[PATCH] drm/amdgpu: fix NULL dereference in gfx_v9_0_kcq() and kiq_init_queue()

[PATCH] drm/amdgpu: fix NULL dereference in gfx_v9_0_kcq() and kiq_init_queue()

[Alexey Nepomnyashih]

A potential NULL pointer dereference may occur when accessing 
tmp_mqd->cp_hqd_pq_control without verifying that tmp_mqd is non-NULL.
This may happen if mqd_backup[mqd_idx] is unexpectedly NULL.

Although a NULL check for mqd_backup[mqd_idx] existed previously, it was
moved to a position after the dereference in a recent commit, which
renders it ineffective.

Add an explicit NULL check for tmp_mqd before dereferencing its members.

Found by Linux Verification Center (linuxtesting.org) with SVACE.

Cc: stable@vger.kernel.org # v5.13+
Fixes: a330b52a9e59 ("drm/amdgpu: Init the cp MQD if it's not be initialized before")
Signed-off-by: Alexey Nepomnyashih <sdl@nppct.ru>
---
 drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
index d7db4cb907ae..134cab16a00d 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
@@ -3817,10 +3817,9 @@ static int gfx_v9_0_kiq_init_queue(struct amdgpu_ring *ring)
 	 * check mqd->cp_hqd_pq_control since this value should not be 0
 	 */
 	tmp_mqd = (struct v9_mqd *)adev->gfx.kiq[0].mqd_backup;
-	if (amdgpu_in_reset(adev) && tmp_mqd->cp_hqd_pq_control){
+	if (amdgpu_in_reset(adev) && tmp_mqd && tmp_mqd->cp_hqd_pq_control) {
 		/* for GPU_RESET case , reset MQD to a clean status */
-		if (adev->gfx.kiq[0].mqd_backup)
-			memcpy(mqd, adev->gfx.kiq[0].mqd_backup, sizeof(struct v9_mqd_allocation));
+		memcpy(mqd, adev->gfx.kiq[0].mqd_backup, sizeof(struct v9_mqd_allocation));
 
 		/* reset ring buffer */
 		ring->wptr = 0;
@@ -3863,7 +3862,7 @@ static int gfx_v9_0_kcq_init_queue(struct amdgpu_ring *ring, bool restore)
 	 */
 	tmp_mqd = (struct v9_mqd *)adev->gfx.mec.mqd_backup[mqd_idx];
 
-	if (!restore && (!tmp_mqd->cp_hqd_pq_control ||
+	if (!restore && tmp_mqd && (!tmp_mqd->cp_hqd_pq_control ||
 	    (!amdgpu_in_reset(adev) && !adev->in_suspend))) {
 		memset((void *)mqd, 0, sizeof(struct v9_mqd_allocation));
 		((struct v9_mqd_allocation *)mqd)->dynamic_cu_mask = 0xFFFFFFFF;
@@ -3874,8 +3873,7 @@ static int gfx_v9_0_kcq_init_queue(struct amdgpu_ring *ring, bool restore)
 		soc15_grbm_select(adev, 0, 0, 0, 0, 0);
 		mutex_unlock(&adev->srbm_mutex);
 
-		if (adev->gfx.mec.mqd_backup[mqd_idx])
-			memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct v9_mqd_allocation));
+		memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct v9_mqd_allocation));
 	} else {
 		/* restore MQD to a clean status */
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
-- 
2.43.0

Re: [PATCH] drm/amdgpu: fix NULL dereference in gfx_v9_0_kcq() and kiq_init_queue()

[Alex Deucher]

On Sat, May 24, 2025 at 2:14 AM Alexey Nepomnyashih <sdl@nppct.ru> wrote:
>
> A potential NULL pointer dereference may occur when accessing
> tmp_mqd->cp_hqd_pq_control without verifying that tmp_mqd is non-NULL.
> This may happen if mqd_backup[mqd_idx] is unexpectedly NULL.
>
> Although a NULL check for mqd_backup[mqd_idx] existed previously, it was
> moved to a position after the dereference in a recent commit, which
> renders it ineffective.

