drv.c b/drivers/accel/amdxdna/amdxdna_pci_drv.c
> index a22c986f1b331..c18ddad9f6459 100644
> --- a/drivers/accel/amdxdna/amdxdna_pci_drv.c
> +++ b/drivers/accel/amdxdna/amdxdna_pci_drv.c
> @@ -253,15 +253,24 @@ static int amdxdna_drm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
>  {
>  	struct drm_file *drm_filp = filp->private_data;
>  	struct amdxdna_client *client = drm_filp->driver_priv;
> -	struct amdxdna_dev *xdna = client->xdna;
> +	struct amdxdna_dev *xdna;
> +	int ret, idx;
>  
>  	if (likely(vma->vm_pgoff >= DRM_FILE_PAGE_OFFSET_START))
>  		return drm_gem_mmap(filp, vma);
>  

[Severity: High]
This is a pre-existing issue, but does this bypass of drm_dev_enter() leave
the use-after-free unresolved for standard GEM mmaps?

The patch adds drm_dev_enter() below this check, but if the condition
(vma->vm_pgoff >= DRM_FILE_PAGE_OFFSET_START) is met, it calls drm_gem_mmap()
which eventually calls amdxdna_gem_obj_mmap() without the device removal
protection. Since amdxdna_remove() frees all client structures regardless of
open file descriptors, abo->client becomes a dangling pointer here.

Additionally, do the IOCTL handlers (like amdxdna_drm_config_hwctx_ioctl in
drivers/accel/amdxdna/amdxdna_ctx.c) and GEM destruction callbacks
(like amdxdna_gem_del_bo_usage in drivers/accel/amdxdna/amdxdna_gem.c)
also need drm_dev_enter() protection? They appear to extract and dereference
filp->driver_priv without synchronization.

> -	if (!xdna->dev_info->ops->mmap)
> -		return -EOPNOTSUPP;
> +	if (!drm_dev_enter(drm_filp->minor->dev, &idx))
> +		return -ENODEV;

[ ... ]

[Severity: High]
This isn't a bug introduced by this patch, but could the error path in
amdxdna_insert_pages() lead to a double put of the GEM object reference?

In drivers/accel/amdxdna/amdxdna_gem.c:amdxdna_insert_pages(), if page
insertion fails:

	if (ret) {
		XDNA_ERR(xdna, "Failed insert pages %d", ret);
		vma->vm_ops->close(vma);
		return ret;
	}

For shmem GEM objects, calling close() drops the object reference. When this
error returns to drm_gem_mmap_obj(), it executes its error cleanup path which
calls drm_gem_object_put() again, causing the same reference to be dropped
twice.

[Severity: High]
This isn't a bug introduced by this patch, but is it safe to destroy the
notifier_wq during device removal?

In drivers/accel/amdxdna/amdxdna_pci_drv.c:amdxdna_remove():

	destroy_workqueue(xdna->notifier_wq);

If userspace retains mapped VMAs after the device is removed, a subsequent
munmap() will trigger amdxdna_hmm_invalidate() in
drivers/accel/amdxdna/amdxdna_gem.c, which queues work to the destroyed
workqueue:

	queue_work(xdna->notifier_wq, &mapp->hmm_unreg_work);

Can this lead to a use-after-free if the workqueue was already freed?

[Severity: Critical]
This isn't a bug introduced by this patch, but does caching the vm_area_struct
pointer lead to a use-after-free during munmap()?

In drivers/accel/amdxdna/amdxdna_gem.c:amdxdna_hmm_register(), the vma
pointer is cached:

	mapp->vma = vma;

The core MM subsystem can split or free VMA objects independently. When
munmap() is called, the notifier queues amdxdna_hmm_unreg_work() to run
asynchronously. By the time amdxdna_umap_release() runs, the core MM has
already freed the VMA, but the driver dereferences it:

	if (is_import_bo(abo) && vma->vm_file && vma->vm_file->f_mapping)

Does this need to be tracked differently to avoid accessing a freed VMA?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260608161418.2748988-1-lizhi.hou@amd.com?part=2