Repository navigation
A release of OOS, an open operating system for the Rigol DHO800/DHO900. It brings fixes from the review of 0.1.13 and from the forum's and the issues' reports, Bluetooth and a WLAN stick, mice and keyboards, a number pad for values, and the DHO802/DHO812 as verified models. Telnet, the native Windows build and the frames draw path are gone.
Experimental. Read the disclaimer and the status first. Measurements are not calibrated to any standard.
Installing and updating
No build needed: Installation.
- USB stick (
DHO800_DHO900_OOS_Install.GEL, Storage > Upgrade in the stock system): installs this version directly, also over an older OOS on the card. The DHO802 and DHO812 no longer needoos-any-model.txt. - From 0.1.13 or older (0.1.7 and newer), over the network: this package is larger (47 MB) than the restart into a new build of 0.1.13 and older can carry.
- With
http://<scope>:8080/update, that restart is expected to stop in a kernel panic. Switch the scope off at the mains: the build you had starts again, and nothing is written. This was seen on a DHO924S on 2026-11-18 with an olderoos-updateand a package of this size. The fix is in 0.1.14 itself, so it helps only from the next update on. - Instead, use a shell on the scope (telnet or SSH in 0.1.13): copy the package to
/tmpand runoos-update install /tmp/oos-0.1.14.oosupd --no-kexec --execute. That writes the image directly, without the restart. Then switch the scope off at the mains, wait 30 seconds and switch it on: that start is the trial, and it confirms itself about 20 s after the start-up. This way has not been tried from 0.1.13 on a unit; it is the path 0.1.13 already has for packages without a kexec trial. - Or use the USB stick above.
- With
- From 0.1.6 or older: install 0.1.7 first.
- Going back: Updates and Going Back.
What ran on a unit
Before the release, on the developer's DHO924S on 2026-11-19, a build of this commit was installed with oos-update install by its kexec start. The update tool used was the one of this build.
*IDN?answeredoos-ui 0.1.14.tools/device-test.py --tier a0: 21 passed, 0 failed, 1 skipped (it needs a plan file). This covers SSH and no telnet, SCPI, the web page, the frame rate, the generator and the protected areas.- The draw path ran at 29 screens a second at 10 µs/div and 1 ms/div, with 1 and 4 channels. At 1 ms/div about every record reached the picture (#220).
- Not tried before the release: the first start from power-on of this build, which confirms an update, because nobody could switch the scope off and on overnight. An earlier build with a package of over 43 MB passed that start on the same unit on 2026-11-18 (the boot test of #177).
Sources of every third-party part in the image are attached (THIRD-PARTY-LICENSES.md), and SHA256SUMS lists every file.
What changed
Released on 2026-11-19. Fixes from the review of 0.1.13 and from the forum's and the issues' reports, Bluetooth and a
WLAN stick with its driver, mice and keyboards, a number pad for values, and the DHO802/DHO812 verified by a user's
install. Telnet, the native Windows build and the frames draw path are gone. The code of oos-ui is split into crates and
tables of pages and services without a change of what the scope does. All of it is tested on the host; marked "(unit)"
where it also ran on a DHO924S. The release build itself ran on a DHO924S from its kexec start (the device tests of tier
a0); its first start from power-on, which confirms an update, was not tried before the release.
