Am 19.04.23 um 09:00 schrieb Mikhail Gavrilov:
Christian?
I'm already looking into this, but can't figure out why we run into problems here.
What happens is that a CS is aborted without sending the job to the scheduler and in this case the cleanup function doesn't seem to work.
Christian.
❯ /usr/src/kernels/6.3.0-0.rc7.56.fc39.x86_64/scripts/faddr2line /lib/debug/lib/modules/6.3.0-0.rc7.56.fc39.x86_64/kernel/drivers/gpu/drm/scheduler/gpu-sched.ko.debug drm_sched_job_cleanup+0x9a drm_sched_job_cleanup+0x9a/0x130: drm_sched_job_cleanup at /usr/src/debug/kernel-6.3-rc7/linux-6.3.0-0.rc7.56.fc39.x86_64/drivers/gpu/drm/scheduler/sched_main.c:808 (discriminator 3) ❯ cat -s -n /usr/src/debug/kernel-6.3-rc7/linux-6.3.0-0.rc7.56.fc39.x86_64/drivers/gpu/drm/scheduler/sched_main.c | head -818 | tail -20 799 /* drm_sched_job_arm() has been called */ 800 dma_fence_put(&job->s_fence->finished); 801 } else { 802 /* aborted job before committing to run it */ 803 drm_sched_fence_free(job->s_fence); 804 } 805 806 job->s_fence = NULL; 807 808 xa_for_each(&job->dependencies, index, fence) { 809 dma_fence_put(fence); 810 } 811 xa_destroy(&job->dependencies); 812 813 } 814 EXPORT_SYMBOL(drm_sched_job_cleanup); 815 816 /** 817 * drm_sched_ready - is the scheduler ready 818 *git blame drivers/gpu/drm/scheduler/sched_main.c -L 800,819dbe48d030b285 drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-17 10:49:16 +0200 800) dma_fence_put(&job->s_fence->finished); dbe48d030b285 drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-17 10:49:16 +0200 801) } else { dbe48d030b285 drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-17 10:49:16 +0200 802) /* aborted job before committing to run it */ d4c16733e7960 drivers/gpu/drm/scheduler/sched_main.c (Boris Brezillon 2021-09-03 14:05:54 +0200 803) drm_sched_fence_free(job->s_fence); dbe48d030b285 drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-17 10:49:16 +0200 804) } dbe48d030b285 drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-17 10:49:16 +0200 805) 26efecf955889 drivers/gpu/drm/scheduler/sched_main.c (Sharat Masetty 2018-10-29 15:02:28 +0530 806) job->s_fence = NULL; ebd5f74255b9f drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-05 12:46:49 +0200 807) ebd5f74255b9f drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-05 12:46:49 +0200 808) xa_for_each(&job->dependencies, index, fence) { ebd5f74255b9f drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-05 12:46:49 +0200 809) dma_fence_put(fence); ebd5f74255b9f drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-05 12:46:49 +0200 810) } ebd5f74255b9f drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-05 12:46:49 +0200 811) xa_destroy(&job->dependencies); ebd5f74255b9f drivers/gpu/drm/scheduler/sched_main.c (Daniel Vetter 2021-08-05 12:46:49 +0200 812) 26efecf955889 drivers/gpu/drm/scheduler/sched_main.c (Sharat Masetty 2018-10-29 15:02:28 +0530 813) } 26efecf955889 drivers/gpu/drm/scheduler/sched_main.c (Sharat Masetty 2018-10-29 15:02:28 +0530 814) EXPORT_SYMBOL(drm_sched_job_cleanup); 26efecf955889 drivers/gpu/drm/scheduler/sched_main.c (Sharat Masetty 2018-10-29 15:02:28 +0530 815) e688b728228b9 drivers/gpu/drm/amd/scheduler/gpu_scheduler.c (Christian König 2015-08-20 17:01:01 +0200 816) /** 2d33948e4e00b drivers/gpu/drm/scheduler/gpu_scheduler.c (Nayan Deshmukh 2018-05-29 11:23:07 +0530 817) * drm_sched_ready - is the scheduler ready 2d33948e4e00b drivers/gpu/drm/scheduler/gpu_scheduler.c (Nayan Deshmukh 2018-05-29 11:23:07 +0530 818) * 2d33948e4e00b drivers/gpu/drm/scheduler/gpu_scheduler.c (Nayan Deshmukh 2018-05-29 11:23:07 +0530 819) * @sched: scheduler instance Daniel, because Christian, looks a little busy. Can you help? The git blame says that you are the author of code which KASAN mentions in its report. The issue is reproducible on all available AMD hardware: 6800M, 6900XT, 7900XTX.
