FlashTester

Website: flashtester.eu.

Finds out whether a USB stick, SD card or other disk really stores what it claims. FlashTester does what h2testw does and adds the things h2testw can't:

• Quick capacity probe: finds a fake drive's real capacity in minutes, not hours.
• Seeded test data: every run writes different data, so a drive that silently stopped accepting writes can't pass on a previous run's data.
• Early abort: while filling a drive, earlier data is re-checked; a fake is caught as soon as it starts overwriting itself.
• Raw mode (whole disk) as well as file mode (free space only, keeps your files).
• Speed map, stalls, sequential benchmark, random 4K IOPS, many-small-files test.
• GUI: pick a drive and a test, watch the block map fill in (the drive as a grid of cells: untested, written, verified, corrupted, aliased, unreadable, with offsets, speed and errors on hover) next to a speed-by-position graph, and read a plain verdict with the next step.
• Trust checks: unplug/replug before verifying, identity changes across replug, placeholder serials, USB 2 link warning, pull-the-plug-mid-write test, retention check later.
• Reports: a plain verdict, a detailed HTML report, JSON, exit codes, and a one-page evidence sheet for refund claims.
• Backup and restore: files in a single ZIP archive, or every sector in a whole-disk image. Backups include SHA-256 checksums; restore checks the input before writing and reads the result back afterward.
• Fix: repartition a fake drive to the capacity that really works.

The GUI

Select any disk to see its identity in the main panel, including firmware, controller hints, device IDs and volumes. Details can be refreshed, copied or saved as HTML and JSON. No separate identification dialog is needed.

Protected drives, including the system drive, offer Drive details, Read speed and Read scan. The speed check defaults to 256 MiB; the scan reads the whole disk unless you enter a size limit. Both can be cancelled. They open the disk read-only without locking or dismounting volumes. Other programs can affect the result and may respond more slowly during a check. Reading does not verify usable capacity.

Other drives also offer file-mode tests in free space and whole-disk write tests. The panel explains what each operation touches. Destructive tests are shown in red and require confirmation. Tests that need administrator rights offer to restart FlashTester elevated. While a test runs, the block map and speed graph fill in; the result includes a verdict and report buttons.

For backup and restore, select a volume for ZIP archives or the whole disk for .img images. Choose Create backup or Restore backup, then enter the backup file path. File restore also asks for a new folder on the selected volume.

Download and run

Download the latest release. Each release includes SHA-256 checksums.

One executable per platform, with no installer. Windows uses system DLLs; Linux and FreeBSD binaries are statically linked. The experimental Haiku binary uses Haiku system libraries. Started without arguments it opens the GUI (where there is one); with arguments it runs on the command line.

Platform	File	Interface
Windows 10 / 11, 64-bit	flashtester.exe	GUI and command line
Linux x86_64, any distribution	flashtester-linux-x86_64	GUI (X11 or XWayland) and command line
Linux aarch64 (64-bit ARM: Raspberry Pi with a 64-bit OS, ARM boards, routers)	flashtester-linux-aarch64	GUI (X11 or XWayland) and command line
FreeBSD 13.5 to 16.0, amd64	flashtester-freebsd-x86_64	command line
Haiku x86_64, experimental	flashtester-haiku-x86_64	command line

Windows

• Run flashtester.exe. Volumes go by drive letter (flashtester test E:), whole disks by the number flashtester list shows (--disk 3).
• File-mode tests run as a normal user. Raw-mode tests and the free-space probe need an administrator terminal; the GUI offers to restart elevated.
• Windows is kept awake while a test runs.
• The GUI needs OpenGL. Where there is none (for example a remote desktop without a GPU), use the command line.

Linux

chmod +x flashtester-linux-x86_64
./flashtester-linux-x86_64                  # GUI, inside an X11 or Wayland (XWayland) desktop
./flashtester-linux-x86_64 test /media/you/STICK
sudo ./flashtester-linux-x86_64 probe --disk sdb --destroy

• Volumes go by mount point, whole disks by kernel name (--disk sdb, --disk nvme1n1, --disk mmcblk1).
• Raw-mode tests and the free-space probe need root, and raw writes need the disk unmounted. In the GUI, "Restart as root" goes through pkexec.
• fix uses sfdisk and mkfs.* (--fs exfat, fat32, ntfs or ext4).
• The free-space probe needs a file system with fallocate and FIEMAP, such as ext4 or xfs.

