Disk Benchmark and I/O Diagnostics

What is it?

Disk diagnostics answer "how fast is this disk, how healthy is it, and where does my I/O time go?" This roadmap adds a diskbench command and window, SMART health reports with a score, per-device I/O statistics and latency histograms, a profiler for the filesystem layer and a background health monitor. None of its seven sections is started. The only measurement in the tree today is a latency histogram inside the VirtIO-blk driver that nothing reads.

How does it work?

VirtIO-blk latency, kept but unused. The VirtIO-blk driver timestamps each request with the CPU timestamp counter when it is submitted and again when it completes, and latency_record() adds the difference to one of four histograms (read, write, flush, discard) declared in blk_internal.h. Each histogram has seven buckets, split at 1 µs, 10 µs, 100 µs, 1 ms, 10 ms and 100 ms, and keeps a total, minimum and maximum. virtio_blk_get_latency_stats() in blk_telemetry.c turns a histogram into an average and estimated 50th, 99th and 99.9th percentiles in microseconds, but no code calls it.

Nothing generic yet. The block-device structure in blkdev.h has no statistics field, so AHCI, NVMe and USB disks record nothing, and there is no /proc/diskstats-style view. The timestamp-counter frequency measured at boot is available from boot_timing_tsc_freq() in boot_timing.c, which the benchmark will need to turn counter ticks into time. No code sends ATA SMART commands or reads the NVMe health log.

Planned design. The order is data first, tools second:

A generic blkdev_stats_t with counters and a latency histogram on every block device, updated by the block layer rather than each driver.
Per-device histograms exposed at /sys/ioqueue and through NtQuerySystemInformation, with a latency command.
SMART attribute reads for SATA drives with a computed health score.
diskbench: sequential and random reads and writes at queue depths 1, 4, 16 and 32, reporting throughput and latency percentiles, then a window that draws the results.
A filesystem profiler that costs nothing when switched off, and a low-priority health monitor thread that warns when SMART attributes worsen.

flowchart LR
    V[VirtIO-blk request] --> H[latency_record: 7-bucket histogram]
    H --> Q[virtio_blk_get_latency_stats]
    Q -. no caller .-> X[nothing reads it]
    subgraph Planned
        S[blkdev_stats_t on every disk] --> IO["/sys/ioqueue + latency command"]
        S --> B[diskbench CLI and window]
        SM[SMART read] --> HM[health monitor]
    end

What are its interfaces?

Interface	Purpose
`virtio_blk_get_latency_stats()`	Average, percentiles, minimum and maximum per VirtIO-blk request type (`blk_telemetry.c`)
`boot_timing_tsc_freq()`	Timestamp-counter frequency measured at boot (`boot_timing.c`)

Everything else in the roadmap is new interface.

How do I use it?

There is nothing to run yet. For block-device error counts on AHCI disks, see the Registry counters described in Block Storage Hardening.

What is not implemented yet?

Statistics on every disk (Generic blkdev_stats_t + Histogram API) and their view (Per-Device Latency Histogram + /sys/ioqueue). The roadmap describes a 64-bucket histogram; the VirtIO one has 7.
Benchmarking (diskbench CLI, Disk Benchmark GUI).
Health (SMART Extended Diagnostics, Storage Health Monitor Daemon), which builds on the SMART read in AHCI SMART Read.
Filesystem profiling (VFS Hot-Path Profiler).

How does it compare with Windows 11 and Linux?

Windows 11 exposes per-disk performance counters and a drive health view in Settings, but has no built-in benchmark; people use CrystalDiskMark and CrystalDiskInfo. Linux has /proc/diskstats, iostat and blktrace, with smartmontools and fio as separate packages, and GNOME Disks includes a graphical read and write benchmark with transfer-rate and access-time graphs. Neither system ships a queue-depth sweep with latency percentiles, which is what this roadmap adds.