Resources needed for output poll workers are destroyed in
nouveau_fbcon_fini() before output poll workers are cleared in
nouveau_display_fini(). This means there is a time between fbcon_fini
and display_fini, where if output poll happens, it crashes.
BUG: KASAN: use-after-free in
__drm_fb_helper_initial_config_and_unlock.cold+0x1f3/0x291
[drm_kms_helper]
Cc: Ben Skeggs <[email protected]>
Cc: Karol Herbst <[email protected]>
Cc: Lyude Paul <[email protected]>
Cc: David Airlie <[email protected]>
Cc: Daniel Vetter <[email protected]>
Cc: Sumit Semwal <[email protected]>
Cc: "Christian König" <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Signed-off-by: Mark Menzynski <[email protected]>
---
drivers/gpu/drm/nouveau/nouveau_drm.c | 17 ++++++++---------
1 file changed, 8 insertions(+), 9 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c
index 561309d447e0..773efdd20d2f 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -588,12 +588,6 @@ nouveau_drm_device_init(struct drm_device *dev)
if (ret)
goto fail_dispctor;
- if (dev->mode_config.num_crtc) {
- ret = nouveau_display_init(dev, false, false);
- if (ret)
- goto fail_dispinit;
- }
-
nouveau_debugfs_init(drm);
nouveau_hwmon_init(dev);
nouveau_svm_init(drm);
@@ -601,6 +595,12 @@ nouveau_drm_device_init(struct drm_device *dev)
nouveau_fbcon_init(dev);
nouveau_led_init(dev);
+ if (dev->mode_config.num_crtc) {
+ ret = nouveau_display_init(dev, false, false);
+ if (ret)
+ goto fail_dispinit;
+ }
+
if (nouveau_pmops_runtime()) {
pm_runtime_use_autosuspend(dev->dev);
pm_runtime_set_autosuspend_delay(dev->dev, 5000);
@@ -641,15 +641,14 @@ nouveau_drm_device_fini(struct drm_device *dev)
pm_runtime_forbid(dev->dev);
}
+ if (dev->mode_config.num_crtc)
+ nouveau_display_fini(dev, false, false);
nouveau_led_fini(dev);
nouveau_fbcon_fini(dev);
nouveau_dmem_fini(drm);
nouveau_svm_fini(drm);
nouveau_hwmon_fini(dev);
nouveau_debugfs_fini(drm);
-
- if (dev->mode_config.num_crtc)
- nouveau_display_fini(dev, false, false);
nouveau_display_destroy(dev);
nouveau_accel_fini(drm);
--
2.32.0
Hmm, I think we might just need to move the drm_kms_helper_poll_enable() call
to the end here instead of all of nouveau_display_init(). I realized this
because in nouveau_display_init() it seems that we actually rely on
nouveau_display_init() to setup hotplug interrupts - which we do actually need
this early on in the driver probe process.
That being said though, drm_kms_helper_poll_enable() shouldn't be required for
MST short HPD IRQs from working so moving that instead should work.
On Wed, 2022-05-04 at 19:18 +0200, Mark Menzynski wrote:
> Resources needed for output poll workers are destroyed in
> nouveau_fbcon_fini() before output poll workers are cleared in
> nouveau_display_fini(). This means there is a time between fbcon_fini
> and display_fini, where if output poll happens, it crashes.