Added:
- A number pad for the values of a channel's scale and offset, the time base and its position and the trigger level
(#22): a tap on the value opens it, a key with the unit (mV,V,µs...) sets it, as:CHANnel<n>:SCALeand
the other remote commands do (First Steps). - USB and Bluetooth mice and keyboards (First Steps): the mouse is the
finger, with a pointer while it is in use that moves in even steps; the keys stand in for the front panel and type
into the on-screen keyboard. oos-ui takes them for itself, so no key reaches the console shell behind the screen
(unit, with a BLE keyboard and touchpad). - Bluetooth through a USB stick (bluetooth.md, ADR 0057): a TP-Link Archer TX10UB Nano is
switched out of its driver-CD mode, its Realtek 8851BU gets its firmware, and BlueZ runs on a system bus of its own;
pairings are kept on the data partition (unit: a BLE keyboard with touchpad paired, used, and reconnected after a
power cycle). - System > Network > Bluetooth (bluetooth.md): find, pair, connect, disconnect and
forget mice and keyboards; a keyboard's passkey is shown on the scope. While a Bluetooth stick is there, aBTbox
in the top bar beside the WLAN box: green while a device is connected, a tap opens the page (unit). - WLAN on the TP-Link Archer TX10UB Nano (Realtek RTL8851BU, 2.4 and 5 GHz; wlan.md, ADR 0027):
rtw89of Linux 6.18.63 is built as modules beside the 6.12 kernel, andoos-wlanpicks it by the stick's id
(unit: a saved WPA2 network joined on 5 GHz, DHCP, about 84 Mbit/s sent). tools/plan-matrix.shalso covers averaging, the logic analyser, roll and scan, the generator, LA thresholds, the
cold-boot dry run and the PLL table of another clock.oos-fpgad made-up-cal DIRwrites a made-up calibration set: every calibration file the register sequences read,
with invented values that mean nothing for a scope.tools/compare-base.shuses it when there are no files of a
unit, so CI now also compares the register values that calibration changes.tools/device-deploy.shchecks every uploaded file with SHA-256 (unit);device-deploy-ssh.shis gone.tools/check.shhas asimlayer (not enforced) for the Docker and QEMU simulations.tools/device-test.py --changed BASEruns only the device tests a change reaches (documents and CI: the six smoke
tests). A failed test runs again (--retries); one that passes then is reported as FLAKY. A failure that a known
run of main has too (--known,--save-known) is reported as KNOWN, except a differing register sequence and a
failed smoke test. Neither FLAKY nor KNOWN fails the run. The register sequences are compared section by section,
without the differences of the setup (among them the last empty line the unit's shell drops)
(device-testing.md).- CI: every commit of a pull request needs a Signed-off-by line (ADR 0026); CI runs
tools/check.shitself. - System > Network > Remote services > Allowed sets
net.allowfrom the scope's screen (unit). The picker offers
every network the scope is on, and Other opens the keyboard. A wrong entry is refused with a message and not
stored. A change takes effect at once, without a restart:- the firewall is reloaded;
- SCPI, VXI-11 and HiSLIP close the connections the new list leaves out, and the web remote checks every request;
- SSH starts or stops with the list.
The Firewall row shows the state in force (host-tested; the earlier build of the row ran on a unit).
- System > Start OOS again, right under Boot mode: two taps within 5 s select OOS for the next power-on once more,
asoos-boot-oos-againdoes (it runs the sameoos-update arm), and a dialog shows the result. With Boot mode
Persist it only says that OOS starts anyway. One-shot users no longer need a shell for this step (host-tested; not
yet run on a unit). - docs/http.md, the reference of the web remote's HTTP routes and the MCP tools
scpiand
screenshot(issue #205): method, query, body, status codes, content type, limits andnet.allowof each, and
the mark of ADR 0055 (proposed): stable, experimental or internal. The behaviour is unchanged.
tools/http-coverage.py, in the layerratchetoftools/check.sh, fails when a route or tool has no entry or
mark there, or an entry names a route the code does not answer.
Removed:
- Telnet. The scope no longer starts a shell without password on port 23, and System > Network > Remote services has
no telnet switch. SSH is the only shell on the network: switch it on under Remote services, set Allowed to your
network, and log in with the password from SSH settings > Show password, then add a key withssh-copy-id
(Remote Control, docs/ssh.md). The serial console keeps its shell.
remote.telnetin an older/oos/system.confis ignored and no longer written.tools/device-deploy.shand
tools/device-test.pyreach the scope over SSH only (OOS_SHELLis gone); a bench unit needs SSH and a key once
(docs/device-testing.md). ADR 0054 (host-tested; not yet run on a unit). - The native Windows build:
build-windows.ps1,run-windows.ps1and the desktop window with its keyboard and mouse
shortcuts and its file dialog. On a computer OOS runs withoos-ui --serveand the web remote in a browser, on
Linux, macOS and Windows inside WSL 2 (ADR 0033). The scope is not affected. - The frames draw path (Standard) and its switch: Acquire > Draw path is gone, the scope always draws with Batch,
which puts every acquired frame into the picture and now holds the screen rate (issue #186, ADR 0059).
acq.pipelinein an older/oos/system.confis ignored and no longer written,OOS_PIPELINEis not read, and
draw_pathin an older setup is passed over.:SYSTem:PIPeline?answersBATC,:SYSTem:PIPeline BATChis
taken and changes nothing,FRAMesandSTANdardare refused with -224. The status line sayspipeline Off
instead ofpipeline Frameswith a source other than the hardware, and System info no longer names the path
(unit: starts in Batch,:SYSTem:PIPelineas above, about 29 pictures a second; the Acquire menu was checked on
the host only).