I don't think it's possible for mqd_backup to be NULL at this point.
We would have failed earlier in init if the mqd backup allocation
failed.

Alex

>
> Add an explicit NULL check for tmp_mqd before dereferencing its members.
>
> Found by Linux Verification Center (linuxtesting.org) with SVACE.
>
> Cc: stable@vger.kernel.org # v5.13+
> Fixes: a330b52a9e59 ("drm/amdgpu: Init the cp MQD if it's not be initialized before")
> Signed-off-by: Alexey Nepomnyashih <sdl@nppct.ru>
> ---
>  drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 10 ++++------
>  1 file changed, 4 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> index d7db4cb907ae..134cab16a00d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> @@ -3817,10 +3817,9 @@ static int gfx_v9_0_kiq_init_queue(struct amdgpu_ring *ring)
>          * check mqd->cp_hqd_pq_control since this value should not be 0
>          */
>         tmp_mqd = (struct v9_mqd *)adev->gfx.kiq[0].mqd_backup;
> -       if (amdgpu_in_reset(adev) && tmp_mqd->cp_hqd_pq_control){
> +       if (amdgpu_in_reset(adev) && tmp_mqd && tmp_mqd->cp_hqd_pq_control) {
>                 /* for GPU_RESET case , reset MQD to a clean status */
> -               if (adev->gfx.kiq[0].mqd_backup)
> -                       memcpy(mqd, adev->gfx.kiq[0].mqd_backup, sizeof(struct v9_mqd_allocation));
> +               memcpy(mqd, adev->gfx.kiq[0].mqd_backup, sizeof(struct v9_mqd_allocation));
>
>                 /* reset ring buffer */
>                 ring->wptr = 0;
> @@ -3863,7 +3862,7 @@ static int gfx_v9_0_kcq_init_queue(struct amdgpu_ring *ring, bool restore)
>          */
>         tmp_mqd = (struct v9_mqd *)adev->gfx.mec.mqd_backup[mqd_idx];
>
> -       if (!restore && (!tmp_mqd->cp_hqd_pq_control ||
> +       if (!restore && tmp_mqd && (!tmp_mqd->cp_hqd_pq_control ||
>             (!amdgpu_in_reset(adev) && !adev->in_suspend))) {
>                 memset((void *)mqd, 0, sizeof(struct v9_mqd_allocation));
>                 ((struct v9_mqd_allocation *)mqd)->dynamic_cu_mask = 0xFFFFFFFF;
> @@ -3874,8 +3873,7 @@ static int gfx_v9_0_kcq_init_queue(struct amdgpu_ring *ring, bool restore)
>                 soc15_grbm_select(adev, 0, 0, 0, 0, 0);
>                 mutex_unlock(&adev->srbm_mutex);
>
> -               if (adev->gfx.mec.mqd_backup[mqd_idx])
> -                       memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct v9_mqd_allocation));
> +               memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct v9_mqd_allocation));
>         } else {
>                 /* restore MQD to a clean status */
>                 if (adev->gfx.mec.mqd_backup[mqd_idx])
> --
> 2.43.0
>

Re: [PATCH] drm/amdgpu: fix NULL dereference in gfx_v9_0_kcq() and kiq_init_queue()

[SDL]

> On Sat, May 24, 2025 at 2:14 AM Alexey Nepomnyashih <sdl@nppct.ru> wrote:
>> A potential NULL pointer dereference may occur when accessing
>> tmp_mqd->cp_hqd_pq_control without verifying that tmp_mqd is non-NULL.
>> This may happen if mqd_backup[mqd_idx] is unexpectedly NULL.
>>
>> Although a NULL check for mqd_backup[mqd_idx] existed previously, it was
>> moved to a position after the dereference in a recent commit, which
>> renders it ineffective.
> I don't think it's possible for mqd_backup to be NULL at this point.
> We would have failed earlier in init if the mqd backup allocation
> failed.
>
> Alex
In scenarios such as GPU reset or power management resume, there is no 
strict
guarantee that amdgpu_gfx_mqd_sw_init() (via ->sw_init()) is invoked before
gfx_v9_0_kiq_init_queue(). As a result, mqd_backup[] may remain 
uninitialized,
and dereferencing it without a NULL check can lead to a crash.

