Repository navigation
Releases
sacdfs 0.6.5
Tested as release candidate 0.6.5-rc.1; the results are in docs/test-plan.md ("Results: 0.6.5
gate"). No change to the program on top of rc.1: everything is in the 0.6.5-rc.1 section below.
The Windows guide's download commands name this release. Audio bytes are those of 0.6.4, except on
images whose tracks overlap, which #402 changes. Known: the first listing of a library on an NFS
share takes one NFS round trip more per disc area for the hole check of #402 (0.18 s to 0.32 s for
416 images; #451).
Downloads
Linux binaries, glibc 2.17 or later (RHEL 7 era and later), needing FUSE and, as a user other than root, FUSE 3's fusermount3 (package fuse3) or FUSE 2's fusermount; root needs neither:
sacdfs-0.6.5-x86_64-linux.tar.gz: generic x86-64, for any x86_64 CPU.sacdfs-0.6.5-x86_64-v3-linux.tar.gz: x86-64-v3 (AVX2, most Intel and AMD CPUs since about 2015): about 20 % less CPU for WAV files. It stops with an illegal-instruction error on older CPUs.sacdfs-0.6.5-aarch64-linux.tar.gz: aarch64 (64-bit ARM).
Linux packages of the binaries above, x86_64 and aarch64 (/usr/bin/sacdfs and a systemd unit, not enabled; they need fuse3; not signed with a package key: check the .sha256 and the attestation; packages guide):
sacdfs-0.6.5-1.x86_64.rpm,sacdfs-0.6.5-1.aarch64.rpm: Fedora, RHEL 9 and 10.sacdfs_0.6.5-1_amd64.deb,sacdfs_0.6.5-1_arm64.deb: Debian 12 and 13, Ubuntu 22.04 and 24.04.sacdfs-0.6.5-1.src.rpm: the source RPM, which rebuilds offline with the distribution's own Rust.
The packages were added after the release, from its published archives and its tag's source, by a later run on main (https://github.com/aclater/sacdfs/actions/runs/44082511107), which their attestation names.
Windows (x86_64, Windows 10 or 11): sacdfs-0.6.5-x86_64-windows.zip, sacdfs-win.exe, which needs WinFsp 2.1 or later (Windows guide). It uses WinFsp - Windows File System Proxy, Copyright (C) Bill Zissimopoulos, https://github.com/winfsp/winfsp. Not code-signed.
macOS 11 or later, mounted by the Mac's own NFS client with nothing else installed (sacdfs-nfs; macOS guide). Not signed or notarised: download with curl, as the guide shows.
sacdfs-0.6.5-aarch64-macos.tar.gz: Apple Silicon.sacdfs-0.6.5-x86_64-macos.tar.gz: Intel.
The macOS archives were added after the release, from this tag's release run (https://github.com/aclater/sacdfs/actions/runs/43951612530); their attestation names the later run on main that published them (https://github.com/aclater/sacdfs/actions/runs/44069048277).
Container image for a NAS (SMB and NFS): ghcr.io/aclater/sacdfs:0.6.5 (generic CPU; amd64 and arm64).
Installer: install.sh finds the system (a NAS or Linux computer) and sets sacdfs up (install guide). Over SSH:
curl -fsSL https://github.com/aclater/sacdfs/releases/latest/download/install.sh \
| sudo sh -s -- --iso-dir /path/to/imagesVerify a download: sha256sum -c <asset>.sha256 and gh attestation verify <asset> -R aclater/sacdfs; the image: gh attestation verify oci://ghcr.io/aclater/sacdfs:0.6.5 -R aclater/sacdfs. See the README, "Install".
Built from ed54adf.
sacdfs 0.6.5-rc.1
The Windows front end sacdfs-win and its release zip, the core's platform layer, the macOS
NFS route (built from source; no macOS archive yet), the NFS fix for hosts with kernels before
6.2, and the pacing and logging changes on main since 0.6.3. The gate results are in
docs/test-plan.md ("Results: 0.6.5 gate").
Added
- A Windows front end,
sacdfs-win(#419). It mounts SOURCE read-only as a drive letter or
a folder through WinFsp, over the same core as the Linux
mount, with the same options and defaults. Names Windows refuses are shown with the SMB
share's look-alike characters, lookups ignore case, and Ctrl+C or a service stop unmounts
cleanly. Every file of the test corpus is byte-identical to the Linux mount, checked on every
pull request by a newwindowsworkflow. Small reads are read ahead, and Windows caches file
contents until sacdfs-win reports a change: reading the corpus 4 KiB at a time takes half the
spike's time, and a second pass a twentieth
(winfsp/README.md). - A Windows download and install guide (#419). Each release has
sacdfs-<version>-x86_64-windows.zip(sacdfs-win.exewith the C runtime built in; it needs
WinFsp installed), with a checksum and a build attestation, built and checked against the
corpus on Windows by the release workflow. The guide covers the install, Explorer, starting it
in the background or at logon through WinFsp.Launcher, a drive for every user, names, and
Microsoft Defender (docs/install/windows.md). sacdfs-nfs: SACD images served over NFSv3, for the system's own NFS client (#420; feature
nfs). The route to macOS, where the Mac's NFS client mounts it as the user with nothing
installed, and a way to serve NFS clients on Linux without FUSE. It listens on 127.0.0.1 only
by default, on a port per user, and answers only its own user's calls (a guard against
mistakes: any local user can read the files). File handles stay valid across a server
restart, open files included, and the volume's size is the images' total, as on Linux. With
a MOUNTPOINT it mounts and, when stopped, unmounts. The protocol layer is vendored from
nfsserve 0.11.0 (BSD-3-Clause), with bounded decoding and a fuzz target
(design). Part of #420.- macOS: mount sacdfs on a Mac (#420).
sacdfs-nfs SOURCE MOUNTPOINTmounts with the Mac's
own NFS client, as the user, with nothing else installed, and takes the same options as the
Linux mount.--launch-agentwrites a launchd agent that mounts at each login. Built from source for now
(macOS guide): the release workflow builds and tests macOS archives
but doesn't publish them until the NFS layer moves to a maintained library. - sacdfs-win: files over 64 MiB are read past Windows's cache manager (#419). A buffered
copy of a 5.2 GB WAV ran at 11 MB/s through it and runs at 52 MB/s now (Windows 11 in a VM;
58 MB/s with--no-kernel-cache); smaller files keep the cache, which reads a second time
20 times as fast. Attributes and listings are cached as before.
Changed
- On Windows, paths in log lines are written as people write them (#419).
