In nouveau_uvmm_bind_job_submit()'s OP_UNMAP_SPARSE arm, op->reg is set
from nouveau_uvma_region_find(), which only looks the region up and takes
no reference; a region's sole reference is its membership in
uvmm->region_mt. Two failure paths leave op->reg set: the -ENOENT check
when the region is busy, and the drm_gpuvm_sm_unmap_ops_create() failure.
The sibling nouveau_uvmm_sm_unmap_prepare() failure just below clears
op->reg; these two do not.
unwind_continue steps back one op, so the failing op is skipped by the
unwind loop and its op->reg stays set. nouveau_uvmm_bind_job_cleanup()
then enters its if (op->reg) branch and calls nouveau_uvma_region_remove()
and nouveau_uvma_region_put() on it, dropping the tree's sole reference
and freeing a region this job never created. The comment above the
cleanup loop documents the broken invariant: op->reg must be NULL on
submit failure.
This frees a live region on an unrelated failure, reachable single-job
when drm_gpuvm_sm_unmap_ops_create() returns -ENOMEM; if another job owns
the same region, its cleanup then removes and puts the freed region, a
use-after-free. Clear op->reg on both failure paths.
Fixes: b88baab82871 ("drm/nouveau: implement new VM_BIND uAPI")
Reported-by: Yuhao Jiang <[email protected]>
Cc: [email protected]
Signed-off-by: Zhenhao Wan <[email protected]>
---
drivers/gpu/drm/nouveau/nouveau_uvmm.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/drivers/gpu/drm/nouveau/nouveau_uvmm.c
b/drivers/gpu/drm/nouveau/nouveau_uvmm.c
index 849bf42c124e..3412200b21fd 100644
--- a/drivers/gpu/drm/nouveau/nouveau_uvmm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_uvmm.c
@@ -1319,6 +1319,7 @@ nouveau_uvmm_bind_job_submit(struct nouveau_job *job,
op->va.range);
if (!op->reg || op->reg->dirty) {
ret = -ENOENT;
+ op->reg = NULL;
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;
}
--
2.34.1
