Re: [PATCH] drm/v3d: Fix double free of query arrays when extension parsing fails
From: 王孝雷
Date: Wed Nov 18 2026 - 22:22:03 EST
Hi Maíra,
Sorry for the early ping — I overlooked the "wait at least a week"
guidance in submitting-patches,
> From: "Maíra Canal"<mcanal@xxxxxxxxxx>
> Date: Wed, Nov 18, 2026, 21:22
> Subject: Re: [PATCH] drm/v3d: Fix double free of query arrays when extension parsing fails
> To: "Wang Xiaolei"<wangxiaolei01@xxxxxxxxxxx>, <mwen@xxxxxxxxxx>, <maarten.lankhorst@xxxxxxxxxxxxxxx>, <mripard@xxxxxxxxxx>, <tzimmermann@xxxxxxx>, <airlied@xxxxxxxxx>, <simona@xxxxxxxx>, <tvrtko.ursulin@xxxxxxxxxx>
> Cc: <dri-devel@xxxxxxxxxxxxxxxxxxxxx>, <linux-kernel@xxxxxxxxxxxxxxx>, <stable@xxxxxxxxxxxxxxx>, "Iago Toral Quiroga"<itoral@xxxxxxxxxx>
> Hi Wang,
>
> Please, be patient with kernel maintainers. Kernel etiquette recommends
> waiting at least a week before pinging maintainers [1].
>
> On 10/11/26 03:13, Wang Xiaolei wrote:
> > v3d_get_cpu_timestamp_query_params() and v3d_copy_query_info() - the
> > latter used by both performance query parsers - free the query array on
> > their error paths through v3d_timestamp_query_info_free() /
> > v3d_performance_query_info_free(), but leave query_info->queries
> > pointing at the freed buffer.
> >
> > The same job is then released by v3d_submit_cpu_ioctl()'s failure path
> > (v3d_submit_cleanup_jobs() -> v3d_job_cleanup() -> v3d_job_put() ->
> > v3d_cpu_job_free()), which calls the same helpers on the same
> > query_info. On these error paths query_info->count was never assigned -
> > it is only written after the parsing loop completes - so the syncobj
> > loop runs zero times and no syncobj is put twice, but the
> > "if (query_info->queries)" test still succeeds and the dangling pointer
> > is handed to kvfree() a second time.
> >
> > Every failure inside the parsing loop takes this path: -ENOENT from
> > drm_syncobj_find(), -EFAULT from get_user(), or -ENOMEM from
> > kvmalloc_array() in v3d_copy_query_info(). The helpers are shared
> > between that error path and the job release path.
> >
> > Clear the pointer once the buffer has been freed.
>
> This is indeed a bug, but the solution doesn't fix the ownership problem
> here. We should actually leave the cleaning process to
> v3d_cpu_job_free() instead of cleaning it in the parser. I wrote this
> draft of what I was thinking on:
>
> diff --git a/drivers/gpu/drm/v3d/v3d_sched.c
> b/drivers/gpu/drm/v3d/v3d_sched.c
> index f29d79b15be3..15bec1e35e39 100644
> --- a/drivers/gpu/drm/v3d/v3d_sched.c
> +++ b/drivers/gpu/drm/v3d/v3d_sched.c
> @@ -96,17 +96,18 @@ v3d_sched_job_free(struct drm_sched_job *sched_job)
> }
>
> void
> -v3d_timestamp_query_info_free(struct v3d_timestamp_query_info *query_info,
> - unsigned int count)
> +v3d_timestamp_query_info_free(struct v3d_timestamp_query_info *query_info)
> {
> - if (query_info->queries) {
> - unsigned int i;
> + unsigned int i;
>
> - for (i = 0; i < count; i++)
> - drm_syncobj_put(query_info->queries[i].syncobj);
> + for (i = 0; i < query_info->count; i++) {
> + if (!query_info->queries[i].syncobj)
> + break;
>
> - kvfree(query_info->queries);
> + drm_syncobj_put(query_info->queries[i].syncobj);
> }
> +
> + kvfree(query_info->queries);
> }
>
> void
> diff --git a/drivers/gpu/drm/v3d/v3d_submit.c
> b/drivers/gpu/drm/v3d/v3d_submit.c
> index eada30a31701..a3e92e097423 100644
> --- a/drivers/gpu/drm/v3d/v3d_submit.c
> +++ b/drivers/gpu/drm/v3d/v3d_submit.c