C:\music\a.iso
instead of\\?\C:\music\a.iso, and\\nas\musicinstead of\\?\UNC\nas\music. Display only:
the paths sacdfs opens keep the verbatim form, under which Windows takes every name literally
(a trailing dot, a device name, more than 260 characters);verifyreports keep it too, since
a resumed run matches rows by exact path. - On Windows, sidecar files are opened relative to their folder's handle, and the seek index
folder is private (#419). A folder of SOURCE renamed or replaced by a link while sacdfs walks
it can no longer send a cover's open elsewhere (NtCreateFilerelative to the folder opened,
never following a link: the Windows counterpart ofopenat), and a seek index folder made by
sacdfs gets an access list for its owner and SYSTEM only, also outside the user's profile. - The TrueNAS catalog app runs 0.6.4 (#284).
deploy/truenas/sacdfsnamed the 0.6.3 image,
which has the three advisories fixed in 0.6.4; the app's version is now 1.0.2. Rendered with
the catalog's tooling and installed on TrueNAS SCALE 25.10.8 in a VM
(store-apps.md). - The core builds for Windows and macOS (#419, #420). The calls that differ between
systems are now in one module,sacdfs::platform, and the Linux behaviour is unchanged.
fuseris Linux-only andnixandsignal-hookare Unix-only. The Linux mount moved to
src/linux_main.rs. On other systems thesacdfsbinary runsverifyandtagsand
doesn't mount
(design). - The mount options and start-up steps that every front end shares are in the library
(sacdfs::serve). The Windows and macOS front ends take the same options, defaults and help
text as the Linux mount. The Linux--helpis unchanged. - Windows: image identity, volume size, sparse areas, background priority and the seek
index's default folder use Windows's own calls (#419). These replace the first port's
stand-ins. The default folder is%LOCALAPPDATA%\sacdfs. Linux is unchanged. - The core's tests also run on Windows and macOS in CI (
core-portable). They cover the
library, theVolumeAPI, the corpus, and thetagsandverifysubcommands, on
windows-latest, macos-latest and macos-15-intel. On macOS,sacdfs verifywrites a report to
/dev/stdoutor/dev/fd/Nas it does on Linux. - The cache of frames is 32 MiB under a memory limit of 256 MiB or more; 16 MiB stays only
below 256 MiB (#271). It was 16 MiB up to a 512 MiB limit. Through the Linux kernel SMB client
that made whole-file WAV copies 12 to 31 % slower than at 32 MiB, and on 2 slow cores at 256 MiB
a player beside four copies stalled up to 3.4 s, against 1.0 s at 32 MiB with 18 % less CPU;
the default had been measured only with local reads. This is also why one WAV copy over SMB
was still 10 to 12 % slower than 0.4.1 for 6 channels (0.4.1's container always passed 32): at
32 MiB, 0.6.3 and main are 5 to 11 % faster than 0.4.1. The larger cache takes about 16 MiB
more, which the memory watch gives back above 85 % of the limit.SACDFS_CACHE_MBor
--cache-mbstill set it
(performance). - The security policy says more about disclosure and about how a report is handled.
SECURITY.md now has the terms of coordinated disclosure (private until a fixed
release or 90 days, a CVE when warranted), what to do when GitHub's private reporting can't
be used, and a wider scope: tag lookups, the installer, the container image, the service
files and the checks on release files. A new section sets out what the maintainer and contributors do with a vulnerability:
a draft advisory from the first moment, the fix in the advisory's private fork, and nothing
in public issues, pull requests, commit messages or CI logs until the fixed release is out.
CONTRIBUTING.md points there. - Every request sacdfs fails with an I/O error is logged with what failed, where and why, and
the lines are rate-limited (part of #339). A read that fails has awarnline with the image,
the track, the inode and the offset, and the cause. Reads of an image refused while it has reads
blocked past the request timeout (#241) are logged as well, and not only when the refusal starts.
Reads of a sidecar file, opens of a sidecar file that fail for a reason other than a missing or
replaced file, and requests whose work ended without an answer are logged too. Before, fuser
answered those withEIOand logged only the request's number. Lines about the same image (or
sidecar, or request) are logged at most once a minute, and at most 50 lines a minute in all;
the next line says how many were left out. So a damaged image or a hung share no longer writes
a line for every read retried, and anEIOwith no line in the log came from the kernel, not
from sacdfs. The soak (scripts/soak.py) has a--log-leveloption and prints the time of
each unexpected read error, to find it in the logs.
Fixed
- NFS from the container on hosts with kernels before 6.2 (#439). On Debian 12 (kernel
6.1), and so OpenMediaVault 7 and CasaOS on Debian 12, every listing of the NFS export failed
with "Unknown error 524". NFS-Ganesha reads each file's POSIX ACL, and those kernels answered
"not supported" for the FUSE mount; Ganesha passed that on. sacdfs now asks the kernel for
FUSE's POSIX ACL support, and the kernel answers "no ACL" itself, as newer kernels already
did. Nothing else changes: still at most oneGETXATTRreaches sacdfs per mount. Checked on
Debian 12 in a VM with docker.io: all 910 corpus files hash the same over NFS, over SMB and
inside the container; SMB is unchanged
(design, "Extended attributes"). - The installer's container route no longer warns that nothing answers on its ports on
ASUSTOR (#340). BusyBoxncthere has no-z, so the check failed although SMB and NFS
answered; it now falls back to the listening sockets. On TrueNAS, its pointer names the
guide's step 3 for the YAML to paste, not the catalog app's folder. - The unlogged EIO of the soak is explained (#339), and the soak expects it. A tags file
edit gives the changed track files a new generation under the same number. Once the kernel
sees the new generation, through any lookup or listing, it marks the old inode bad and fails a
descriptor still open on the old version withEIO, without asking sacdfs, so nothing is
logged.scripts/soak.pynow count...
Test build for #456: 32-bit x86, not a release
Test builds for #456. These are not releases, and not a supported 32-bit version.
These are one-off 32-bit x86 (i686) builds of sacdfs, made only to measure DST decoding and WAV conversion speed on an ASUSTOR AS-302T. Each is built static for i686-unknown-linux-musl with -C target-cpu=pentium4 (SSE2 at most).
- They were built by hand, not by the release workflow, so they have no build attestation. Check each file with its SHA-256 below.
- They are unsupported and get no updates. Please follow the steps and the safety check in #456 before running one.
Current build: use this one.
sacdfs-0.6.5-7cd37ea-i686-test
- sacdfs main at 7cd37ea: the third build plus decode-ahead at the start of a track (#481).
- Same scarletbook 0.6.1 and scarletbook-pcm 97733cc as the third build.
- SHA-256:
bc2e8141eb43b68d6147fd3e7818e63bda591aef8786e3224b4d39de0bbf1c9d
Earlier builds, kept so earlier results can be checked. Please don't use them for new tests.
sacdfs-0.6.5-df16b20-i686-test (third build)
- sacdfs main at df16b20, with the released scarletbook 0.6.1 (32-bit DST decoder) and scarletbook-pcm at scarletbook-rs main 97733cc (#374, #377: faster WAV conversion on 32-bit; not a release).
- SHA-256:
2fdd9960f036fdda847c665fc14552c744b8bcd1857cf94bc560f37996059a3c
sacdfs-0.6.5-d237b20-sb-d3b7aed-i686-test (second build)
- sacdfs at d237b20, with scarletbook pinned to scarletbook-rs main d3b7aed (the 32-bit DST decoder before it was released in 0.6.1).
- SHA-256:
f7e630540ee4ae1a71e5ad1cda53d274b4691d0b7a606d4894e255bcc516415d
sacdfs-0.6.5-d237b20-i686-test (first build)
- sacdfs at d237b20, with scarletbook 0.6.0.
- SHA-256:
13e70d37280cdb23d8881f97b74133dd3bbf683feacbff48296bb9a4c1fd0eac
sacdfs 0.6.4
A security release: 0.6.3 plus fixes for three advisories, and the hardening of the installer,
the container and the systemd unit. Audio bytes are unchanged from 0.6.3. The pacing changes on
main since 0.6.3 (#363, #362) are not in it; they come in 0.6.5. The gate results are in
docs/test-plan.md ("Results: 0.6.4 gate").
Security
- Open files: a client of the mount could make images disappear from their folders
(GHSA-phfg-58h9-f2pp, High).
