[PATCH] drm/msm/disp/dpu1: avoid clearing hw interrupts if hw_intr is null during drm uninit
Avoid clearing irqs and derefernce hw_intr when hw_intr is null.
BUG: Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
Call trace: dpu_core_irq_uninstall+0x50/0xb0 dpu_irq_uninstall+0x18/0x24 msm_drm_uninit+0xd8/0x16c msm_drm_bind+0x580/0x5fc try_to_bring_up_master+0x168/0x1c0 __component_add+0xb4/0x178 component_add+0x1c/0x28 dp_display_probe+0x38c/0x400 platform_probe+0xb0/0xd0 really_probe+0xcc/0x2c8 __driver_probe_device+0xbc/0xe8 driver_probe_device+0x48/0xf0 __device_attach_driver+0xa0/0xc8 bus_for_each_drv+0x8c/0xd8 __device_attach+0xc4/0x150 device_initial_probe+0x1c/0x28
Fixes: a73033619ea ("drm/msm/dpu: squash dpu_core_irq into dpu_hw_interrupts") Signed-off-by: Vinod Polimera quic_vpolimer@quicinc.com --- drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c index c515b7c..ab28577 100644 --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c @@ -599,6 +599,9 @@ void dpu_core_irq_uninstall(struct dpu_kms *dpu_kms) { int i;
+ if (!dpu_kms->hw_intr) + return; + pm_runtime_get_sync(&dpu_kms->pdev->dev); for (i = 0; i < dpu_kms->hw_intr->total_irqs; i++) if (!list_empty(&dpu_kms->hw_intr->irq_cb_tbl[i]))
Quoting Vinod Polimera (2022-04-25 23:02:11)
Avoid clearing irqs and derefernce hw_intr when hw_intr is null.
Presumably this is only the case when the display driver doesn't fully probe and something probe defers? Can you clarify how this situation happens?
BUG: Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
Call trace: dpu_core_irq_uninstall+0x50/0xb0 dpu_irq_uninstall+0x18/0x24 msm_drm_uninit+0xd8/0x16c msm_drm_bind+0x580/0x5fc try_to_bring_up_master+0x168/0x1c0 __component_add+0xb4/0x178 component_add+0x1c/0x28 dp_display_probe+0x38c/0x400 platform_probe+0xb0/0xd0 really_probe+0xcc/0x2c8 __driver_probe_device+0xbc/0xe8 driver_probe_device+0x48/0xf0 __device_attach_driver+0xa0/0xc8 bus_for_each_drv+0x8c/0xd8 __device_attach+0xc4/0x150 device_initial_probe+0x1c/0x28
Fixes: a73033619ea ("drm/msm/dpu: squash dpu_core_irq into dpu_hw_interrupts")
The fixes tag looks odd. In dpu_core_irq_uninstall() at that commit it is dealing with 'irq_obj' which isn't a pointer. After commit f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr") dpu_core_irq_uninstall() starts using 'hw_intr' which is allocated on the heap. If we backported this patch to a place that had a73033619ea without f25f656608e3 it wouldn't make any sense.
Signed-off-by: Vinod Polimera quic_vpolimer@quicinc.com
drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c index c515b7c..ab28577 100644 --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c @@ -599,6 +599,9 @@ void dpu_core_irq_uninstall(struct dpu_kms *dpu_kms) { int i;
if (!dpu_kms->hw_intr)return;pm_runtime_get_sync(&dpu_kms->pdev->dev); for (i = 0; i < dpu_kms->hw_intr->total_irqs; i++)
On 27/04/2022 00:50, Stephen Boyd wrote:
Quoting Vinod Polimera (2022-04-25 23:02:11)
Avoid clearing irqs and derefernce hw_intr when hw_intr is null.
Presumably this is only the case when the display driver doesn't fully probe and something probe defers? Can you clarify how this situation happens?