Changed:
- The DHO802 and DHO812 count as verified (models.md 2.1): a user installed OOS 0.1.13 on a DHO802
from the USB stick and it started, answered SCPI and the web page (issue #19). The first install from the stock
system now takes both models withoutoos-any-model.txt. - Batch draw path: at 10 us/div the screen shows about 29 pictures a second with one,
two or four channels, where it showed 14 to 19 (unit). The workers fold their maps once a screen on the little
cores, and the frame loop swaps them in. Frames that are read and not drawn now count as dropped in:SYSTem:STAT?. - System > Overclock is only in the menu with
adc.overclock = onin/oos/system.conf, or whilelimits.confalready
sets another clock: above 1.25 GHz every step boots, but records at the full rate lose samples (unit). Every step
above 1.25 GHz is marked "(test)". - The waveform is drawn with the Batch path (the only one, see Removed): from 100 ns to 500 us/div it draws 1.2 to
3.4 times as many of the acquired frames, 5269 instead of 1563 a second with one channel at 100 ns/div (unit). - system.conf: of two lines with the same key the last one wins, and
#starts a comment only at the start of a line
or after a space or tab, asoos-confreads it. trig.serial_hw,acq.frame0,zone.batchandmask.map.*are also taken from the image defaults
(/etc/oos/system.conf).- The computer's mDNS announcer refuses queries with forward name pointers, as the scope does.
oos-fpgad: a wrong FPGA version names the value read inroll,selfcalandafg; a bad number names its value.
Theroll [--scan]dry run also lists the fill pulse, sweep restart and stop writes.- Image build: a new version in
rootfs/deps.lockor a changed build script rebuilds wpa_supplicant, e2fsck, kexec,
the GPU helper and Mesa by itself. Mesa comes from archive.mesa3d.org (same SHA-256), no Ubuntu source package is
needed. Build images are made again when their Dockerfile changes. - Build: the Rust release (
rust-toolchain.toml, 1.100.0) and the Docker base images (by digest, in
tools/lib/docker.sh) are pinned.rustup toolchain installin the tree installs that release, with the target for
the scope; CI and the image build use the same one (ADR 0045). - Card tools:
card-install.py,switch.py,sdcard-layout.pyand the image'soos-oneshotcheck the same stock
partition-table marks, now includingrecovery. - Docs: overclocking above 1.25 GHz loses about 5 % of the samples at the full rate (unit), so frequencies read too
high; 1.25 GHz stays (docs/models.md 4.1). Unit profiles forOOS_UNIT=<name>are described in docs/models.md 4.2;
rootfs/unit/is private and ignored by git. - SCPI setup and hold (
SETUP,HOLD,SETHOLD,FSETUP,FHOLDin:MEASure:ITEM?) take the data first and the
clock second, as the tiles say (A = data, B = clock). Before, they took the clock first: a script that sends the
clock first gets swapped results now. These are OOS items; the stock programming guide has none. - A change of trigger type keeps the user's pulse width, pattern, levels and limits on the hardware, Live and Replay.
The demo's starting values are set on the demo only. Slope, Runt and Window keep the lower level at or below the
upper one;:TRIGger:RUNT:UPPer/LOWerrefuses a threshold past the other with -224. - Choosing a decoder protocol on Live or a Replay keeps the sources, as on the hardware. Before, they moved to the
demo's channels. - Touch lock has one place on the screen: the Touch lock action in the start menu, next to the other front-panel
keys (and the TOUCH key itself). Utility > Touch, the same switch a second time, is gone.