Upgrade to 0.6.4. - Installer and
sacdfs-ctl: files that other users could change were run as root
(GHSA-wfjg-pqhm-4mm2, High).
Upgrade to 0.6.4, and on QNAP, or wherever sacdfs's folder is inside a shared folder, run the
installer again (see "Upgrading" below). - Tags, cache and report files were written and read unsafely
(GHSA-825f-7xrc-fqf7, Medium).
Upgrade to 0.6.4.
Upgrading
- Installs made with the installer: run it again as you first did. On QNAP, and wherever
sacdfs's folder was inside a shared folder, it moves sacdfs's files to a folder only root can
change (on QNAP/share/CACHEDEV1_DATA/.sacdfs, with the DSF and WAV files in
/share/CACHEDEV1_DATA/sacd) and removes the old copy; then share the newsacdfolder and
point your players at it. By hand: where the files go
and the QNAP guide. - Docker, Podman, TrueNAS, Unraid: take the image
ghcr.io/aclater/sacdfs:0.6.4, and for least
privilege the newcap_drop,cap_addandsecurity_optlines from
deploy/nas-compose.yaml (or thedocker runoptions in
docs/deployment.md). An older Compose file keeps working, with
the per-program limits applied inside the container. - systemd: copy the new
sacdfs.servicefrom the release archive; a systemd too old for an
option logsUnknown lvalueand runs the unit without it. sacdfs tags importexits with status 2 when it left out a picture or value the mount
would not use (listed asnot kept:); scripts that delete extracts after an import should
keep those extracts.
Added
- The installer stops on a system it can't support before it downloads anything, and prints
the details to ask for it (#382). Before, the container route never checked the processor
(on i686 or 32-bit ARM it pulled an image that can't run), a 64-bit kernel with 32-bit
programs (Raspberry Pi OS 32-bit on a Pi 4 or 5) got the aarch64 build, and an old glibc or
musl was only found when the downloaded binary failed to start. Now the processor and the
programs' word size are checked for every route, and glibc 2.17 or later (not musl) for the
binary routes, before the root check and any download; FUSE is checked before the download
too. On those stops, a missing FUSE,--method composewith no engine, and a binary that
doesn't start after all, it prints a short block (installer version, the check, processor,
memory, OS and NAS version, kernel, C library, FUSE, systemd, container engine; no host name,
address, serial number, user name or folder) and a link to the unsupported system form. A Mac
or Windows PC still gets the "Linux only" message, with no block.install.sh --reportprints
the block on any Linux system, as a normal user, and changes nothing, for beta test reports. - sacdfs raises its open files limit to the hard limit at start and logs a warning when that
limit is below 4096;--max-open-imagesis kept under half of it. - A fuzz target for the audio read path (#390).
fuzz/fuzz_targets/read_path.rsreads
the DSF and WAV files of damaged corpus images (plain DSD and DST, 2048-, 2054- and 2064-byte
sectors) at any offset and length, with read-ahead and small caches whose buffers are reused
across images, and checks each read against a reader with no cache. It runs in the fuzz job
with the other targets, from seeds made byexamples/fuzz_seeds.rs. Locally it ran for the
hours in its pull request, with no finding.
Changed
-
The systemd unit limits what sacdfs may do, and the mount stays visible (#387). The unit in
the release archive now keeps sacdfs to the rights to mount and to read the images
(CapabilityBoundingSet=CAP_SYS_ADMIN CAP_DAC_OVERRIDE). It also sets:- no new privileges;
- no device but
/dev/fuse; - sockets for the journal and MusicBrainz only;
- a system call deny list;
- native calls only;
- no namespaces, realtime scheduling, setuid files, writable-and-executable memory, clock
or hostname changes.
Options that give the service its own mount namespace (
ProtectSystem=,PrivateTmp=and the
like) stay out, because the mount would then be seen by no other program; the unit's comments
and docs/deployment.md say why each was kept or left out.
systemd-analyze securitygoes from 9.6 to 4.1. A systemd too old for an option logs
Unknown lvalueand runs the unit without it. Checked on Fedora 44, RHEL 10.1, TerraMaster
TOS 7.0, Debian 12 with UGREEN's kernel and ReadyNAS OS 6.10 (systemd 230), each in a VM, not on
the vendor's hardware. -
The image upgrades Fedora's own packages when it is built, and a weekly scan checks the
published image (#399). The runtime stage starts from the digest-pinned Fedora 44 base as
before, then runsdnf upgradebefore it installs Samba and the rest, so the base's packages
get Fedora 44's updates too. A scan of the 0.6.3 image with grype found 39 known
vulnerabilities, all fixed in Fedora 44 updates: 16 High in pcre2 and 23 Medium in dnf5,
libgcc, libstdc++, libgomp, libmodulemd and tpm2-tss. None were in Samba, NFS-Ganesha or fuse3.
An image built with the upgrade has none. The image is about 35 MB larger uncompressed (the
upgraded files sit on top of the pinned base layer). A new workflow,image-scan.yml, scans the
latest published image every week, and on demand, and fails on High or Critical findings that
have a fix. -
The container gives each program only the capabilities it needs (#386). The Compose files,
the TrueNAS app and the Unraid template now drop the runtime's whole default set
(cap_drop: ALL) and add back seven:SYS_ADMIN,DAC_OVERRIDE,DAC_READ_SEARCH,SETUID,
SETGID,NET_BIND_SERVICEandSETPCAP, withno-new-privileges. Inside, the entry point
starts each program withsetprivand a reduced bounding set: sacdfs keepsSYS_ADMINand
DAC_OVERRIDE, smbdSETUID,SETGIDandNET_BIND_SERVICE, and NFS-Ganesha
DAC_READ_SEARCHalone; neither server can mount, and only NFS-Ganesha can open files by
handle. At its start the entry point also drops whatever none of them need, so a container
started with an older Compose file holds no more either, and withNFS=0no process keeps
DAC_READ_SEARCH. Every capability in the list was checked to be needed by leaving it out.
Checked with the CI share test (Docker and rootful Podman, AppArmor) and on Fedora 44 in a VM
with SELinux enforcing (Docker, rootful and rootless Podman, both Compose tools); copying the
corpus over SMB and NFS took as long as before. To upgrade: take the newcap_drop,
cap_addandsecurity_optlines from deploy/nas-compose.yaml, or
thedocker runoptions in docs/deployment.md. An older file
keeps working, with the per-program limits applied. Under SELinux, the container mounts with
its normal label: docs/deployment.md lists the labels it can
read the images under. -
Open files are bounded. The sidecar files open at once are capped below the open files
limit (8 to 4096); an open past the cap fails withEMFILEfor that client alone, logged at most
once a minute. A folder listing that left out an image for a passing reason (no descriptors or
memory left,EIO) is made again after a wait (1 s, doubling to 60 s) and isn't kept by the
kernel meanwhile; an image whose opens fail withEIOthree times in a row is left out until
its file changes. Images are opened only if they are regular files. -
Where the installer puts sacdfs, and how it is checked. On QNAP sacdfs's files now live in
/share/CACHEDEV1_DATA/.sacdfs, outside every shared folder; the installer refuses a
--prefixor--mountthat other users can change, and folder names with characters other
than letters, digits, spaces and. _ - + , @ = & ( ) [ ].sacdfs-ctlreads its settings
file as data, not shell code (older files still work), and as root refuses a settings file,
program or folder that other users can change. Each release'sinstall.shcarries the
checksums of its archives;curlis used HTTPS-only. On Synology the Task Scheduler tasks are
found by name. -
Tags, state and report files are written through a new temporary file created exclusively,
then renamed into place; they are read without following a symbolic link and with a size
bound, keep the umask's mode, and a parse error says where in the file it is. Paths and messages on the
terminal, in the log and insacdfs verify's report have control characters escaped, the
frame keys in the summaries ofverifyandtags importeach on their own, and a report row
lists at most 32 other frame keys of at most 128 characters each, then "+N more". -
Names in the daemon's log lines can't split or disguise a line (#391). Image paths, folder
names, disc text and error causes were logged as they were, so a name with a newline split a
line in two, an escape sequence changed how a terminal s...
sacdfs 0.6.3
Security: this release is affected by GHSA-phfg-58h9-f2pp (High), GHSA-wfjg-pqhm-4mm2 (High) and GHSA-825f-7xrc-fqf7 (Medium). They are fixed in 0.6.4; use the latest release. Upgrade as the install guide describes.