BUG: Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
Call trace: dpu_core_irq_uninstall+0x50/0xb0 dpu_irq_uninstall+0x18/0x24 msm_drm_uninit+0xd8/0x16c msm_drm_bind+0x580/0x5fc try_to_bring_up_master+0x168/0x1c0 __component_add+0xb4/0x178 component_add+0x1c/0x28 dp_display_probe+0x38c/0x400 platform_probe+0xb0/0xd0 really_probe+0xcc/0x2c8 __driver_probe_device+0xbc/0xe8 driver_probe_device+0x48/0xf0 __device_attach_driver+0xa0/0xc8 bus_for_each_drv+0x8c/0xd8 __device_attach+0xc4/0x150 device_initial_probe+0x1c/0x28
Fixes: a73033619ea ("drm/msm/dpu: squash dpu_core_irq into dpu_hw_interrupts")
The fixes tag looks odd. In dpu_core_irq_uninstall() at that commit it is dealing with 'irq_obj' which isn't a pointer. After commit f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr") dpu_core_irq_uninstall() starts using 'hw_intr' which is allocated on the heap. If we backported this patch to a place that had a73033619ea without f25f656608e3 it wouldn't make any sense.
I'd agree here. The following tag would be correct:
Fixes: f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr")
Signed-off-by: Vinod Polimera quic_vpolimer@quicinc.com
drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c index c515b7c..ab28577 100644 --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c @@ -599,6 +599,9 @@ void dpu_core_irq_uninstall(struct dpu_kms *dpu_kms) { int i;
if (!dpu_kms->hw_intr)return;pm_runtime_get_sync(&dpu_kms->pdev->dev); for (i = 0; i < dpu_kms->hw_intr->total_irqs; i++)
Looks like our new CI has given all the answers we need :) which is a great win for the CI in my opinion.
Take a look at this report : https://gitlab.freedesktop.org/drm/msm/-/jobs/22015361
This issue seems to be because this change https://github.com/torvalds/linux/commit/169466d4e59ca204683998b7f45673ebf0e... is missing in our tree.
Without this change, what happens is that we are not hitting the return 0 because we check for ENODEV.
/* * External bridges are mandatory for eDP interfaces: one has to * provide at least an eDP panel (which gets wrapped into panel-bridge). * * For DisplayPort interfaces external bridges are optional, so * silently ignore an error if one is not present (-ENODEV). */ rc = dp_parser_find_next_bridge(dp_priv->parser); if (!dp->is_edp && rc == -ENODEV) return 0;
So, I think we should do both:
1) Since we are running CI on the tree, backport this change so that this error path doesnt hit?
2) Add this protection as well because this shows that we can indeed hit this path in EDEFER cases causing this crash.
Thanks
Abhinav
On 4/27/2022 3:53 AM, Dmitry Baryshkov wrote:
On 27/04/2022 00:50, Stephen Boyd wrote:
Quoting Vinod Polimera (2022-04-25 23:02:11)
Avoid clearing irqs and derefernce hw_intr when hw_intr is null.
Presumably this is only the case when the display driver doesn't fully probe and something probe defers? Can you clarify how this situation happens?
BUG: Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
Call trace: dpu_core_irq_uninstall+0x50/0xb0 dpu_irq_uninstall+0x18/0x24 msm_drm_uninit+0xd8/0x16c msm_drm_bind+0x580/0x5fc try_to_bring_up_master+0x168/0x1c0 __component_add+0xb4/0x178 component_add+0x1c/0x28 dp_display_probe+0x38c/0x400 platform_probe+0xb0/0xd0 really_probe+0xcc/0x2c8 __driver_probe_device+0xbc/0xe8 driver_probe_device+0x48/0xf0 __device_attach_driver+0xa0/0xc8 bus_for_each_drv+0x8c/0xd8 __device_attach+0xc4/0x150 device_initial_probe+0x1c/0x28
Fixes: a73033619ea ("drm/msm/dpu: squash dpu_core_irq into dpu_hw_interrupts")
The fixes tag looks odd. In dpu_core_irq_uninstall() at that commit it is dealing with 'irq_obj' which isn't a pointer. After commit f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr") dpu_core_irq_uninstall() starts using 'hw_intr' which is allocated on the heap. If we backported this patch to a place that had a73033619ea without f25f656608e3 it wouldn't make any sense.
