| Message ID | cover.1613828726.git.eafanasova@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Introduce MMIO/PIO dispatch file descriptors (ioregionfd) | expand |
On 21/02/21 13:04, Elena Afanasova wrote: > This patchset introduces a KVM dispatch mechanism which can be used > for handling MMIO/PIO accesses over file descriptors without returning > from ioctl(KVM_RUN). This allows device emulation to run in another task > separate from the vCPU task. > > This is achieved through KVM vm ioctl for registering MMIO/PIO regions and > a wire protocol that KVM uses to communicate with a task handling an > MMIO/PIO access. > > TODOs: > * Implement KVM_EXIT_IOREGIONFD_FAILURE > * Add non-x86 arch support > * Add kvm-unittests > * Flush waiters if ioregion is deleted Hi ELena, as a quick thing that jumped at me before starting the review, you should add a test for the new API in tools/testing/selftests/kvm, as well as documentation. Ideally, patch 4 would also add a testcase that fails before and passes afterwards. Also, does this work already with io_uring? Paolo > v3: > - add FAST_MMIO bus support > - add KVM_IOREGION_DEASSIGN flag > - rename kvm_ioregion read/write file descriptors > - split ioregionfd signal handling support into two patches > - move ioregion_interrupted flag to ioregion_ctx > - reorder ioregion_ctx fields > - rework complete_ioregion operations > - add signal handling support for crossing a page boundary case > - change wire protocol license > - fix ioregionfd state machine > - remove ioregionfd_cmd info and drop appropriate macros > - add comment on ioregionfd cmds/replies serialization > - drop kvm_io_bus_finish/prepare() > > Elena Afanasova (5): > KVM: add initial support for KVM_SET_IOREGION > KVM: x86: add support for ioregionfd signal handling > KVM: implement wire protocol > KVM: add ioregionfd context > KVM: enforce NR_IOBUS_DEVS limit if kmemcg is disabled > > arch/x86/kvm/Kconfig | 1 + > arch/x86/kvm/Makefile | 1 + > arch/x86/kvm/vmx/vmx.c | 40 ++- > arch/x86/kvm/x86.c | 273 +++++++++++++++++- > include/linux/kvm_host.h | 28 ++ > include/uapi/linux/ioregion.h | 30 ++ > include/uapi/linux/kvm.h | 25 ++ > virt/kvm/Kconfig | 3 + > virt/kvm/eventfd.c | 25 ++ > virt/kvm/eventfd.h | 14 + > virt/kvm/ioregion.c | 529 ++++++++++++++++++++++++++++++++++ > virt/kvm/ioregion.h | 15 + > virt/kvm/kvm_main.c | 36 ++- > 13 files changed, 996 insertions(+), 24 deletions(-) > create mode 100644 include/uapi/linux/ioregion.h > create mode 100644 virt/kvm/eventfd.h > create mode 100644 virt/kvm/ioregion.c > create mode 100644 virt/kvm/ioregion.h >
On Sun, 2021-02-21 at 18:06 +0100, Paolo Bonzini wrote: > On 21/02/21 13:04, Elena Afanasova wrote: > > This patchset introduces a KVM dispatch mechanism which can be used > > for handling MMIO/PIO accesses over file descriptors without > > returning > > from ioctl(KVM_RUN). This allows device emulation to run in another > > task > > separate from the vCPU task. > > > > This is achieved through KVM vm ioctl for registering MMIO/PIO > > regions and > > a wire protocol that KVM uses to communicate with a task handling > > an > > MMIO/PIO access. > > > > TODOs: > > * Implement KVM_EXIT_IOREGIONFD_FAILURE > > * Add non-x86 arch support > > * Add kvm-unittests > > * Flush waiters if ioregion is deleted > > Hi ELena, > Hi Paolo, Thank you for your answer. > as a quick thing that jumped at me before starting the review, you > should add a test for the new API in tools/testing/selftests/kvm, as > well as documentation. Ideally, patch 4 would also add a testcase > that > fails before and passes afterwards. > Ok > Also, does this work already with io_uring? > I have a few kvm-unittests and QEMU testdev patch for testing base functionality. I haven't tried to run them with io_uring (only run with socket and pipes). Will do. > Paolo > > > v3: > > - add FAST_MMIO bus support > > - add KVM_IOREGION_DEASSIGN flag > > - rename kvm_ioregion read/write file descriptors > > - split ioregionfd signal handling support into two patches > > - move ioregion_interrupted flag to ioregion_ctx > > - reorder ioregion_ctx fields > > - rework complete_ioregion operations > > - add signal handling support for crossing a page boundary case > > - change wire protocol license > > - fix ioregionfd state machine > > - remove ioregionfd_cmd info and drop appropriate macros > > - add comment on ioregionfd cmds/replies serialization > > - drop kvm_io_bus_finish/prepare() > > > > Elena Afanasova (5): > > KVM: add initial support for KVM_SET_IOREGION > > KVM: x86: add support for ioregionfd signal handling > > KVM: implement wire protocol > > KVM: add ioregionfd context > > KVM: enforce NR_IOBUS_DEVS limit if kmemcg is disabled > > > > arch/x86/kvm/Kconfig | 1 + > > arch/x86/kvm/Makefile | 1 + > > arch/x86/kvm/vmx/vmx.c | 40 ++- > > arch/x86/kvm/x86.c | 273 +++++++++++++++++- > > include/linux/kvm_host.h | 28 ++ > > include/uapi/linux/ioregion.h | 30 ++ > > include/uapi/linux/kvm.h | 25 ++ > > virt/kvm/Kconfig | 3 + > > virt/kvm/eventfd.c | 25 ++ > > virt/kvm/eventfd.h | 14 + > > virt/kvm/ioregion.c | 529 > > ++++++++++++++++++++++++++++++++++ > > virt/kvm/ioregion.h | 15 + > > virt/kvm/kvm_main.c | 36 ++- > > 13 files changed, 996 insertions(+), 24 deletions(-) > > create mode 100644 include/uapi/linux/ioregion.h > > create mode 100644 virt/kvm/eventfd.h > > create mode 100644 virt/kvm/ioregion.c > > create mode 100644 virt/kvm/ioregion.h > >
On Sun, Feb 21, 2021 at 03:04:36PM +0300, Elena Afanasova wrote: > This patchset introduces a KVM dispatch mechanism which can be used > for handling MMIO/PIO accesses over file descriptors without returning > from ioctl(KVM_RUN). This allows device emulation to run in another task > separate from the vCPU task. > > This is achieved through KVM vm ioctl for registering MMIO/PIO regions and > a wire protocol that KVM uses to communicate with a task handling an > MMIO/PIO access. > > TODOs: > * Implement KVM_EXIT_IOREGIONFD_FAILURE > * Add non-x86 arch support > * Add kvm-unittests > * Flush waiters if ioregion is deleted * Add ioctl docs to api.rst * Add wire protocol docs to <linux/ioregionfd.h> Great, looks like userspace can really start trying out ioregionfd now - most features are implemented. I will do a deeper review of the state machine when you send the next revision. Stefan