On Tue, 2025-02-25 at 11:12 +0800, Qianyi Liu wrote: > Hello Philipp, > > Thank you for your patient reply. Let's first clarify the issue and > send a new > patch if necessary. > > As soon as it enters the drm_sched_entity_kill function, the entity > ->last_scheduled reference count is incremented by 1. If there are > still jobs in > the current entity, it will enter the while loop, assuming there is > only one job > left. If entity->last_scheduled has already been signaled, it will > enter > drm_sched_entity_kill_jobs_cb, but because null is passed in, the > last_scheduled reference count will not be correctly reduced by 1. > > Because the prev pointer has been updated to &s_fence->finished, the > dma_fence_put in the last line only reduces the reference count of > s_fence->finished. > The reference count of entity->last_scheduled was not reduced by > 1, causing a memory leak. > > We should subtract 1 from the reference count of the prev when > dma_fence_add_callback > fails, which is called balance. > > Best Regards. > QianYi.
OK, thanks for clarification. I think, next to the other feedback, it would be good to have a brief version of the above explanation in your commit message, since this clearly describes the problem and the proposed solution. Please address the feedback by Matt and myself in a v2. (and btw., please don't remove the original e-mail content when answering – that's uncommon on-list. Only when the threads are getting huge one removes parts not addressed by the answer from the quoted content ;) Thanks P.
