Null Block Driver

The Rust null block driver rnull is an effort to implement a drop in replacement for null_blk in Rust.

A null block driver is a good opportunity to evaluate Rust bindings for the block layer. It is a small and simple driver and thus should be simple to reason about. Further, the null block driver is not usually deployed in production environments. Thus, it should be fairly straight forward to review, and any potential issues are not going to bring down any production workloads.

Being small and simple, the null block driver is a good place to introduce the Linux kernel storage community to Rust. This will help prepare the community for future Rust projects and facilitate a better maintenance process for these projects.

Statistics from the commit log of the C null_blk driver (before move) show that the C null block driver has had a significant amount of memory safety related problems in the past. 41% of fixes merged for the C null block driver are fixes for memory safety issues. This makes the null block driver a good candidate for rewriting in Rust.

The driver is implemented entirely in safe Rust, with all unsafe code fully contained in the abstractions that wrap the C APIs.

Features

Implemented features:

blk-mq support
Direct completion
SoftIRQ completion
Timer completion
Read and write requests
Optional memory backing

Features available in the C null_blk driver that are currently not implemented in this work:

Bio-based submission
NUMA support
Block size configuration
Multiple devices
Dynamic device creation/destruction
Queue depth configuration
Queue count configuration
Discard operation support
Cache emulation
Bandwidth throttling
Per node hctx
IO scheduler configuration
Blocking submission mode
Shared tags configuration (for >1 device)
Zoned storage support
Bad block simulation
Poll queues

Resources

6.12-rc2 Rebase (`rnull-v6.12-rc2`)

Changes from rnull-v6.10:

Make QueueData references pinned.

Performance

Setup

AMD Ryzen 5 7600
32 GB 4800 MT/s DDR5 on one channel
1x Samsung 990 Pro 1TB (PCIe 4.0 x4 16 GT/S)
NixOS 24.05

Results

Plot shows (mean_iops_r - mean_iops_c) / mean_iops_c
40 samples for each configuration
Difference of means modeled with t-distribution
P95 confidence intervals

6.11 Rebase (`rnull-v6.11`)

Changes from rnull-v6.10:

None.

Performance

Setup

AMD Ryzen 5 7600
32 GB 4800 MT/s DDR5 on one channel
1x Samsung 990 Pro 1TB (PCIe 4.0 x4 16 GT/S)
NixOS 24.05

Results

Plot shows (mean_iops_r - mean_iops_c) / mean_iops_c
40 samples for each configuration
Difference of means modeled with t-distribution
P95 confidence intervals

6.11-rc2 Rebase (`rnull-v6.11-rc2`)

Changes from rnull-v6.10:

Base abstractions merged upstream 🥳
Use atomic queue limits C API for setting queue limits.

Performance

Setup

AMD Ryzen 5 7600
32 GB 4800 MT/s DDR5 on one channel
1x Samsung 990 Pro 1TB (PCIe 4.0 x4 16 GT/S)
NixOS 24.05

Results

Plot shows (mean_iops_r - mean_iops_c) / mean_iops_c
40 samples for each configuration
Difference of means modeled with t-distribution
P95 confidence intervals

6.10 Rebase (`rnull-v6.10`)

Changes from rnull-v6.10-rc3:

None

Performance

Setup

AMD Ryzen 5 7600
32 GB 4800 MT/s DDR5 on one channel
1x Samsung 990 Pro 1TB (PCIe 4.0 x4 16 GT/S)
NixOS 24.05

Results

Plot shows (mean_iops_r - mean_iops_c) / mean_iops_c
40 samples for each configuration
Difference of means modeled with t-distribution
P95 confidence intervals

6.10-rc3 Rebase (`rnull-v6.10-rc3`)

Changes from rnull-v6.9:

Add ForeignBorrowed.
Move GenDisk to a builder pattern instead of typestate pattern.
Move block size validation from driver to abstractions.
Pin NullBlkModuel.
Refactor Request::try_set_end.
Rewrite atomic functions in terms of core library functions.
Fix a bug in timer completions where an offset was not calculated correctly.
Refactor TagSet initialization in terms of core::mem::zeroed() instead of Opaque::try_ffi_init

Performance

Setup

AMD Ryzen 5 7600
32 GB 4800 MT/s DDR5 on one channel
1x Samsung 990 Pro 1TB (PCIe 4.0 x4 16 GT/S)
NixOS 24.05

Results

Plot shows (mean_iops_r - mean_iops_c) / mean_iops_c
40 samples for each configuration
Difference of means modeled with t-distribution
P95 confidence intervals

6.9 Rebase (`rnull-v6.9`)

Changes from rnull-v6.8:

Do not rely on C refcounting of Request
Use ARef to track Request lifetime
Use Page instead of Folio to track memory for memory backed mode
Use typestate pattern to track state of GenDisk
Panic when requests cannot be completed
Remove associated type RequestDataInit and use return position impl trait instead
Call Request::start implicitly
Split helper function C file