> @@ -734,7 +734,6 @@ v3d_get_cpu_timestamp_query_params(struct drm_file
> *file_priv,
> struct drm_v3d_timestamp_query timestamp;
> struct v3d_timestamp_query_info *query_info =
> &job->timestamp_query;
> unsigned int i;
> - int err;
>
> if (!v3d_validate_cpu_job(file_priv, job))
> return -EINVAL;
> @@ -747,7 +746,7 @@ v3d_get_cpu_timestamp_query_params(struct drm_file
> *file_priv,
>
> job->job_type = V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY;
>
> - query_info->queries = kvmalloc_objs(struct v3d_timestamp_query,
> + query_info->queries = kvzalloc_objs(struct v3d_timestamp_query,
> timestamp.count);
> if (!query_info->queries)
> return -ENOMEM;
> @@ -758,32 +757,22 @@ v3d_get_cpu_timestamp_query_params(struct drm_file
> *file_priv,
> for (i = 0; i < timestamp.count; i++) {
> u32 offset, sync;
>
> - if (get_user(offset, offsets++)) {
> - err = -EFAULT;
> - goto error;
> - }
> + if (get_user(offset, offsets++))
> + return -EFAULT;
>
> query_info->queries[i].offset = offset;
>
> - if (get_user(sync, syncs++)) {
> - err = -EFAULT;
> - goto error;
> - }
> + if (get_user(sync, syncs++))
> + return -EFAULT;
>
> query_info->queries[i].syncobj =
> drm_syncobj_find(file_priv,
> sync);
> - if (!query_info->queries[i].syncobj) {
> - err = -ENOENT;
> - goto error;
> - }
> + if (!query_info->queries[i].syncobj)
> + return -ENOENT;
> }
> query_info->count = timestamp.count;
>
> return 0;
> -
> -error:
> - v3d_timestamp_query_info_free(&job->timestamp_query, i);
> - return err;
> }
>
> It doesn't even compile, it's just a draft to give you a more concrete
> example of what I'm thinking on.
And thanks for the pointer on the ownership problem — you are right,
the parser should not be the one releasing the arrays. Following your
draft, I'll rework it so that v3d_cpu_job_free() is the only place that frees
them: the parsers will just return on error, ->count will be published
before any entry is filled (so a partially built array is still cleanable),
the arrays will be zero allocated, and the same treatment will be
applied to the performance path and to all of the parsers, not only
the timestamp one. I'll send a v2 once it builds.
thanks
xiaolei
>
> [1]
> https://docs.kernel.org/process/submitting-patches.html#don-t-get-discouraged-or-impatient
>
> Best regards,
> - Maíra
>
> >
> > Fixes: 0e50fcc20bd8 ("drm/v3d: Fix potential memory leak in the timestamp extension")
> > Fixes: 32df4abc44f2 ("drm/v3d: Fix potential memory leak in the performance extension")
> > Cc: stable@xxxxxxxxxxxxxxx
> > Cc: Maíra Canal <mcanal@xxxxxxxxxx>
> > Cc: Iago Toral Quiroga <itoral@xxxxxxxxxx>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxx>
> > Signed-off-by: Wang Xiaolei <wangxiaolei01@xxxxxxxxxxx>
> > ---
> > drivers/gpu/drm/v3d/v3d_sched.c | 2 ++
> > 1 file changed, 2 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
> > index 1924e3440dc9..a96023b5dce0 100644
> > --- a/drivers/gpu/drm/v3d/v3d_sched.c
> > +++ b/drivers/gpu/drm/v3d/v3d_sched.c
> > @@ -106,6 +106,7 @@ v3d_timestamp_query_info_free(struct v3d_timestamp_query_info *query_info,
> > drm_syncobj_put(query_info->queries[i].syncobj);
> >
> > kvfree(query_info->queries);
> > + query_info->queries = NULL;
> > }
> > }
> >
> > @@ -122,6 +123,7 @@ v3d_performance_query_info_free(struct v3d_performance_query_info *query_info,
> > }
> >
> > kvfree(query_info->queries);
> > + query_info->queries = NULL;
> > }
> > }
> >
>