If you installed with
install.sh, run the current one (releases/latest/download/install.sh), not this release's: this release's installer has the problem and puts it back.
Tested as release candidate 0.6.3-rc.1; the results are in docs/test-plan.md ("Results: 0.6.3-rc.1
gate"). No change on top of rc.1: everything is in the 0.6.3-rc.1 section below. A player beside
four copies on 2 slow cores over SMB stalls far less than on 0.6.2 or 0.6.1 (worst read 3.4 s
against 12.3 s and 7.7 s in the gate), though it can still stall there; on Windows the player had
no read over 100 ms beside four copies. The known issue listed for rc.1 (2–10 % more CPU for a tag
scan through a default Linux SMB mount) did not reproduce in the gate or in a second check: the
same CPU as 0.6.2. Known: a second player starting beside four copies waits longer for its first
reads than on 0.6.2 (#363); a macOS client copying four files beside a player on a busy server
stalls the player for seconds, as 0.6.2 does.
Downloads
Linux binaries, glibc 2.17 or later (RHEL 7 era and later), needing FUSE and, as a user other than root, FUSE 3's fusermount3 (package fuse3) or FUSE 2's fusermount; root needs neither:
sacdfs-0.6.3-x86_64-linux.tar.gz: generic x86-64, for any x86_64 CPU.sacdfs-0.6.3-x86_64-v3-linux.tar.gz: x86-64-v3 (AVX2, most Intel and AMD CPUs since about 2015): about 20 % less CPU for WAV files. It stops with an illegal-instruction error on older CPUs.sacdfs-0.6.3-aarch64-linux.tar.gz: aarch64 (64-bit ARM).
Container image for a NAS (SMB and NFS): ghcr.io/aclater/sacdfs:0.6.3 (generic CPU; amd64 and arm64).
Installer: install.sh finds the system (a NAS or Linux computer) and sets sacdfs up (install guide). Over SSH:
curl -fsSL https://github.com/aclater/sacdfs/releases/latest/download/install.sh \
| sudo sh -s -- --iso-dir /path/to/imagesVerify a download: sha256sum -c <asset>.sha256 and gh attestation verify <asset> -R aclater/sacdfs; the image: gh attestation verify oci://ghcr.io/aclater/sacdfs:0.6.3 -R aclater/sacdfs. See the README, "Install".
Built from 8414919.
sacdfs 0.6.3-rc.1
Security: this release is affected by GHSA-phfg-58h9-f2pp (High), GHSA-wfjg-pqhm-4mm2 (High) and GHSA-825f-7xrc-fqf7 (Medium). They are fixed in 0.6.4; use the latest release. Upgrade as the install guide describes.
If you installed with
install.sh, run the current one (releases/latest/download/install.sh), not this release's: this release's installer has the problem and puts it back.
Fixed
- Installer: on macOS, or any system other than Linux, it now stops at once and says sacdfs
runs on Linux only. Before, it reported "Linux detected", downloaded the Linux release and
stopped only when that binary failed to run, with a hint about glibc. It changed nothing then
either. Now it stops before any download, and points to playing the discs from a Linux NAS
or computer over the network. - Installer: a download cut short now fails with an error (#352). If the script piped into
shends before its last line, it stops with status 1 and says the download was cut short,
instead of exiting 0 having done nothing. The install guide now gives thewgetform as
download, then run (wget -O install.sh … && sh install.sh …), so a failed download stops
beforesh: piped intosh, an empty download still exits 0, which no script can detect. CI
checks both in the BusyBox and Fedora containers. - Installer:
--uninstallon QNAP switches "Run user defined processes" off again when the
installer switched it on (#353). The installer records that it did, and--uninstallswitches
it off only then, and only ifautorun.shruns nothing else; otherwise it says why it left it on. - Tag tests no longer race on the tags store. Tests only.
tags::installsets one store for the
whole process; intests/fallback.rsandtests/tags.rsthe test that sets it could overlap the
others, and one CI run on arm64 read a title from it on a disc that has none. The tests in each of
those files now run one at a time. - A player beside copies on a server too slow for them all no longer stalls longer than in
0.6.1 (#357). Behind a network client, copies reading at about a player's rate counted as
players, so the player lost its precedence and every reader read ahead a player's window,
which overflowed the cache: most frames were converted twice. A reader behind a network client
now keeps reading for throughput while an earlier player reads, until its client pauses, and a
player counts while it waits for a read. Over SMB, four copies beside a player on 2 slow cores:
worst read 3.9 s (0.6.1: 7.5 s, 0.6.2: 12.8 s), copies 12 % faster than 0.6.1 with 13 % less
CPU. Still stalls there (four copies at real time need more than those cores). Slower: a second
player starting beside four copies on 4 cores waits longer for its first reads than on 0.6.2
(worst 1.3 s against 0.8 s; 0.6.1 2.6 s). Over NFS, copies whose reads reach sacdfs in order
are not covered (performance).
Audio bytes are unchanged: only when frames are decoded ahead changes.
Documentation
- A Linux SMB mount for a player or for copies: add
rsize=262144(part of #266). Measured on
a NAS with a Linux 6.19 client: by default the client's 4 MiB reads and 8 MiB read-ahead made a
header-and-tags scan cost sacdfs 1.8 MB of audio per file; withrsize=262144262 KB and about
a seventh of the CPU, copies were faster (101 to 109 MB/s against 51 to 81) and player reads
unchanged.cache=nonecuts it to 99 KB, with a round trip per player read; the share setting
smb2 max read = 262144does whatrsizedoes for every Linux client. No default changed. New
scripts/tag-scan.pyrepeats the scan
(performance, deployment, Linux client and
Navidrome guides, clients).
Known issues
- A tag scan through a Linux SMB mount with default settings uses 2–10 % more CPU than on
0.6.2, a side effect of the #357 change (measured on a 4-CPU container, 222 files; copies and
player reads are equal or better). Withrsize=262144on the mount, as the docs now advise,
there is no difference, and the scan costs about a seventh of the CPU of a default mount.
Downloads
Linux binaries, glibc 2.17 or later (RHEL 7 era and later), needing FUSE and, as a user other than root, FUSE 3's fusermount3 (package fuse3) or FUSE 2's fusermount; root needs neither:
sacdfs-0.6.3-rc.1-x86_64-linux.tar.gz: generic x86-64, for any x86_64 CPU.sacdfs-0.6.3-rc.1-x86_64-v3-linux.tar.gz: x86-64-v3 (AVX2, most Intel and AMD CPUs since about 2015): about 20 % less CPU for WAV files. It stops with an illegal-instruction error on older CPUs.sacdfs-0.6.3-rc.1-aarch64-linux.tar.gz: aarch64 (64-bit ARM).
