Re: [PATCH 1/2] accel/ivpu: Fix deadlock in ivpu_ms_cleanup()

[Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>]

On 3/27/2025 6:38 PM, Lizhi Hou wrote:
> 
> On 3/26/25 01:06, Jacek Lawrynowicz wrote:
>> Hi,
>>
>> On 3/25/2025 9:50 PM, Lizhi Hou wrote:
>>> On 3/25/25 04:43, Maciej Falkowski wrote:
>>>> From: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
>>>>
>>>> Fix deadlock in ivpu_ms_cleanup() by preventing runtime resume after
>>>> file_priv->ms_lock is acquired.
>>>>
>>>> During a failure in runtime resume, a cold boot is executed, which
>>>> calls ivpu_ms_cleanup_all(). This function calls ivpu_ms_cleanup()
>>>> that acquires file_priv->ms_lock and causes the deadlock.
>>>>
>>>> Fixes: cdfad4db7756 ("accel/ivpu: Add NPU profiling support")
>>>> Cc: <stable@vger.kernel.org> # v6.11+
>>>> Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
>>>> Signed-off-by: Maciej Falkowski <maciej.falkowski@linux.intel.com>
>>>> ---
>>>>    drivers/accel/ivpu/ivpu_ms.c | 6 ++++++
>>>>    1 file changed, 6 insertions(+)
>>>>
>>>> diff --git a/drivers/accel/ivpu/ivpu_ms.c b/drivers/accel/ivpu/ivpu_ms.c
>>>> index ffe7b10f8a76..eb485cf15ad6 100644
>>>> --- a/drivers/accel/ivpu/ivpu_ms.c
>>>> +++ b/drivers/accel/ivpu/ivpu_ms.c
>>>> @@ -4,6 +4,7 @@
>>>>     */
>>>>      #include <drm/drm_file.h>
>>>> +#include <linux/pm_runtime.h>
>>>>      #include "ivpu_drv.h"
>>>>    #include "ivpu_gem.h"
>>>> @@ -281,6 +282,9 @@ int ivpu_ms_get_info_ioctl(struct drm_device *dev, void *data, struct drm_file *
>>>>    void ivpu_ms_cleanup(struct ivpu_file_priv *file_priv)
>>>>    {
>>>>        struct ivpu_ms_instance *ms, *tmp;
>>>> +    struct ivpu_device *vdev = file_priv->vdev;
>>>> +
>>>> +    pm_runtime_get_sync(vdev->drm.dev);
>>> Could get_sync() be failed here? Maybe it is better to add warning for failure?
>> Yes, this could fail but we already have detailed warnings in runtime resume callback (ivpu_pm_runtime_resume_cb()).
> 
> Will the deadlock still happens if this function fails?

No. The deadlock was caused by runtime resume in free_instance().
pm_runtime_get_sync() will always bump PM usage counter, so there will be no resume regardless if it fails or not.

>>>>          mutex_lock(&file_priv->ms_lock);
>>>>    @@ -293,6 +297,8 @@ void ivpu_ms_cleanup(struct ivpu_file_priv *file_priv)
>>>>            free_instance(file_priv, ms);
>>>>          mutex_unlock(&file_priv->ms_lock);
>>>> +
>>>> +    pm_runtime_put_autosuspend(vdev->drm.dev);
>>>>    }
>>>>      void ivpu_ms_cleanup_all(struct ivpu_device *vdev)
>> Regards,
>> Jacek
>>