Making GPU Resets Less Painful on Linux
summary
Modern GPUs hang unpredictably; Linux handles this by driver-initiated timeouts that perform anything from an engine reset to a full device reset that destroys VRAM. Work adds a proposed DRM get_reset_state API, kernel+Mesa hooks and a vendor ioctl (AMD capture callback) to report which buffer/instruction triggered the hang. Recovery already succeeds in most cases if user space recreates contexts (Vulkan returns DEVICE_LOST; GL uses robustness), but actionable hang logs, standardized dev coredumps and IGT tests are still needed.
AI-written summary · about
chapters (10)
- 0:00GPU resets on Linux
- 1:25GPU hangs problematic
- 3:53Kernel-userspace after reset
- 6:00Standardized reset reporting
- 7:29AMD GPU recovery challenges
- 8:54Root causes of GPU hangs
- 10:42Debugging hangs with limited info
- 12:31Roadmap for reliable resets
- 14:13Q&A: Compute workloads
- 16:51Q&A: Detection and recovery
related talks
39mnouveau and kernel GPU VMA management
20mLinux Graphics 101
44mGraphics: A Frame's Journey
39mOverview of a Modern DRM Driver
41mWalking Through the Linux-Based Graphics Stack- 39mOverview of a Modern DRM Driver
46mDemystifying Graphics and Media Work Synchronisation
39mNavigating the Linux Graphics Stack