Container image for a NAS (SMB and NFS): ghcr.io/aclater/sacdfs:0.6.3-rc.1 (generic CPU; amd64 and arm64).
Installer: install.sh finds the system (a NAS or Linux computer) and sets sacdfs up (install guide). Over SSH:
curl -fsSL https://github.com/aclater/sacdfs/releases/latest/download/install.sh \
| sudo sh -s -- --iso-dir /path/to/imagesVerify a download: sha256sum -c <asset>.sha256 and gh attestation verify <asset> -R aclater/sacdfs; the image: gh attestation verify oci://ghcr.io/aclater/sacdfs:0.6.3-rc.1 -R aclater/sacdfs. See the README, "Install".
Built from 163686c.
sacdfs 0.6.2
Security: this release is affected by GHSA-phfg-58h9-f2pp (High), GHSA-wfjg-pqhm-4mm2 (High) and GHSA-825f-7xrc-fqf7 (Medium). They are fixed in 0.6.4; use the latest release. Upgrade as the install guide describes.
If you installed with
install.sh, run the current one (releases/latest/download/install.sh), not this release's: this release's installer has the problem and puts it back.
Tested as release candidate 0.6.2-rc.1; the results are in docs/test-plan.md ("Results: 0.6.2-rc.1
gate"). On top of rc.1, the release archive's Compose file now names the release's own image, and CI
runs in Red Hat-derived containers. Everything else is in the 0.6.2-rc.1 section below. Known: with
a CPU too slow to decode several copies at real time, a player beside four copies over SMB stalls
longer than on 0.6.1 (#357); it stalls on 0.6.1 too.
Changed
- CI builds and tests in Red Hat-derived containers. The jobs that build or test sacdfs (lint,
tests on x86-64 and arm64, MSRV, coverage, fuzzing, the weekly memcheck)
run in a UBI 10 container (Fedora 44 for the lint job), pinned by digest, as a normal user, with
the FUSE tests mounting as before. cargo-deny and zizmor are installed in a UBI container,
pinned by version and hash, instead of running their vendors' images (cargo-deny's is Alpine).
CI only; nothing shipped changes (development).
Fixed
- The release archive's Compose file names that release's image.
nas-compose.yamlin the
archive is now stamped with the release version, asinstall.shis. In 0.6.2-rc.1 it named
0.6.1, so the installer's container route ran 0.6.1 unless given--version.
Downloads
Linux binaries, glibc 2.17 or later (RHEL 7 era and later), needing FUSE and, as a user other than root, FUSE 3's fusermount3 (package fuse3) or FUSE 2's fusermount; root needs neither:
sacdfs-0.6.2-x86_64-linux.tar.gz: generic x86-64, for any x86_64 CPU.sacdfs-0.6.2-x86_64-v3-linux.tar.gz: x86-64-v3 (AVX2, most Intel and AMD CPUs since about 2015): about 20 % less CPU for WAV files. It stops with an illegal-instruction error on older CPUs.sacdfs-0.6.2-aarch64-linux.tar.gz: aarch64 (64-bit ARM).
Container image for a NAS (SMB and NFS): ghcr.io/aclater/sacdfs:0.6.2 (generic CPU; amd64 and arm64).
Installer: install.sh finds the system and sets sacdfs up (docs/install/README.md): curl -fsSL https://github.com/aclater/sacdfs/releases/latest/download/install.sh | sudo sh -s -- --iso-dir /path/to/images.
Verify a download: sha256sum -c <asset>.sha256 and gh attestation verify <asset> -R aclater/sacdfs; the image: gh attestation verify oci://ghcr.io/aclater/sacdfs:0.6.2 -R aclater/sacdfs. See the README, "Install".
Built from 08a2c6d.
sacdfs 0.6.2-rc.1
Security: this release is affected by GHSA-phfg-58h9-f2pp (High), GHSA-wfjg-pqhm-4mm2 (High) and GHSA-825f-7xrc-fqf7 (Medium). They are fixed in 0.6.4; use the latest release. Upgrade as the install guide describes.
If you installed with
install.sh, run the current one (releases/latest/download/install.sh), not this release's: this release's installer has the problem and puts it back.
Added
- Installer: one line sets sacdfs up (part of #340).
install.sh, plain POSIX sh that runs under
BusyBox, finds out what it runs on (Synology DSM, QNAP, ASUSTOR ADM, TerraMaster TOS, UGREEN,
ReadyNAS, OpenMediaVault, TrueNAS, Unraid or plain Linux) and does what that system's guide does:
the Compose file where Docker or Podman with Compose answers; otherwise the release binary with the
systemd unit, or withsacdfs-ctland the NAS's own boot hook (Task Scheduler,autorun.sh,
init.d). On TrueNAS and Unraid it only points to their app screens.
curl -fsSL https://github.com/aclater/sacdfs/releases/latest/download/install.sh | sudo sh -s -- --iso-dir /path.
It checks every download against the release's checksum, tries the binary before installing, never
writes under--iso-dir, changes nothing on a second run, and has--dry-runand--uninstall. Not
root, it stops before any change and prints the line to run withsudo. The whole script is in
functions called from its last line, so a download cut short runs nothing. Each release carries it
as an asset (checksum and attestation) and in the archives, withnas-compose.yaml. Checked through
its whole cycle on DSM 7.4.1, QuTScloud 5.2.9, ADM 5.1.4, TOS 7.0, ReadyNAS OS 6.10.10 and Debian 12 on
UGREEN's kernel in VMs, on Fedora 45 (systemd and Compose), and in BusyBox, Fedora 44 and UBI 7
containers; a new CI workflow (installer.yml) runs those containers and the runner's systemd and
Compose routes on every change todeploy/(install guide).
Changed
- A platform-neutral core, split from the Linux FUSE front end (part of #22). The first step of
native macOS and Windows support.docs/design.md("The core and its front ends") describes the
API the front ends call, what stays Linux-only, and the decisions the Windows and macOS front
ends build on (handles, threads, name rules, end-of-file reads). The code moved with no change in
behaviour: every request handler is now insacdfs::vfs::Volume, which answers through replies
the front end supplies, and the FUSE front end (src/fuse.rs, formerlysrc/fs.rs) only converts
fuser's types. A new test (tests/vfs.rs) reads every file of a DSD and a DST image through the
core without a mount and checks it against the reader. These changes touch the structure of the
read path, so the release that includes them needs the full gate and a soak.
Fixed
- WAV copies through the Linux kernel SMB client keep their pace again, and a player beside
starting copies stays first (#271). Samba serves the kernel client's 4 MiB reads concurrently,
so they reach sacdfs up to 56 MiB out of order; since 0.5.0 each of those restarted the reader's
pace, and a copy kept switching between reading for throughput and playing. A jump is now a read
more than 64 MiB away, as in the withdrawn #323. Unlike #323, a reader whose reads land out of
order (behind a network client, whose own read-ahead keeps a player 8 to 16 MiB ahead) reads for
throughput only once it is 32 MiB ahead of playing time, and until then lets a player that was
there before it go first. Beside a 6-channel player at real time, four copies starting together no
longer stall it (0 of 16 runs over 100 ms, worst 14.8 ms, against 4 of 16 and up to 1.4 s with
#323; 0.6.0: 0 of 8, 15.0 ms), and finish 4 % (4 cores) to 11 % (2 cores) sooner than on 0.6.0.