>
> BUG: KASAN: use-after-free in
> __drm_fb_helper_initial_config_and_unlock.cold+0x1f3/0x291
> [drm_kms_helper]
>
> Cc: Ben Skeggs <[email protected]>
> Cc: Karol Herbst <[email protected]>
> Cc: Lyude Paul <[email protected]>
> Cc: David Airlie <[email protected]>
> Cc: Daniel Vetter <[email protected]>
> Cc: Sumit Semwal <[email protected]>
> Cc: "Christian König" <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
> Signed-off-by: Mark Menzynski <[email protected]>
> ---
> drivers/gpu/drm/nouveau/nouveau_drm.c | 17 ++++++++---------
> 1 file changed, 8 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
> b/drivers/gpu/drm/nouveau/nouveau_drm.c
> index 561309d447e0..773efdd20d2f 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
> @@ -588,12 +588,6 @@ nouveau_drm_device_init(struct drm_device *dev)
> if (ret)
> goto fail_dispctor;
>
> - if (dev->mode_config.num_crtc) {
> - ret = nouveau_display_init(dev, false, false);
> - if (ret)
> - goto fail_dispinit;
> - }
> -
> nouveau_debugfs_init(drm);
> nouveau_hwmon_init(dev);
> nouveau_svm_init(drm);
> @@ -601,6 +595,12 @@ nouveau_drm_device_init(struct drm_device *dev)
> nouveau_fbcon_init(dev);
> nouveau_led_init(dev);
>
> + if (dev->mode_config.num_crtc) {
> + ret = nouveau_display_init(dev, false, false);
> + if (ret)
> + goto fail_dispinit;
> + }
> +
> if (nouveau_pmops_runtime()) {
> pm_runtime_use_autosuspend(dev->dev);
> pm_runtime_set_autosuspend_delay(dev->dev, 5000);
> @@ -641,15 +641,14 @@ nouveau_drm_device_fini(struct drm_device *dev)
> pm_runtime_forbid(dev->dev);
> }
>
> + if (dev->mode_config.num_crtc)
> + nouveau_display_fini(dev, false, false);
> nouveau_led_fini(dev);
> nouveau_fbcon_fini(dev);
> nouveau_dmem_fini(drm);
> nouveau_svm_fini(drm);
> nouveau_hwmon_fini(dev);
> nouveau_debugfs_fini(drm);
> -
> - if (dev->mode_config.num_crtc)
> - nouveau_display_fini(dev, false, false);
> nouveau_display_destroy(dev);
>
> nouveau_accel_fini(drm);
--
Cheers,
Lyude Paul (she/her)
Software Engineer at Red Hat
Also JFYI Mark - I just realized that your email client is defaulting to
sending HTML mail when you reply to me, that doesn't really make it onto the
mailing list at all so you probably want to fix that.
Also - I misspoke here again, I think it should actually be a call to
drm_kms_helper_poll_fini() instead since that will both cancel the output poll
worker and update dev->mode_config.poll_enabled to prevent the output poll
worker from starting up again. The documentation doesn't say it's fine to call
drm_kms_helper_poll_disable() after that, but it seems to be safe from looking
at the code - and I think someone just generally forgot to write docs for
drm_kms_helper_poll_fini()…
On Mon, 2022-05-09 at 13:59 -0400, Lyude Paul wrote:
> On Mon, 2022-05-09 at 15:13 +0200, Mark Menzynski wrote:
> > I think there are no direct issues with initialization in the state how it
> > is now. I suspect it's because "drm_kms_helper_poll_enable()" starts the
> > first worker thread with a delay, which gives enough time to initialize
> > required resources. I changed the initialization part to keep it
> > consistent with the finish part, which is the one causing troubles.
>
> I think I may have misspoke on what the issue was here since I was about to
> leave work. The MST bit might not actually be an issue, however I think
> nouveau_fbcon_init() being called before nouveau_display_init() would be an
> issue since nouveau_fbcon_init() can trigger a modeset - and we can't
> perform modesets before nouveau_display_init() has set things up.
>
> Looking at the docs for drm_kms_helper_poll_disable() - I think the actual
> fix here (to ensure that we still call drm_kms_helper_poll_disable() at the
> right time during suspend) would be to actually add another call to
> drm_kms_helper_poll_disable() into nouveau_fbcon_fini() before we call
> nouveau_fbcon_accel_fini() and everything after. This should make sure that
> we stop the output polling work early on driver unload, and since the docs
> mention that it's OK to disable polling more then once with
> drm_kms_helper_poll_disable() I don't see any issues with that.
>
>
> >
> > I am not sure where I could move "drm_kms_helper_poll_enable/disable()",
> > since it is defined in "drm/drm_probe_helper.c", which is only included in
> > "nouveau_display.c" and "nouveau_connector.c". Both creating a new
> > function in "nouveau_display.c", and including "probe_helper.h" and using
> > poll_enable in a different file like "nouveau_fbcon.c" seem like too big
> > changes for such small fix. I don't know.
> >
> > Can this new proposed order break something in the finish part as well?
> > Maybe it would be just better to change the order of "nouveau_drm_finish"
> > and keep the current order of "noueau_drm_init"?
