Re: [lvc-project] [PATCH v2] drm/amd/pm: check return value of amdgpu_irq_add_id()

[Alex Deucher]

On Thu, Oct 3, 2024 at 10:28 AM Fedor Pchelkin <pchelkin@xxxxxxxxx> wrote:
>
> On Wed, 02. Oct 16:01, Igor Artemiev wrote:
> > amdgpu_irq_ad_id() may fail and the irq handlers will not be registered.
> > This patch adds error code check.
> >
> > Found by Linux Verification Center (linuxtesting.org) with static
> > analysis tool SVACE.
> >
> > Signed-off-by: Igor Artemiev <Igor.A.Artemiev@xxxxxxx>
> > ---
> > v2: Remove the cast to struct amdgpu_device as Christian König
> > <christian.koenig@xxxxxxx> suggested.
> >
> >  .../drm/amd/pm/powerplay/hwmgr/smu_helper.c   | 19 ++++++++++++++++---
> >  1 file changed, 16 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/pm/powerplay/hwmgr/smu_helper.c b/drivers/gpu/drm/amd/pm/powerplay/hwmgr/smu_helper.c
> > index 79a566f3564a..50a3085c00aa 100644
> > --- a/drivers/gpu/drm/amd/pm/powerplay/hwmgr/smu_helper.c
> > +++ b/drivers/gpu/drm/amd/pm/powerplay/hwmgr/smu_helper.c
> > @@ -647,28 +647,41 @@ int smu9_register_irq_handlers(struct pp_hwmgr *hwmgr)
> >  {
> >       struct amdgpu_irq_src *source =
> >               kzalloc(sizeof(struct amdgpu_irq_src), GFP_KERNEL);
> > +     int ret;
> >
> >       if (!source)
> >               return -ENOMEM;
> >
> >       source->funcs = &smu9_irq_funcs;
> >
> > -     amdgpu_irq_add_id((struct amdgpu_device *)(hwmgr->adev),
> > +     ret = amdgpu_irq_add_id(hwmgr->adev,
> >                       SOC15_IH_CLIENTID_THM,
> >                       THM_9_0__SRCID__THM_DIG_THERM_L2H,
> >                       source);
> > -     amdgpu_irq_add_id((struct amdgpu_device *)(hwmgr->adev),
> > +     if (ret)
> > +             goto err;
> > +
> > +     ret = amdgpu_irq_add_id(hwmgr->adev,
> >                       SOC15_IH_CLIENTID_THM,
> >                       THM_9_0__SRCID__THM_DIG_THERM_H2L,
> >                       source);
> > +     if (ret)
> > +             goto err;
> >
> >       /* Register CTF(GPIO_19) interrupt */
> > -     amdgpu_irq_add_id((struct amdgpu_device *)(hwmgr->adev),
> > +     ret = amdgpu_irq_add_id(hwmgr->adev,
> >                       SOC15_IH_CLIENTID_ROM_SMUIO,
> >                       SMUIO_9_0__SRCID__SMUIO_GPIO19,
> >                       source);
> > +     if (ret)
> > +             goto err;
> >
> >       return 0;
> > +
> > +err:
> > +     kfree(source);
>
> Oh, the calltrace looks like:
>
> hwmgr_sw_init()
>   phm_register_irq_handlers()
>     ->register_irq_handlers()
>     smu9_register_irq_handlers()
>
> And the return value of phm_register_irq_handlers() is not processed and
> the error is not reported anywhere, so I guess there is a risk of
> use-after-free: the source pointer may have been already registered by
> some of amdgpu_irq_add_id() calls before the error occured.
>
> The similar code exists in smu7_register_irq_handlers(), maybe should be
> fixed as well.
>
> Alex, is https://gitlab.freedesktop.org/agd5f/linux a public repo this
> patch should go in? I'd suggest to drop the patch and ask Igor to do a
> complete fix or, if dropping is not possible now, fix it by another patch.
> For the latter one I can do this myself but it would be nice to refer to
> the current patch via a git hash (it's probably not published yet in your
> repo).

Dropped.  Thanks.

Alex

>
> > +
> > +     return ret;
> >  }
> >
> >  void *smu_atom_get_data_table(void *dev, uint32_t table, uint16_t *size,
> > --
> > 2.39.2