Slower: one copy beside a player takes 5 to 8 % longer than on 0.6.0, which counted that copy
as a player most of the time; it now gets the player-first rules of #230 throughout. One reader
copying a whole area over SMB is as fast as 0.4.1 again: stereo WAV 22.86 s (0.6.0 25.22 s, 0.4.1
23.06 s), six channels 73.84 s (81.41 s, 75.60 s). Local reads take no new rule (and were not
affected by #323's). The change decides only when frames are decoded ahead, not what they hold.
Needs a soak in the next release
(performance). sacdfs-ctl startcould report a start that worked as a failure. It looked at the new process's
name right after starting it; beforesetsidhad started sacdfs, the name was still the shell's, so a
slow first start (seen right after the binary was replaced) gave "sacdfs did not mount", removed the
pid file, and left sacdfs running untracked. It now waits on the process itself. It also uses
kill -s TERMandkill -s KILL, the POSIX spelling.- An unmount from outside no longer ends sacdfs with status 1 (#322). When the mount was
unmounted or detached from outside (fusermount3 -u,fusermount3 -u -z,umount -l) while a
request was still queued, typically the release of a file just closed, the kernel could end the
connection as aborted instead of unmounted, and sacdfs exited with "Software caused connection
abort (os error 103)" and status 1: about one detach in ten when the last open file was closed
right after it. Under load a plain unmount right after a close can meet the same race, the
likely cause of the restart test's rare failure. sacdfs now takes that error as the end of the
connection, as it takes the usual one (before, only after its own detach on a signal): the
kernel returns it for no other reason, as sacdfs does not ask forFUSE_ABORT_ERROR. A new
integration test detaches 40 mounts from outside with a file open; without the fix it fails
within the first five. The restart test now prints the new sacdfs's log when it fails. - A stop signal right after the start is no longer lost (#334). Lifecycle change. A SIGTERM,
SIGINT or SIGHUP that arrived in the microseconds while sacdfs set up its signal handlers was
dropped by the signal library: sacdfs neither died nor stopped, mounted, and ran on until
killed (a service stopped right after its start). The handlers now go in with those signals
blocked, before any thread exists, so such a signal waits and stops sacdfs without mounting.
They also go in earlier, right after the fork into the background instead of just before the
mount: a signal in between now ends sacdfs with status 0 and "stopped by signal before
mounting" (the background parent still reports it as a failed start) rather than killing it.
A new test (tests/stop_signals.rs, no libtest harness) sends SIGTERM into the registration
2000 times; about 2 % were lost before. The mount test that found it now prints sacdfs's log
and what each of its threads waits on when sacdfs does not stop.
Downloads
Linux binaries, glibc 2.17 or later (RHEL 7 era and later), needing FUSE and, as a user other than root, FUSE 3's fusermount3 (package fuse3) or FUSE 2's fusermount; root needs neither:
sacdfs-0.6.2-rc.1-x86_64-linux.tar.gz: generic x86-64, for any x86_64 CPU.sacdfs-0.6.2-rc.1-x86_64-v3-linux.tar.gz: x86-64-v3 (AVX2, most Intel and AMD CPUs since about 2015): about 20 % less CPU for WAV files. It stops with an illegal-instruction error on older CPUs.sacdfs-0.6.2-rc.1-aarch64-linux.tar.gz: aarch64 (64-bit ARM).
Container image for a NAS (SMB and NFS): ghcr.io/aclater/sacdfs:0.6.2-rc.1 (generic CPU; amd64 and arm64).
Installer: install.sh finds the system and sets sacdfs up (docs/install/README.md): curl -fsSL https://github.com/aclater/sacdfs/releases/latest/download/install.sh | sudo sh -s -- --iso-dir /path/to/images.
Verify a download: sha256sum -c <asset>.sha256 and gh attestation verify <asset> -R aclater/sacdfs; the image: gh attestation verify oci://ghcr.io/aclater/sacdfs:0.6.2-rc.1 -R aclater/sacdfs. See the README, "Install".
Built from 757ab06.
sacdfs 0.6.1
Security: this release is affected by GHSA-phfg-58h9-f2pp (High), GHSA-wfjg-pqhm-4mm2 (High) and GHSA-825f-7xrc-fqf7 (Medium). They are fixed in 0.6.4; use the latest release. Upgrade as the install guide describes.
Tested on main against 0.6.0; the results are in docs/test-plan.md ("Results: 0.6.1 gate").
The pace change for WAV copies over SMB (#271, #323) was taken out before the release because it
could stall a player beside four starting copies (#338); 0.6.1 keeps 0.6.0's pace rules.
Changed
- SMB directory leases stay off: why they can't be fresh (part of #325). Docs and a comment in the
container's entry point only; no code or default changed, so this needs no soak. Samba (4.24.7) breaks a
directory lease only for a change made through Samba, never for a kernel change notification, and a FUSE
file system can't raise one for an entry it adds; so neither FUSE invalidations from sacdfs nor
kernel change notify = yeskeep a leased listing fresh. Measured on 0.6.0: with leases on, a new image
didn't show within 155 s on the Linux kernel client (also withkernel change notify = yes, and also on
a plain folder shared by the same Samba) nor on Windows 11 listing every second; with leases off it showed
in 0.4 to 1.7 s on Linux and 5.1 s on Windows.docs/deployment.mdnow says that turning leases on for
Windows makes Windows miss new images too, and that a Linux kernel client mounted withnohandlecache
stays fresh with leases on (1.6 s). - Windows thumbnail pass: where the extra folder opens come from (part of #312). Docs only. On
Windows 11 in a VM, the pass over 196 files sends 2,104 to 2,166 requests on 0.6.0 against 508 to 514
on 0.5.0 and 921 to 941 on 0.4.1; the 800 extra areOPENDIRandRELEASEDIRpairs. They come from
the container'ssmb3 directory leases = no(#205), not from theopendirchange: with leases on,
0.6.0 counts 496. On a 5,040-file tree the pass took about 1.5 times as long (121 to 126 s against 80 to
86 s).docs/deployment.mdsays so, and thatSMB_EXTRA_GLOBAL=smb3 directory leases = yesturns the
leases back on for a Windows-only share, at the cost the container's default avoids for the Linux kernel
client. No code or default changed. - DST areas read faster on one thread: scarletbook is built in one codegen unit (#270). Release
builds now compile scarletbook withcodegen-units = 1. Its DST decoder then runs in as many cycles
as in the library's own builds, about 4 % fewer than before. Six-channel DST on one thread went from
1.11-1.15× to 0.99× of the library on a NAS-class CPU, which meets the 1.1× target.
Audio bytes are unchanged. A clean release build takes about 10 s longer. Details are in
docs/performance.md, "One codegen unit for scarletbook". - sacdfs exits after a grace period on SIGTERM even while clients hold files open (#276). This
changes the lifecycle next to the read path, not the reads, and the next release needs a soak.