OpenWrt

• Use flashtester-linux-aarch64 on most current routers and flashtester-linux-x86_64 on x86 builds. Routers with MIPS or 32-bit ARM CPUs are not supported.
• A USB stick only appears as a disk once its driver is installed: kmod-usb-storage, and for most USB 3 sticks also kmod-usb-storage-uas (without it, a UAS stick shows up on the USB bus but never becomes /dev/sdX).
• fix needs sfdisk and a formatter: exfat-mkfs for exFAT, e2fsprogs for ext4.
• Install packages with apk add (OpenWrt 25.12 and later) or opkg install (older releases).

FreeBSD

• Command line only. Built against FreeBSD 14.5; runs on 13.5 to 16.0.
• Volumes go by mount point (UFS, FAT and ZFS), whole disks by device name (--disk da0, --disk ada1, --disk nda0).
• Raw-mode tests need root. FreeBSD refuses raw writes while anything on the disk is mounted, used as swap, or held by ZFS or another GEOM layer (mirror, geli).
• fix uses gpart with newfs_msdos (--fs fat32, the default) or newfs (--fs ufs); the base system has no exFAT or NTFS formatter.
• No free-space probe: FreeBSD cannot report where a file's blocks are on the disk. Use the raw probe on an empty drive, or a full test.

Safety

System disks allow identification, read scans and whole-disk read-speed checks. Writes and file-mode tests remain blocked, with no override, even when a file benchmark is called without --write.

Protection covers disks holding the Windows folder, boot partition or page files; on Linux, disks under /, /boot, /boot/efi, /usr or /var (through LVM, RAID and device-mapper) or holding swap; on FreeBSD, the same mount points, swap and the crash-dump device (through partitions, labels, geli, gmirror, ZFS pools and file-backed memory disks). If the OS disk cannot be identified, write access fails closed. Network shares and FUSE pools such as mergerfs are refused because they cannot be mapped to one disk. Whole-disk writes erase data and require --destroy (or confirmation in the GUI) and administrator or root access. File-mode tests use free space, but a fake drive can redirect those writes onto existing files.

Backup and restore

File backups use one ZIP archive with a checksum manifest inside it. Ordinary archive tools can extract the files/ folder. FlashTester restores its own archives into a new folder, refusing to overwrite an existing one.

flashtester backup E:\photos --output C:\Backups\photos.zip
flashtester restore C:\Backups\photos.zip --to E:\restored-photos
flashtester backup --disk 3 --output C:\Backups\drive.img
flashtester restore C:\Backups\drive.img --disk 3 --destroy

Create the backup's parent directory first. Store the backup on a different physical disk from the source or restore destination; a second partition on the same disk is refused. The computer's system disk can hold backup files, but it cannot be the source or restore destination. Live system backups need snapshot support, which this version does not have.

ZIP backups preserve file contents, folder structure, empty folders and file modification times. Close programs using the source: file backups are not snapshots. Links, special files and names that cannot be represented safely across platforms are refused. Permissions, ownership, alternate data streams and other filesystem metadata are not preserved. Windows volume-root backups exclude $RECYCLE.BIN and System Volume Information, and list those exclusions in the result. Use a disk image when filesystem metadata matters.

Disk images include every byte, including partitions, boot records and unused space, so the backup disk needs at least the source disk's reported capacity in free space. Keep drive.img and drive.img.json together. Restore requires the same logical sector size and a destination at least as large as the original disk. It copies the original partition layout without expanding it; bytes beyond the image on a larger disk remain unchanged. A larger GPT destination may need its backup partition table relocated with a partitioning tool.

Whole-disk backup and restore need administrator or root access. Windows locks the source volumes during imaging; close programs using them. Linux and FreeBSD require unmounting the volumes first. FreeBSD also requires a writable source device to hold a GEOM write claim that prevents mounts, though the backup does not write data. Whole-disk backup is unavailable on Haiku because exclusive source access is not implemented there; its file backups and restores currently require BFS volumes.

Unreadable data stops the operation. Cancelled or failed output is kept for inspection and is not reported as a successful backup or restore. Restore checks the archive or image before it starts writing, then checks the restored data. File checks can be served from the operating system's cache; they do not replace testing the media or checking it again after reconnecting it. These commands do not recover already-corrupted files or test usable capacity.

Keep backup files unchanged during restore. Disk-image restore rechecks each chunk against the initial validation before writing it. A changed chunk stops the restore, but earlier chunks may already have been written. Completed output is flushed, including directory entries on Unix; this cannot guarantee that a drive or controller survives power loss without losing data.

Quick start

flashtester list                       # disks, identity, link speed, what may be tested
flashtester test E:                    # fill free space, verify (like h2testw)
flashtester probe --disk 3 --destroy   # real capacity in minutes (admin, destroys data)
flashtester fix --disk 3 --destroy     # probe, then partition+format to the real size

On Linux use a mount point and a kernel name (test /media/you/STICK, --disk sdb), on FreeBSD a device name (--disk da0).