I'd agree here. The following tag would be correct:
Fixes: f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr")
Signed-off-by: Vinod Polimera quic_vpolimer@quicinc.com
drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c index c515b7c..ab28577 100644 --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c @@ -599,6 +599,9 @@ void dpu_core_irq_uninstall(struct dpu_kms *dpu_kms) { int i;
+ if (!dpu_kms->hw_intr) + return;
pm_runtime_get_sync(&dpu_kms->pdev->dev); for (i = 0; i < dpu_kms->hw_intr->total_irqs; i++)
On Mon, 2 May 2022 at 04:38, Abhinav Kumar quic_abhinavk@quicinc.com wrote:
Looks like our new CI has given all the answers we need :) which is a great win for the CI in my opinion.
Take a look at this report : https://gitlab.freedesktop.org/drm/msm/-/jobs/22015361
This issue seems to be because this change https://github.com/torvalds/linux/commit/169466d4e59ca204683998b7f45673ebf0e... is missing in our tree.
Without this change, what happens is that we are not hitting the return 0 because we check for ENODEV.
/* * External bridges are mandatory for eDP interfaces: one has to * provide at least an eDP panel (which gets wrapped into panel-bridge). * * For DisplayPort interfaces external bridges are optional, so * silently ignore an error if one is not present (-ENODEV). */ rc = dp_parser_find_next_bridge(dp_priv->parser); if (!dp->is_edp && rc == -ENODEV) return 0;
So, I think we should do both:
- Since we are running CI on the tree, backport this change so that
this error path doesnt hit?
- Add this protection as well because this shows that we can indeed hit
this path in EDEFER cases causing this crash.
I have been waiting for v2 for the last week or so. It should include a fixed Fixes tag and an updated description (which should note that this happens in the error path, etc) as requested by Stephen.
Thanks
Abhinav
On 4/27/2022 3:53 AM, Dmitry Baryshkov wrote:
On 27/04/2022 00:50, Stephen Boyd wrote:
Quoting Vinod Polimera (2022-04-25 23:02:11)
Avoid clearing irqs and derefernce hw_intr when hw_intr is null.
Presumably this is only the case when the display driver doesn't fully probe and something probe defers? Can you clarify how this situation happens?
BUG: Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
Call trace: dpu_core_irq_uninstall+0x50/0xb0 dpu_irq_uninstall+0x18/0x24 msm_drm_uninit+0xd8/0x16c msm_drm_bind+0x580/0x5fc try_to_bring_up_master+0x168/0x1c0 __component_add+0xb4/0x178 component_add+0x1c/0x28 dp_display_probe+0x38c/0x400 platform_probe+0xb0/0xd0 really_probe+0xcc/0x2c8 __driver_probe_device+0xbc/0xe8 driver_probe_device+0x48/0xf0 __device_attach_driver+0xa0/0xc8 bus_for_each_drv+0x8c/0xd8 __device_attach+0xc4/0x150 device_initial_probe+0x1c/0x28
Fixes: a73033619ea ("drm/msm/dpu: squash dpu_core_irq into dpu_hw_interrupts")
The fixes tag looks odd. In dpu_core_irq_uninstall() at that commit it is dealing with 'irq_obj' which isn't a pointer. After commit f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr") dpu_core_irq_uninstall() starts using 'hw_intr' which is allocated on the heap. If we backported this patch to a place that had a73033619ea without f25f656608e3 it wouldn't make any sense.
I'd agree here. The following tag would be correct:
Fixes: f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr")
Signed-off-by: Vinod Polimera quic_vpolimer@quicinc.com
drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c index c515b7c..ab28577 100644 --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c @@ -599,6 +599,9 @@ void dpu_core_irq_uninstall(struct dpu_kms *dpu_kms) { int i;
if (!dpu_kms->hw_intr)return;pm_runtime_get_sync(&dpu_kms->pdev->dev); for (i = 0; i < dpu_kms->hw_intr->total_irqs; i++)
On 5/2/2022 1:05 AM, Dmitry Baryshkov wrote:
On Mon, 2 May 2022 at 04:38, Abhinav Kumar quic_abhinavk@quicinc.com wrote:
Looks like our new CI has given all the answers we need :) which is a great win for the CI in my opinion.