The mount is still detached at once; sacdfs now serves the files still open for--stop-grace
seconds (default 10, container:SACDFS_STOP_GRACE,0for the old wait-for-ever) and then
saves its state and exits with status 0. Clients with a file still open get I/O errors. It exits
only when its mount is gone from the mount table, so a stop never leavesTransport endpoint is not connected; if the detach fails it logs an error and keeps serving. The stop timeouts stay
above the grace: the container'sstop_grace_periodand the Unraid--stop-timeoutgo from
20 s to 30 s,TimeoutStopSecandsacdfs-ctlstay at 30 s (sacdfs-ctlwaitsgrace + 15s
for a larger--stop-grace). Tests: an integration test with a held-open file, and
deploy/ci-held-stop-test.shin the container share test (files held open over SMB and NFS and
inside the container). - Install guides rewritten to be simpler (part of #310).
docs/install/README.mdis now a chooser
with three questions; the Docker and Podman, Linux binary and Linux client guides follow one shape
(what you need, numbered steps, check it worked, three common problems, Details), with the test
history moved to Details. The NAS and player guides follow in later pull requests. - Player install guides rewritten (part of #310). Roon, Plex, Jellyfin, Navidrome, Lyrion and MPD now follow the common guide shape (what you need, numbered steps, check, three problems, Details), start from the Docker/Podman and Linux client guides, and were run on 0.6.0 for Jellyfin, Navidrome, Lyrion and MPD. Lyrion no longer needs the root settings with the SMB mount, and MPD needs no cover renaming.
- NAS install guides rewritten (part of #310). Each vendor guide (Synology, QNAP, ASUSTOR, TerraMaster, UGREEN, ReadyNAS) is one page with one path, the 0.6.0 release archive and the vendor's own start-at-boot method, followed as written on 0.6.0 in the NAS VMs; TrueNAS SCALE, Unraid and OpenMediaVault keep their facts in the same layout, marked untested on 0.6.0.
nas-binary.mdis the short generic guide; test history, options and the longer troubleshooting moved to each page's Details. Found while checking: TOS has no/usr/local/bin, so the unit steps make it.
Fixed
- Mount tests: a failing stop test no longer hangs the run (#335). Tests only. The two tests
that stop a daemon (SIGSTOP) now hold it in a guard that, when the test fails, resumes it
(SIGCONT), stops and reaps it, and prints its debug log: a failed assertion no longer leaves
a stopped daemon that blocks the rest of the run. The cause of those failures under load was
the #305 race through tests outside the lock:bad_arguments_fail_cleanly,
start_up_messages_name_the_fixanda_source_or_cache_under_the_mount_point_is_refused
need no FUSE but spawn sacdfs, so they did not take the shared lock, and a child they forked
could inherit a stop test's file on the mount it was about to stop. The child then blocked in
exec, the mount stayed busy, and the stop test failed to unmount it. They now hold the lock
shared (common::spawns_children!). - Mount tests no longer hang under load (#305). A child forked by one test holds a copy of the
open files of the other tests' threads until it execs, and closing a file on a FUSE mount waits
for its daemon. While a test held a daemon stopped (SIGSTOP), such a child, and the test that
spawned it, blocked inexec. The two tests that stop a daemon now run exclusive: every FUSE
test holds a shared lock for its whole run (common::skip_without_fuse!), and they hold it
exclusive (skip_without_fuse!(exclusive)). The folder-times test also waits for each folder to
move, not only the root, which failed one run in four under load. - Real-disc tests: a truncated image fails (#294). Before the tests run, each image is checked
against its own TOC (every area's TOC copies and audio must lie inside the file); a copy cut
short fails the run with "truncated: ...". A test on truncated copies of a corpus image runs in
CI. A corrupt image of the right size is still not caught. - Soak: designed art-churn errors no longer fail the verdict (#279).
ESTALEon acover-N.jpg
the churner replaced or removed, andENOENTorESTALEon an area'scover.jpgin the
art-churn album after a cover removal, are counted as replaced sidecars. Every other errno, and
these on any other path, still fail the soak;scripts/test_soak.pypins both. - macOS: precomposed (NFC) lookups of accented names (part of #308). Docs only. Names are
served precomposed and macOS lists them decomposed; the macOS SMB client answers a precomposed
lookup itself, without asking the server, once it has cached the folder's listing. All 111 names
with a differing NFC and NFD spelling failed that way, not 111 of 357 at random. The setting
dir_cache_off=yesinnsmb.conffixed it on macOS 27.2 in a VM.docs/deployment.md(macOS)
says so. No code or Samba setting changed. Checked on macOS in a VM, not on Apple hardware.
Documentation
- What a player's import costs over SMB, and the settings that change it (part of #266).
Through a Linux kernel SMB mount, Navidrome's scan made sacdfs serve 3.5 times the bytes, and
use 5 times the CPU, of a direct mount. The cause is the client reading ahead around each
file's header and tags.cache=noneon the client mount, oroplocks = noon the share,
brings it down to the direct mount's cost;smb2 max read = 262144halves it. Copies and
playback were measured with each. The defaults are unchanged
(performance, Navidrome guide).
Downloads
Linux binaries, glibc 2.17 or later (RHEL 7 era and later), needing FUSE and, as a user other than root, FUSE 3's fusermount3 (package fuse3) or FUSE 2's fusermount; root needs neither:
sacdfs-0.6.1-x86_64-linux.tar.gz: generic x86-64, for any x86_64 CPU.sacdfs-0.6.1-x86_64-v3-linux.tar.gz: x86-64-v3 (AVX2, most Intel and AMD CPUs since about 2015): about 20 % less CPU for WAV files. It stops with an illegal-instruction error on older CPUs.sacdfs-0.6.1-aarch64-linux.tar.gz: aarch64 (64-bit ARM).
Container image for a NAS (SMB and NFS): ghcr.io/aclater/sacdfs:0.6.1 (generic CPU; amd64 and arm64).
Verify a download: sha256sum -c <asset>.sha256 and gh attestation verify <asset> -R aclater/sacdfs; the image: gh attestation verify oci://ghcr.io/aclater/sacdfs:0.6.1 -R aclater/sacdfs. See the README, "Install".
Built from 03c2026...
sacdfs 0.6.0
Security: this release is affected by GHSA-phfg-58h9-f2pp (High), GHSA-wfjg-pqhm-4mm2 (High) and GHSA-825f-7xrc-fqf7 (Medium). They are fixed in 0.6.4; use the latest release. Upgrade as the install guide describes.
Tested as release candidate 0.6.0-rc.1 (2026-11-10). The final version carries only the version
change on top of the rc.1 code. The results are in docs/clients.md and docs/test-plan.md.
Added
-
Soak: the memory steps explained, and judged without them (part of #247). The steps of 3 to
9 MiB in a soak's resident size are the kernel swapping pages out and in on a busy host: four
parallel runs fell by the same amounts at the same samples while the malloc heap did not change.
scripts/soak.pynow recordsVmSwap(swap_kb) and judges resident plus swapped memory,
and has--smapsand--malloc-info-everyto trace a step to its mapping. -
The release archives also ship
sacdfs-ctl,sacdfs.serviceandsacdfs.conf.example
(part of #289), next to the binary, so a NAS install doesn't need a second download. -
Safe start and stop for the release binary on a NAS (#289).
deploy/systemd/sacdfs.service
is a unit for/usr/local/bin/sacdfs(settings in/etc/default/sacdfs, loads FUSE, waits up to
30 s for the unmount, clears a leftover mount);deploy/sacdfs-ctldoesstart,stop,status
andrestartwith BusyBoxshwhere there is no systemd. Stopping is SIGTERM, and SIGKILL only
after the wait. The NAS steps warn againstkill -9, which leaves a dead mount, and the
TerraMaster guide uses the shipped unit. CI runsshellcheckon the script and
systemd-analyze verifyon the unit. The NAS steps now say how to start it at boot with the
vendor's own tools: an ASUSTORinit.dscript, a QNAPautorun.shand a Synology Task Scheduler
task, checked in VMs. The QNAPautorun.shstarted sacdfs by itself
after a restart of the QuTScloud VM, on a hand-made volume. -
Unraid and TrueNAS app-store entries, prepared (part of #284).
deploy/unraid/sacdfs.xmlis a
Community Applications template (device, capabilities and AppArmor setting in Extra Parameters,
not privileged; ports 10445 and 12049; read-only images and index cache paths; SMB and NFS
settings), loadable today from Unraid's user templates.deploy/truenas/sacdfs/is a TrueNAS
apps-catalog app definition (form, template, test values) that renders with the catalog's tooling.