Commands

Command	What it does	Admin
list	Disks and volumes, protection status, USB ID and link speed	no
info E: / info --disk N	Vendor, product, serial, USB VID:PID, link, volumes, suspicious identity	no
backup E:\photos --output C:\Backups\photos.zip	Save files in one ZIP archive with checksums	no
restore photos.zip --to E:\restored	Check and restore an archive into a new folder	no
backup --disk N --output drive.img	Save the whole disk and a checksum manifest	yes
restore drive.img --disk N --destroy	Check, restore and read back a whole-disk image	yes
flash-id "AD D3 10 A6 34 00"	Offline NAND part lookup, preserving all matching candidates	no
flash-layouts / decode-flash reply.bin --identity-layout 0x463c20	List recovered identity layouts / decode a saved binary response	no
test E:	Fill free space with seeded data, read it all back. --size 8G, --replug, --passes N, --cleanup	no
test --disk N --destroy	Same over the whole disk (raw)	yes
endurance E: [--passes N]	Write + verify again and again with new seeds (0 = until error / Ctrl+C)	raw only
verify E:	Re-read data from an earlier test (retention). --loops N (0 = endless), --h2w for h2testw files	raw only
probe --disk N --destroy	Quick fake-capacity probe on the raw disk	yes
probe E:	Quick probe inside free space (files are created, mapped to disk positions, and deleted)	yes
scan --disk N	Read-only surface scan: unreadable and slow regions	yes
iops E: / iops --disk N	Random 4K at queue depth 1 and 32 (raw writes need --write --destroy)	raw only
bench E: [--size 1G] / bench --disk N	Sequential write+read speed with a speed graph (speed only, not a capacity check; raw is read-only unless --write --destroy)	raw only
small-files E: [--count N]	Create, rename, delete and verify thousands of small files	no
plug-pull E:	Guided: pull the drive out while it writes; flushed data must survive	no
fix --disk N --destroy	Probe, then one partition of the real capacity (--size 7.5G / full, `--fs exfat	fat32
clean E:	Delete FlashTester's test data	no
render result.json --html r.html	Turn a saved JSON result into an HTML report / --evidence-out sheet	no

Options for every command: --json (result on stdout), --report file.html|file.json, --evidence file.html, -q (no progress).

For backup and restore, report and evidence paths must be new files. This prevents a report from replacing a backup, its manifest or another existing file. Use a different report filename for each run. If saving a report fails, the command returns exit code 3 even when the backup or restore itself completed; the printed result shows both outcomes.

Exit codes: 0 OK, 1 problem found (fake, defective, read-only, unreliable), 2 refused (protected operation, missing --destroy, not admin), 3 error, 130 cancelled.

Flash identification

The CLI supports offline NAND lookup and saved-response decoding on Windows, Linux and FreeBSD. These commands use only the data you supply and do not query a drive.

The executable embeds 1,457 NAND records recovered from USB Flash Drive Information Extractor. Supply four or six ID bytes in device order; six bytes rank exact extended-ID matches first. All primary-ID matches remain visible, including conflicting extended IDs. Raw database records, source indices, original die-count tokens and capacity values are preserved in JSON. The source capacity values are not measured usable capacity, and no capacity or die-count filter hides candidates.

flashtester flash-id "AD D3 10 A6 34 00" --json
flashtester flash-id "AD:D3:10:A6" --report nand.html
flashtester decode-flash nand.bin --nand-layout phison --report decoded.json
flashtester decode-flash nand.bin --nand-layout smi-new
flashtester decode-flash model.bin --smi-model
flashtester flash-layouts
flashtester decode-flash identity.bin --identity-layout 0x463c20 --device-info saved-info.json

NAND reply layouts are phison, smi-legacy (ID at 0x20), smi-legacy30 (legacy subtype 3, ID at 0x30), and smi-new. The 16 identity profiles decode resolved VID/PID offsets and inquiry vendor/product/revision fields; Alcor's inquiry offsets are derived from the reply's descriptor lengths. Unresolved inquiry fields are reported as unavailable, not as conflicting values. USB string contents, opaque validation fields and most model mappings remain undecoded. --device-info optionally compares a capture with a previously saved info --json result. Choosing a layout does not establish the controller family. The SMI model-marker search is limited to the recovered region of the first 512 reply bytes. Captures must be binary files of at most 64 KiB; JSON reports preserve their bytes and comparison identity, including when decoding fails. Identification evidence sheets describe offline decoding and do not claim a capacity test was performed.

Normal drive info shows USB-ID hints for known chips and broader Silicon Motion, Phison, Alcor, Chipsbank, ITE and Innostor families. A USB ID can be copied and does not confirm the chip inside. The recovered parsers have synthetic test coverage, but no authentic hardware captures have validated them. Identification does not change capacity-test verdicts.