Fixed:
- A remote client (SCPI on port 5555, VXI-11, HiSLIP) that reads its answers slowly or not at all no longer slows
the screen: answers go out as the network takes them, and the client's next command waits until its answers are
out. Before, such a client could hold the screen and the keys for up to 30 s per answer (#192; host tests). - A
cal_vertical.hexwith 64-byte records (one block group, seen on a DHO804) is read at 64 bytes a record instead
of stopping oos-ui at every start; a calibration file whose scale keys are out of order is not used, and the scope
starts uncalibrated instead of stopping (issue #37; host tests, and the usual 56-byte file still loads on a DHO924S). - The WLAN box of the top bar is as wide as the LAN box at every text size: at the normal size its address ran out
of it (#23). - SSC Freq Dev and SSC Modulation Rate are measured on the modulation (issue #196). The deviation was the span of
single cycles, mostly their timing error (about 9,400 ppm for a 5,000 ppm test clock, now about 5,500); it is now
the span of the rate averaged over 1/32 of the modulation period, which reads a sine 0.2 % and a triangle 3 % under
its span. The rate was the nearest spectral bin; it is now interpolated between bins, and with fewer than four
modulation periods in the record fitted with a sine (33 kHz with three periods read 29.3 kHz, now 33.0 kHz). - Double pulse: Ipeak and Irrm are the turn-on peak, not the end of the inductor's ramp later in the second pulse, so
Trr, Qrr and the diode's d/dt have values. The descriptions of Td(off), Tf and Toff name the current fall they
measure, not a voltage rise. - Measurements without an SI unit take no SI prefix: a ratio reads
0.498, not498 m, a count1, not1.00, and
ppm,rpmandrpm/sare plain decimals (1500 rpm, not1.50 krpm). - The trigger level marker stays when the source channel is on AC coupling (#46): the scope couples ahead of
the ADC, so the level is a level on the AC trace as drawn. Only the demo, which triggers before it takes the mean
out, still hides it (on the unit: assumed, not yet checked). - Eye Width no longer reads short by the picture's own spread (issue #195): a crossing without jitter lights two
columns of the eye and read half a column of jitter, 6 % of the unit interval in a 96-column eye. That spread is
now taken out, so a signal without jitter reads 1 UI and one with 20 ps RJ on 1 ns reads 880 ps. - The PLL clock recovery of the eye in
oos-dspends for every loop bandwidth (issue #194): one near the edge rate
drove the recovered period to zero or below, and the clock's ticks were added until memory ran out. The scope itself
only asks for the default bandwidth (rate / 1667), which never got there. - The G key lights up while the generator's output is on, and the LA key while the logic analyser is on; both stayed
dark before (contributed by halma1). - DCC decodes the signal of a live track (contributed by halma1, tested on a layout with an ESU ECoS): a phase slip no
longer breaks the bit stream, a 0 inside a long preamble is no longer a broken packet, and the decoders of a live
record read the acquisition's own samples instead of the screen's, so the 58 us halves of a DCC bit survive. A
decoder over a record of more than 262,144 points runs in the background and can show its result one record later. - Slow time bases put about every record in the picture again (issue #220): at 1 ms/div a batch holds the frame the
scope kept from before the arm and one more, and the first was always left out (issue #5), so only 18 of 37 records
a second were drawn. In a batch of up to four frames that frame is now kept when its edge is at the trigger point;
with other trigger types, and in longer batches, it is still left out (unit: 206 of 222 records at 1 ms/div). - SCPI and the web page come up when their port was still taken at the start (by the oos-ui before, ending): the
port is tried again every second. Before, both stayed off until the next start, while mDNS still announced them. - Update: an image with a larger initramfs (a 48 MB package) stopped in a kernel panic at the restart into it
("unknown-block(0,0)"); a power cycle brought the previous build back, nothing was written. The restart now places
the new initramfs above 1 GiB through kexec-tools, also when the device tree stays the same (unit). An update to such
an image needs this build installed first. - Measure in roll and scan, and on compressed deep records: the frequency, the period and the other time values of
a slow signal read up to twice the true value now and then (25 Hz read as 50 Hz at 0.2 s/div). Their points are
min/max pairs, and on an edge each pair spans more than the hysteresis of the edge search; the time values are now
measured on the mean of neighbouring points. Max, Min, Pk-pk, Top and Base still come from the pairs (unit, #118). - Colour grade drawn by the GPU: the densest pixels (a summed count from 32768 up) were left out on a driver that
keepsmediumpintegers in 16 bits; the shaders now compute counts inhighp(host, Mesa 26). - Bode: the generator amplitude knob reaches 10 mVpp (it stopped at 1 Vpp).
