Repository navigation
Since v0.3.0:
- readme: CUDA through virtio-nvgpu, and the aperture it needs (@wanjohiryan)
- docs: a short README, NVIDIA first, with how many guests fit (@wanjohiryan)
- security: a short SECURITY at the root, with a reporting address (@wanjohiryan)
- chore: drop PROGRESS.md (@wanjohiryan)
- security: it has no measurable performance cost (@wanjohiryan)
- ci: send the draft's since date in UTC, so authors resolve to @login (@wanjohiryan)
- docs: Without requiring scheduling from us (@wanjohiryan)
- feat: Some fixes and optimizations (#13) (@DatCaptainHorse)
- perf(gpu): keep the worker fast on CPUs it shares with the host (@DatCaptainHorse)
- chore(patches): rebase the virglrenderer patches onto 123e0bc (@DatCaptainHorse)
- fix(gpu): tell a guest its VRAM budget, and let the kernel hold it there (@DatCaptainHorse)
- fix(virtiofs): map the guest's owner onto the directory's owner, not ours (@DatCaptainHorse)
- fix(patches): the VRAM budget no longer needs VM_ALWAYS_VALID stripped (@DatCaptainHorse)
- fix(gpu): charge a guest's GPU time as work, at the clock it ran at (@DatCaptainHorse)
- docs(gpu): the limit is a ceiling; sharing the card is the kernel's (@DatCaptainHorse)
- docs(gpu): a guest's VRAM budget is protected, not capped (@DatCaptainHorse)
- docs: a shared GPU needs the kernel's FAIR scheduler (@DatCaptainHorse)
- Ship the jailer in the jail image (#15) (@wanjohiryan)
- fix(jailer): bind /dev/null into every jail (#16) (@wanjohiryan)
- feat(stats): device memory and spills across GPU drivers (@DatCaptainHorse)
- build: drop virglrenderer patch 0001 (strip VM_ALWAYS_VALID) (@DatCaptainHorse)
- fix(jailer): keep the groups that own the box's device nodes (#18) (@wanjohiryan)
- chore: Bump to v0.3.1 (@wanjohiryan)