Block Devices

Fuchsia Block device drivers are, like other drivers on the system, implemented as userspace services which are accessible via IPC. Programs using block devices will have one or more handles to these underlying drivers. Similar to filesystem clients, which may send “read” or “write” requests to servers by encoding these requests within RPC messages, programs may act as clients to block devices, and may transmit RPC messages to a “device host” (referred to as “devhost” within Zircon). The devhost process then transforms these requests into driver-understood “I/O transactions”, where they are actually transmitted to the particular block device driver, and eventually to real hardware.

Particular block device drivers (USB, AHCI / SATA, Ramdisk, GPT, etc) implement the ZX_PROTOCOL_BLOCK_CORE prototol, which allows clients to queue transactions and query the block device.

Fast Block I/O

Block device drivers are often responsible for taking large portions of memory, and queueing requests to a particular device to either “read into” or “write from” a portion of memory. Unfortunately, as a consequence of transmitting messages of a limited size from an RPC protocol into an “I/O transaction”, repeated copying of large buffers is often required to access block devices.

To avoid this performance bottleneck, the block device drivers implement another mechanism to transmit reads and writes: a fast, FIFO-based protocol which acts on a shared VMO. Filesystems (or any other client wishing to interact with a block device) can acquire FIFOs from a block device, register a “transaction buffer”, and pass handles to VMOs to the block device. Instead of transmitting “read” or “write” messages with large buffers, a client of this protocol can instead send a fast, lightweight control message on a FIFO, indicating that the block device driver should act directly on the already-registered VMO. For example, when writing to a file, rather than passing bytes over IPC primitives directly, and copying them to a new location in the block device’s memory, a filesystem (representing the file as a VMO) could simply send a small FIFO message indicating “write N bytes directly from offset X of VMO Y to offset Z on a disk”. When combined with the “mmap” memory-mapping tools, this provides a “zero-copy” pathway directly from client programs to disk (or in the other direction) when accessing files.