> >
> > On Thu, May 5, 2022 at 9:57 PM Lyude Paul <[email protected]> wrote:
> > > Hmm, I think we might just need to move the drm_kms_helper_poll_enable()
> > > call
> > > to the end here instead of all of nouveau_display_init(). I realized
> > > this
> > > because in nouveau_display_init() it seems that we actually rely on
> > > nouveau_display_init() to setup hotplug interrupts - which we do
> > > actually need
> > > this early on in the driver probe process.
> > >
> > > That being said though, drm_kms_helper_poll_enable() shouldn't be
> > > required for
> > > MST short HPD IRQs from working so moving that instead should work.
> > >
> > > On Wed, 2022-05-04 at 19:18 +0200, Mark Menzynski wrote:
> > > > Resources needed for output poll workers are destroyed in
> > > > nouveau_fbcon_fini() before output poll workers are cleared in
> > > > nouveau_display_fini(). This means there is a time between fbcon_fini
> > > > and display_fini, where if output poll happens, it crashes.
> > > >
> > > > BUG: KASAN: use-after-free in
> > > > __drm_fb_helper_initial_config_and_unlock.cold+0x1f3/0x291
> > > > [drm_kms_helper]
> > > >
> > > > Cc: Ben Skeggs <[email protected]>
> > > > Cc: Karol Herbst <[email protected]>
> > > > Cc: Lyude Paul <[email protected]>
> > > > Cc: David Airlie <[email protected]>
> > > > Cc: Daniel Vetter <[email protected]>
> > > > Cc: Sumit Semwal <[email protected]>
> > > > Cc: "Christian König" <[email protected]>
> > > > Cc: [email protected]
> > > > Cc: [email protected]
> > > > Cc: [email protected]
> > > > Cc: [email protected]
> > > > Cc: [email protected]
> > > > Signed-off-by: Mark Menzynski <[email protected]>
> > > > ---
> > > > drivers/gpu/drm/nouveau/nouveau_drm.c | 17 ++++++++---------
> > > > 1 file changed, 8 insertions(+), 9 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
> > > > b/drivers/gpu/drm/nouveau/nouveau_drm.c
> > > > index 561309d447e0..773efdd20d2f 100644
> > > > --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
> > > > +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
> > > > @@ -588,12 +588,6 @@ nouveau_drm_device_init(struct drm_device *dev)
> > > > if (ret)
> > > > goto fail_dispctor;
> > > >
> > > > - if (dev->mode_config.num_crtc) {
> > > > - ret = nouveau_display_init(dev, false, false);
> > > > - if (ret)
> > > > - goto fail_dispinit;
> > > > - }
> > > > -
> > > > nouveau_debugfs_init(drm);
> > > > nouveau_hwmon_init(dev);
> > > > nouveau_svm_init(drm);
> > > > @@ -601,6 +595,12 @@ nouveau_drm_device_init(struct drm_device *dev)
> > > > nouveau_fbcon_init(dev);
> > > > nouveau_led_init(dev);
> > > >
> > > > + if (dev->mode_config.num_crtc) {
> > > > + ret = nouveau_display_init(dev, false, false);
> > > > + if (ret)
> > > > + goto fail_dispinit;
> > > > + }
> > > > +
> > > > if (nouveau_pmops_runtime()) {
> > > > pm_runtime_use_autosuspend(dev->dev);
> > > > pm_runtime_set_autosuspend_delay(dev->dev, 5000);
> > > > @@ -641,15 +641,14 @@ nouveau_drm_device_fini(struct drm_device *dev)
> > > > pm_runtime_forbid(dev->dev);
> > > > }
> > > >
> > > > + if (dev->mode_config.num_crtc)
> > > > + nouveau_display_fini(dev, false, false);
> > > > nouveau_led_fini(dev);
> > > > nouveau_fbcon_fini(dev);
> > > > nouveau_dmem_fini(drm);
> > > > nouveau_svm_fini(drm);
> > > > nouveau_hwmon_fini(dev);
> > > > nouveau_debugfs_fini(drm);
> > > > -
> > > > - if (dev->mode_config.num_crtc)
> > > > - nouveau_display_fini(dev, false, false);
> > > > nouveau_display_destroy(dev);
> > > >
> > > > nouveau_accel_fini(drm);
> > >
>
--
Cheers,
Lyude Paul (she/her)
Software Engineer at Red Hat