Take a look at this report : https://gitlab.freedesktop.org/drm/msm/-/jobs/22015361
This issue seems to be because this change https://github.com/torvalds/linux/commit/169466d4e59ca204683998b7f45673ebf0e... is missing in our tree.
Without this change, what happens is that we are not hitting the return 0 because we check for ENODEV.
/* * External bridges are mandatory for eDP interfaces: one has to * provide at least an eDP panel (which gets wrapped intopanel-bridge). * * For DisplayPort interfaces external bridges are optional, so * silently ignore an error if one is not present (-ENODEV). */ rc = dp_parser_find_next_bridge(dp_priv->parser); if (!dp->is_edp && rc == -ENODEV) return 0;
So, I think we should do both:
- Since we are running CI on the tree, backport this change so that
this error path doesnt hit?
- Add this protection as well because this shows that we can indeed hit
this path in EDEFER cases causing this crash.
I have been waiting for v2 for the last week or so. It should include a fixed Fixes tag and an updated description (which should note that this happens in the error path, etc) as requested by Stephen.
Prior to the above CI report, we did not know what is the error path in which this happens and why.
Till then we were just speculating.
Now that we do, we can certainly add it and post a v2.
Thanks
Abhinav
On 4/27/2022 3:53 AM, Dmitry Baryshkov wrote:
On 27/04/2022 00:50, Stephen Boyd wrote:
Quoting Vinod Polimera (2022-04-25 23:02:11)
Avoid clearing irqs and derefernce hw_intr when hw_intr is null.
Presumably this is only the case when the display driver doesn't fully probe and something probe defers? Can you clarify how this situation happens?
BUG: Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
Call trace: dpu_core_irq_uninstall+0x50/0xb0 dpu_irq_uninstall+0x18/0x24 msm_drm_uninit+0xd8/0x16c msm_drm_bind+0x580/0x5fc try_to_bring_up_master+0x168/0x1c0 __component_add+0xb4/0x178 component_add+0x1c/0x28 dp_display_probe+0x38c/0x400 platform_probe+0xb0/0xd0 really_probe+0xcc/0x2c8 __driver_probe_device+0xbc/0xe8 driver_probe_device+0x48/0xf0 __device_attach_driver+0xa0/0xc8 bus_for_each_drv+0x8c/0xd8 __device_attach+0xc4/0x150 device_initial_probe+0x1c/0x28
Fixes: a73033619ea ("drm/msm/dpu: squash dpu_core_irq into dpu_hw_interrupts")
The fixes tag looks odd. In dpu_core_irq_uninstall() at that commit it is dealing with 'irq_obj' which isn't a pointer. After commit f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr") dpu_core_irq_uninstall() starts using 'hw_intr' which is allocated on the heap. If we backported this patch to a place that had a73033619ea without f25f656608e3 it wouldn't make any sense.
I'd agree here. The following tag would be correct:
Fixes: f25f656608e3 ("drm/msm/dpu: merge struct dpu_irq into struct dpu_hw_intr")
Signed-off-by: Vinod Polimera quic_vpolimer@quicinc.com
drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c index c515b7c..ab28577 100644 --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_interrupts.c @@ -599,6 +599,9 @@ void dpu_core_irq_uninstall(struct dpu_kms *dpu_kms) { int i;
if (!dpu_kms->hw_intr)return;pm_runtime_get_sync(&dpu_kms->pdev->dev); for (i = 0; i < dpu_kms->hw_intr->total_irqs; i++)
dri-devel@lists.freedesktop.org
-
Abhinav Kumar -
Dmitry Baryshkov -
Stephen Boyd -
Vinod Polimera