Repository navigation
Closed
Labels
driversDevice drivers (virtio, block, GPU, keyboard)Device drivers (virtio, block, GPU, keyboard)enhancementNew feature or requestNew feature or requestkernelKernel-space changesKernel-space changesuserspaceUserspace libraries and programsUserspace libraries and programs
Description
Summary
Move device drivers (keyboard, block, GPU) out of the kernel into userspace services, building on the service protocol framework (#26). Drivers would register as discoverable services (e.g., service:/keyboard) and clients would connect using typed protocols.
Current state
- All device drivers run in kernel space: virtio-keyboard, virtio-block, virtio-gpu
- Drivers directly access MMIO regions and handle interrupts in kernel context
- No mechanism to map device memory or deliver interrupts to userspace
Desired behaviour
Userspace drivers run as regular processes with special kernel-granted capabilities:
- MMIO mapping: Map PCI BAR regions into driver process address space
- Interrupt delivery:
EVENT_IRQposted to driver's mailbox when device interrupt fires - DMA buffers: Allocate physically contiguous memory mapped into userspace with known physical addresses
- Device ownership: Exclusive access per PCI device (prevent multiple drivers claiming the same device)
- Service registration: Driver registers as
service:/keyboard(or similar) for client discovery
Required kernel primitives
- Userspace MMIO mapping: New syscall to map a PCI BAR into the calling process. Requires permission checks (only granted to driver processes).
- Userspace interrupt delivery: When a device interrupt fires, the kernel posts an
EVENT_IRQto the owning driver's mailbox instead of calling a kernel handler. Uses the mailbox infrastructure from Service manager step 2: Timer resource (mailbox-integrated one-shot timers) #28. - DMA buffer allocation: Allocate physically contiguous pages and map them into userspace. Return both virtual and physical addresses. Needed for virtio descriptor rings and data buffers.
- Device ownership: A registry tracking which process owns which PCI device. Prevents conflicts and ensures cleanup on process exit.
- IOMMU support (for real hardware): Program the IOMMU to restrict DMA to allocated buffers only. Not needed for QEMU/virtio but essential for real hardware security.
Migration path
- Implement kernel primitives (MMIO mapping, IRQ delivery, DMA allocation)
- Port one simple driver first (keyboard) as proof of concept
- Port block driver (more complex due to DMA)
- Port GPU driver (most complex)
- Remove in-kernel driver code once userspace versions are stable
Key files
panda-kernel/src/drivers/— current in-kernel drivers (to be replaced)panda-kernel/src/syscall/— new syscalls for MMIO, IRQ, DMApanda-kernel/src/pci/— device ownership tracking- New userspace driver packages in
userspace/drivers/
Testing
- Kernel test: MMIO mapping syscall maps correct physical range
- Kernel test: IRQ delivery posts event to mailbox
- Kernel test: DMA allocation returns physically contiguous memory
- Integration test: userspace keyboard driver receives key events
- Integration test: userspace block driver reads/writes correctly
- Run existing tests against userspace drivers to verify functional equivalence
Documentation
- Document each kernel primitive (MMIO, IRQ, DMA) syscall API
- Document the driver development guide (how to write a userspace driver)
- Document the migration plan and compatibility considerations
- Document IOMMU requirements for real hardware
Metadata
Metadata
Assignees
Labels
driversDevice drivers (virtio, block, GPU, keyboard)Device drivers (virtio, block, GPU, keyboard)enhancementNew feature or requestNew feature or requestkernelKernel-space changesKernel-space changesuserspaceUserspace libraries and programsUserspace libraries and programs
Activity
michaelmelanson commentedon May 15, 2026
Superseded by #88 (IOMMU support) and #89 (userspace device driver model), which expand significantly on this design — ELF-based driver metadata, bus-agnostic subscription model, device tokens, hot-plug as the default case, and a standalone
crates/iommu/crate. See those issues for the full plan.Scrub migration-history comments now that M3-M5 is complete (task #44)