Both were started with rootful Podman and served the SMB share; neither was run on Unraid or
TrueNAS, and neither is submitted to a store. Seedocs/install/store-apps.md.
The store icon is now an original drawing (deploy/icon.svg,deploy/icon.png), and
ca_profile.xml(the Community Applications repository profile) is in the repository root.
The catalog app passes the TrueNAS apps repository's own render, metadata, port and validate
checks (part of #284). -
One Compose file for every NAS container manager (#283).
deploy/nas-compose.yamlruns the
published image with ports 10445 and 12049 (a NAS's own SMB and NFS usually hold 445 and 2049),
the health check and a commented setting list. The one line to edit is the path of your images.
The Synology, QNAP, UGREEN, TerraMaster, Unraid, OpenMediaVault and ASUSTOR guides point to it
instead of repeating Compose blocks, and each says where it was actually run: Docker Compose on
Podman, Docker 25 with Compose 2.24 on ASUSTOR ADM in a VM, and Docker 29 with Compose 5.1 from
the App Center on TerraMaster TOS in a VM, and Docker 27 with Compose 5.1 from Container Station
on QuTScloud in a VM (on a hand-made volume: QuTScloud wouldn't make a pool); the other vendors'
container managers are untested. -
Plain start-up messages for NAS users (#282). Before scanning or mounting, sacdfs checks
what the mount needs and says in one line what is missing and what to type: FUSE not loaded
(modprobe fuse, instead ofcannot mount on X: No such file or directory), the folder to
mount on missing (mkdir -p X), no mount helper for a user other than root, no right to mount
in a container (--cap-add SYS_ADMIN). A mount refused by AppArmor or SELinux, and
--allow-otherwithoutuser_allow_other, are explained the same way. The container's entry
point checks/dev/fuse,CAP_SYS_ADMIN, ports 445 and 2049 already taken, and the/srvbind
for players on the same host before it starts. -
The RF64 test is read by other programs, and the file-handle tests also run on arm64 (#220).
wav::rf64_fileused to be ignored and asserted only the file's size. It now runs in every test
job: it writes the 4.6 GB sparse file and hasscripts/check_rf64.py(Python, shares no code with
sacdfs) check the chunk layout and sizes, andffprobewhen installed check the channels, rate,
duration and tags. A newfile handles (arm64)job in the NFS workflow runs theopen_by_handle_at
tests as root on a native arm64 runner, where the normal arm64 job skips them. Tests and CI only; no
change to the program.
Changed
- Listings over SMB cost sacdfs a quarter fewer requests (#205). The kernel now keeps the
listing of every folder until it changes, including areas that are re-read for their art, which
it listed again at every open before. The container's Samba no longer asks sacdfs for the
filesystem's size at each folder (vfs_default:VFS_OPEN_HOW_RESOLVE_NO_XDEV = noon the share),
and no longer offers directory leases (smb3 directory leases = no), with which the Linux kernel
client did not show a new image while it kept listing that folder. Through the container and the
Linux kernel client, a repeatedls -lRof 2,484 entries costs 2,010 requests instead of 2,776,
and a library rescan 1,582 instead of 2,345. Names, sizes, times and bytes are unchanged. See
docs/deployment.md ("What a listing costs over SMB"), with the settings for your own Samba. - Root's mount lets other users read it (#282). Run as root, sacdfs now mounts with
allow_otherby default, so Samba guests, NFS clients and apps can read the mount instead of
getting "Permission denied".--no-allow-other(-o no_allow_other) keeps it to root. Users
other than root are unchanged. The mount is read-only; the decision and its trade-off are in
docs/design.md. - Unmount hints name a command that exists (#282):
umountas root (DSM, ADM and QuTScloud
have nofusermount3), elsefusermount3 -uorfusermount -u, whichever is installed.
Fixed
- Two
sacdfs-ctl startat once started two sacdfs (#289). ASUSTOR ADM runs its start script
twice at boot; both starts passed the "already running" check and stacked two mounts, one of
whichstopnever reached. A lock now serialisesstartandstop. - A background mount that failed printed its error twice (#282).
- The container could take 10 s to stop (#283). A stop within a few seconds of the start (a
restart right afterup, say) sometimes waited 10 s for the servers to release the mount,
becausefuser -mlisted every process on the container's filesystem once the mount was gone.
Podman's default stop timeout then killed the container ("resorting to SIGKILL"). The entry point
now waits only while the mount exists: 0 of 75 stops took over a second, against 14 of 75 before (Podman on Fedora and Docker on ADM).
Documentation
-
A player beside whole-file scans, checked on a NAS's own CPU (#230). The 0.5.0 image as an app on
a NAS's E-cores, client on another host over SMB and NFS: a 6-channel WAV player held real time with
one or four whole-file scanners on 2 and 4 cores (0.4.1: 0.35 to 0.65× real time), and players that
buffer 8 to 64 MiB ahead weren't hurt by counting as scanners while they settle. One run in 7 with
four scanners on 2 cores fell to 0.88× while the NAS was busy. Tables indocs/deployment.mdand
docs/performance.md. -
#185 (MusicBrainz lookup) is complete except fingerprinting. Fallback naming,
tags fetchand
--musicbrainzare shipped; the AcoustID fingerprint lookup, which needs an API key and a
Chromaprint dependency decision, moved to #296 (Later). The user guide and design notes say so. -
macOS and SMB restarts (part of #204): the install guides now say, where a user updates, that a
macOS client must mount the SMB share again after an upgrade or restart, while Linux and Windows
clients carry on.docs/deployment.mdrecords that the SMB server identity follows the container's
host name, how to keep it across a recreate (hostname: sacdfsornetbios name, checked with a
raw SMB2 negotiate and the gvfs client), and why that is not the default. macOS was not re-tested. -
NAS install guides for people new to SSH and Linux: a new
docs/install/nas-binary.mdruns the
release binary on a NAS one command at a time, each with what you should see, and a table of the
messages met when something goes wrong and what to do. It keeps sacdfs on the data volume, adds
--allow-other(without it, SMB guests and other users got "Permission denied"), and usescp,
notinstall, which on TerraMaster TOS is a different program. The Synology, QNAP, ASUSTOR,
TerraMaster and UGREEN guides were rewritten around it and run again as written with the 0.5.0
release, in VMs, not on the vendors' hardware. ASUSTOR gains a tested Compose file (other ports, the
health check written in); TerraMaster a systemd service that mounted again after a restart, and a
warning that its NFS service empties/etc/exports. Why Synology's ...