Database provenance is embedded from src/flash-data/provenance.json. To regenerate the data from the original research pack, run python test/import-flash-db.py PATH_TO_RESEARCH_PACK. No external DLL, online database or extra runtime file is needed.

Verdicts

• OK: everything written came back intact.
• FAKE: less capacity than claimed. Data written past the real end is lost, or wraps around and overwrites the start of the drive.
• DEFECTIVE: data lost or corrupted inside the capacity.
• READ-ONLY: the drive accepts writes but keeps its old data.
• UNRELIABLE: the data verified, but there were disconnects, read errors or unstable reads. Try another port, cable or reader.

Each lost 512-byte sector is classified like h2testw does (slightly changed under 8 bits, corrupted, aliased = holds another position's data), plus stale (holds data from an earlier run: the write never landed) and unreadable. Bad chunks are re-read: the same wrong bytes each time point at the flash; different bytes or errors that clear on retry point at the connection.

How the quick probe works

Cheap fakes ignore the high address bits, so writes past the real capacity wrap around. The probe writes 4 KB blocks at about 1000 positions on a power-of-two grid across the claimed capacity, once in ascending and once in descending order, each time with a new random tag, and reads them back. A wrapped write lands exactly on another probe, which then holds the wrong position's data. That gives the real size directly. Fakes that drop writes or return errors past their end show a tail of failed probes, and the boundary is narrowed down with finer grids. Anything stranger is caught by the full test.

Limits

• File mode tests free space only. Data already on the drive isn't checked, and on a fake drive filling free space can overwrite existing files (the fake wraps onto them). Back up first.
• Software only sees errors the controller's error correction couldn't fix. There is no health or remaining-life percentage, by design.
• Repartitioning a fake drive to its real size makes it usable; it doesn't make it trustworthy.
• FreeBSD has no free-space probe (see FreeBSD above).

Tested

• Windows 11: USB sticks, an NVMe SSD in a USB enclosure, and throwaway VHDs (test\vhd-test.ps1).
• Linux on hardware: Debian 13 (OpenMediaVault) and Fedora 44 on x86_64; Armbian on ODROID C2 and N2+, Ubuntu 26.04 on AWS Graviton4 and OpenWrt 23.05 on a GL.iNet Flint 3e router, all aarch64.
• Linux in QEMU/KVM with a real USB stick passed through, installed or booted live from a USB stick: AlmaLinux 10, Alpine 3.24, Arch, Bazzite, Chimera, Debian (on RAID, LUKS and LVM), EndeavourOS, Gentoo, Guix System 1.5, Kali 2026.2, Linux Mint 22, Manjaro, NixOS 26.05, OpenWrt (x86), openSUSE Tumbleweed, Rocky Linux 10, Slackware, SystemRescue, Tiny Core 17, Ubuntu 22.04 and 26.04, Void (glibc and musl) and Zorin OS 18. Kernels 5.15 to 7.2.
• Docker: works with docker run --privileged -v /dev:/dev. Without the host's /dev, FlashTester can't tell which disks the host uses and refuses all of them.
• FreeBSD in QEMU/KVM: 13.5, 14.4, 14.5, 15.0, 15.1 and 16.0-CURRENT.
• Fake drives (wraparound, dropped writes, read-only) are covered by unit tests on a simulated drive.

Building

Rust with the GNU toolchain (no Visual Studio needed):

rustup toolchain install stable-x86_64-pc-windows-gnu
rustup default stable-x86_64-pc-windows-gnu
cargo test --release
cargo build --release                          # target\release\flashtester.exe
# Static Linux binaries from any host (Rust's bundled lld, see .cargo/config.toml):
rustup target add x86_64-unknown-linux-musl aarch64-unknown-linux-musl
cargo build --release --target x86_64-unknown-linux-musl
cargo build --release --target aarch64-unknown-linux-musl
# Static FreeBSD binary: the same lld, linked against FreeBSD's own startup files and
# static libraries. Unpack them from FreeBSD 14.5's base.txz into target/freebsd-sysroot
# and strip their debug sections first, as .github/workflows/release.yml does.
rustup target add x86_64-unknown-freebsd
cargo build --release --target x86_64-unknown-freebsd
powershell -File test\check-imports.ps1        # only Windows DLLs imported
powershell -ExecutionPolicy Bypass -File test\vhd-test.ps1   # elevated; uses a throwaway VHD

Licence

MIT. The GUI embeds the Ubuntu font (Regular and Medium; Copyright 2011 Canonical Ltd.), which is under the Ubuntu Font Licence 1.0.