fix(snapshot): fix memleak in lv_snapshot

[terry0012]

Description of the feature or fix

When we take a snapshot for a lv_obj, if there is a LV_LAYER_TYPE_TRANSFORM object in the children.
A new layer named new_layer will be malloced in refr_obj function.
But it will not be freed after lv_snapshot_take_to_buf done as we call lv_draw_dispatch_layer(NULL, &layer);

Because the new_layer is freed only when disp is not NULL as the below code.

                /*Remove the layer from  the display's*/
                if(disp) {
                    lv_layer_t * l2 = disp->layer_head;
                    while(l2) {
                        if(l2->next == layer_drawn) {
                            l2->next = layer_drawn->next;
                            break;
                        }
                        l2 = l2->next;
                    }

                    if(disp->layer_deinit) disp->layer_deinit(disp, layer_drawn);
                    lv_free(layer_drawn);
                }

Notes

• Update the Documentation if needed.
• Add Examples if relevant.
• Add Tests if applicable.
• If you added new options to lv_conf_template.h run lv_conf_internal_gen.py and update Kconfig.
• Run scripts/code-format.py (astyle version v3.4.12 needs to be installed) and follow the Code Conventions.
• Mark the Pull request as Draft while you are working on the first version, and mark is as Ready when it's ready for review.
• When changes were requested, re-request review to notify the maintainers.
• Help us to review this Pull Request! Anyone can approve or request changes.

[kisvegabor]

Wonderful, thank you!

Could you add a test too? Having a screenshot for a transformed layer and an assert for the memory leak would be great.

[terry0012]

Wonderful, thank you!

Could you add a test too? Having a screenshot for a transformed layer and an assert for the memory leak would be great.

OK.
I add a transform obj under active screen in function test_snapshot_should_not_leak_memory.
And test_snapshot can detect the memory leak without this patch.

By the way, I found test_snapshot will stuck in LVGL_CI_USING_SYS_HEAP mode if lv_obj_set_style_transform_rotation is added.
I will continue to check it further.

[kisvegabor]

I've checked it an it seems it really stuck in lv_snapshot_take only when lv_obj_set_style_transform_rotation is enabled and only in the tests. Running it normally (not in tests) works well even with the very same lv_conf.h as in the tests.

[terry0012]

I've checked it an it seems it really stuck in lv_snapshot_take only when lv_obj_set_style_transform_rotation is enabled and only in the tests. Running it normally (not in tests) works well even with the very same lv_conf.h as in the tests.

Yes. it's a dead lock problem.

[terry0012]

In the test case, we have three draw task:

• draw rect for screen obj on snapshot layer (Task 1)
• draw label for label obj on a transform layer (Task 2)
• draw transform layer on snapshot layer (Task 3)

For Task 3, the init state is LV_DRAW_TASK_STATE_WAITING as it need to wait Task 2 done.
But when we call lv_draw_dispatch_layer in lv_snapshot_take_to_draw_buf, it only traverse the snapshot layer to change the task state.

At last, when lv_draw_dispatch_wait_for_request called, it'll call lv_draw_get_next_available_task to search a draw task, but Task 3 is not available because it is in LV_DRAW_TASK_STATE_WAITING state.

At this moment, layer.draw_task_head is not NULL , LVGL thread is waiting a signal, Render thread is also waiting a signal as u->task_act is NULL. Then, a dead lock happened.

[kisvegabor]

Thank you for the investigation! I see no issue with running a full dispatching here.

[kisvegabor]

Thank you!