Most other uses of mqd_backup[] in the driver explicitly check for NULL,
indicating that uninitialized entries are an expected condition and 
should be handled
accordingly.

Alexey

Re: [PATCH] drm/amdgpu: fix NULL dereference in gfx_v9_0_kcq() and kiq_init_queue()

[Alex Deucher]

On Wed, Jun 4, 2025 at 3:30 PM SDL <sdl@nppct.ru> wrote:
>
>
> > On Sat, May 24, 2025 at 2:14 AM Alexey Nepomnyashih <sdl@nppct.ru> wrote:
> >> A potential NULL pointer dereference may occur when accessing
> >> tmp_mqd->cp_hqd_pq_control without verifying that tmp_mqd is non-NULL.
> >> This may happen if mqd_backup[mqd_idx] is unexpectedly NULL.
> >>
> >> Although a NULL check for mqd_backup[mqd_idx] existed previously, it was
> >> moved to a position after the dereference in a recent commit, which
> >> renders it ineffective.
> > I don't think it's possible for mqd_backup to be NULL at this point.
> > We would have failed earlier in init if the mqd backup allocation
> > failed.
> >
> > Alex
> In scenarios such as GPU reset or power management resume, there is no
> strict
> guarantee that amdgpu_gfx_mqd_sw_init() (via ->sw_init()) is invoked before
> gfx_v9_0_kiq_init_queue(). As a result, mqd_backup[] may remain
> uninitialized,
> and dereferencing it without a NULL check can lead to a crash.
>
> Most other uses of mqd_backup[] in the driver explicitly check for NULL,
> indicating that uninitialized entries are an expected condition and
> should be handled
> accordingly.

sw_init() is only called once at driver load time.  everything is
allocated at that point.  If that fails, the driver would not have
loaded in the first place.  I don't think it's possible for it to be
NULL.

Alex

>
> Alexey

Re: [PATCH] drm/amdgpu: fix NULL dereference in gfx_v9_0_kcq() and kiq_init_queue()

[SDL]

04.06.2025 22:34, Alex Deucher пишет:
> On Wed, Jun 4, 2025 at 3:30 PM SDL <sdl@nppct.ru> wrote:
>>
>>> On Sat, May 24, 2025 at 2:14 AM Alexey Nepomnyashih <sdl@nppct.ru> wrote:
>>>> A potential NULL pointer dereference may occur when accessing
>>>> tmp_mqd->cp_hqd_pq_control without verifying that tmp_mqd is non-NULL.
>>>> This may happen if mqd_backup[mqd_idx] is unexpectedly NULL.
>>>>
>>>> Although a NULL check for mqd_backup[mqd_idx] existed previously, it was
>>>> moved to a position after the dereference in a recent commit, which
>>>> renders it ineffective.
>>> I don't think it's possible for mqd_backup to be NULL at this point.
>>> We would have failed earlier in init if the mqd backup allocation
>>> failed.
>>>
>>> Alex
>> In scenarios such as GPU reset or power management resume, there is no
>> strict
>> guarantee that amdgpu_gfx_mqd_sw_init() (via ->sw_init()) is invoked before
>> gfx_v9_0_kiq_init_queue(). As a result, mqd_backup[] may remain
>> uninitialized,
>> and dereferencing it without a NULL check can lead to a crash.
>>
>> Most other uses of mqd_backup[] in the driver explicitly check for NULL,
>> indicating that uninitialized entries are an expected condition and
>> should be handled
>> accordingly.
> sw_init() is only called once at driver load time.  everything is
> allocated at that point.  If that fails, the driver would not have
> loaded in the first place.  I don't think it's possible for it to be
> NULL.
>
> Alex
Thanks for the review! I agree with your point.

Alexey