Re: [PATCH] drm/nouveau/uvmm: fix UAF in bind job cleanup

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] Unwinding a successfully prepared OP_MAP_SPARSE operation in nouveau_uvmm_bind_job_submit() causes a guaranteed NULL pointer dereference.
- [High] Unwinding a successfully prepared OP_UNMAP_SPARSE operation leaves the region permanently dirty, causing uninterruptible hangs for future operations.
--

commit 203035037bce29ea5a47c8cbbdd0cff202e6dbea
Author: Peiyang He <peiyang_he@smail.nju.edu.cn>
Subject: drm/nouveau/uvmm: fix UAF in bind job cleanup

This patch clears op->reg on the unwind_continue paths for OP_UNMAP_SPARSE.
This prevents a stale region pointer from being passed to cleanup after an
async sparse unmap fails, avoiding a use-after-free.

> diff --git a/drivers/gpu/drm/nouveau/nouveau_uvmm.c b/drivers/gpu/drm/nouveau/nouveau_uvmm.c
> index f5e4756b4de4a..aad14316cbefd 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_uvmm.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_uvmm.c
> @@ -1318,6 +1318,7 @@ nouveau_uvmm_bind_job_submit(struct nouveau_job *job,
>  			op->reg = nouveau_uvma_region_find(uvmm, op->va.addr,
>  							   op->va.range);
>  			if (!op->reg || op->reg->dirty) {
> +				op->reg = NULL;
>  				ret = -ENOENT;
>  				goto unwind_continue;
>  			}
> @@ -1327,6 +1328,7 @@ nouveau_uvmm_bind_job_submit(struct nouveau_job *job,
>  								op->va.range);
>  			if (IS_ERR(op->ops)) {
>  				ret = PTR_ERR(op->ops);
> +				op->reg = NULL;
>  				goto unwind_continue;
>  			}

[Severity: High]
This isn't a bug introduced by this patch, but looking at the unwind path
triggered by goto unwind_continue, does the unwind loop cause a null pointer
dereference for OP_MAP_SPARSE operations?

When unwinding a successfully prepared OP_MAP_SPARSE operation, the code
unconditionally calls drm_gpuva_ops_free(&uvmm->base, op->ops). However,
op->ops is never allocated for OP_MAP_SPARSE and remains NULL.

The drm_gpuva_ops_free() function does not check for NULL and evaluates
&(ops)->list inside the drm_gpuva_for_each_op_safe() iterator:

drm_gpuvm.c:drm_gpuva_ops_free() {
    struct drm_gpuva_op *op, *next;

    drm_gpuva_for_each_op_safe(op, next, ops) {
        list_del(&op->entry);
        ...

Could this crash if a VM_BIND job submits an OP_MAP_SPARSE operation
followed by a malformed operation that fails?

[Severity: High]
This is also a pre-existing issue in the same unwind path, but does
unwinding a successfully prepared OP_UNMAP_SPARSE operation leave the region
permanently dirty?

The prepare phase for OP_UNMAP_SPARSE calls nouveau_uvma_region_dirty(op->reg).
If a subsequent operation fails, the unwind loop restores the region to the
maple tree and unconditionally clears op->reg = NULL without clearing the
dirty flag or signaling the completion:

nouveau_uvmm.c:nouveau_uvmm_bind_job_submit() unwind path {
    ...
    case OP_UNMAP_SPARSE:
        __nouveau_uvma_region_insert(uvmm, op->reg);
        nouveau_uvmm_sm_unmap_prepare_unwind(uvmm, &op->new, op->ops);
        break;
    ...
    drm_gpuva_ops_free(&uvmm->base, op->ops);
    op->ops = NULL;
    op->reg = NULL;
    ...

Because op->reg is cleared here, the nouveau_uvmm_bind_job_cleanup() function
skips it, meaning nouveau_uvma_region_complete(op->reg) is never called.

Will future tasks attempting to map over this leaked region hang forever
waiting on &reg->complete in bind_validate_map_common()?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/7F26D29D0BD60619+20260901063255.661503-1-peiyang_he@smail.nju.edu.cn?part=1