- Measure: a burst idle, setup or hold time above 5 s (set over SCPI) steps down to 5 s instead of turning off, also
in a measurement's own settings. - Waveform import: a damaged file is refused with "truncated waveform" wherever a field lies past its end.
- Fast-frame load (
frame-NNN.bin): a frame whose time step or start is not a finite number, or with fewer than two
samples per channel, is refused, as the waveform import already refused such anOOSWfile. oos-fpgadwithout a command, or withhelp, prints its usage again (it printed nothing).- Tests no longer depend on the speed of the computer in the UI either (wall-time limits replaced).
- The expand button in the plot's corner enlarges the panel also while HDMI is on Waveform or Layout. Before, it
only changed the density there and the panel stayed as it was (forum report). Full chosen under Density still
keeps the panel's bars while the external display has the waveform. - A typo in
net.allow(192.168.1.0/,2001:db8:::1) made nft refuse the whole firewall, and the scope came up
without one while telnet listened. Such entries are left out and logged; a ruleset nft refuses anyway closes the
scope to everything but loopback. - STOP (unit): the stopped record is no longer counted twice in the statistics, trend and catalog, and a change of view while
stopped adds nothing. Clear while stopped keeps the hardware stopped (it went on acquiring and overwrote the stored
record). RUN after STOP starts with a frame from after the RUN press. - Deep records read as min/max pairs: Mean, RMS, Std dev and AC RMS statistics no longer count pair values while the
reading is empty; with the LA on or three channels the measurements no longer stand still waiting for a raw read
the hardware never makes; after a Clear a held deep measurement no longer shows the cleared record. - The acquisition thread no longer holds the shared state while it decodes a raw window (up to ~24 MB), and the
newest frame of a batch gets the ADC-core correction like every other record. - Leaving a reference, Take over, Offline or SIM session brings back the live set-up from before it, whichever way it
is left (Live, Record = With live, a channel key, a channel as trigger source, Sources). A set-up recalled during a
session ends it and stays. Leaving the Bus trigger after a recall keeps the recalled decoder on. In Live, switching
a channel on from Sources leaves the trigger and Run/Stop alone. - Issue #5, stray traces (unit): frame 0 of a batch is left out in Batch draw mode too. Frame 1 is the first
trigger after the arm, and now and then it holds a piece of older memory. With an edge trigger it is left out
when its edge is not at the trigger point. SINGLE returned no frame from before the arm in 40 tries. :REF:LOAD "/proc/kmsg"froze the screen for good (unit). Files under /proc, /sys and /dev are refused now. So is
any file larger than the 128 MiB import limit, before it is read.- Waveform import (unit for the first two):
- A CSV with more columns than reference slots is refused while it is read. It used to take gigabytes first.
- An upload over the web page is parsed beside the frame loop: 32 MB took 6 s, and SCPI answered within 43 ms
meanwhile. - A frame file (
frame-NNN.bin) whose counts do not fit its size no longer crashes oos-ui.
- Web remote:
/updateand/mcpfrom outsidenet.alloware refused at the request head, before the body is read.
Only one large body is read at a time. - HiSLIP (desktop): a tiny maximum message size no longer splits a reply into one message per byte.
- The zoom window no longer overflows at extreme delays.
- The console banner and the Network page said SSH takes a provisioned key only. It takes the password or a key.
- With hardware segments recording, the DVM, the levels and the ADC core trim wait for the recording to end.
- STOP on a deep record (e.g. 10 M at 1 ms/div): the views read again as min/max pairs are drawn, while Mean, RMS,
Std dev and the other readouts stay measured on the raw record of the STOP. Before, they turned to "----". - Reference > Clear also removes the readouts of the cleared slots. Before, an import's Vpp, Freq, RMS and Rise rows
stayed empty (unit). - A
system.confthat leaves out every decoder no longer crashes Decode > Protocol. - Measurements: line and switching ripple without the rail's DC level on records of partial cycles; double pulse
I_L, td_on, tr, td_off, tf, di/dt, Ipeak - I_L and Qoss with a ramping load current; high time, low time and time
outside without noise from slow edges; Ones counts bits, not samples; setup/hold violations judged per clock edge,
and the Setup + Hold tile says it counts violations; Eye Width@BER in seconds; Hall shaft speed and angle
mechanical (divided by the pole pairs) and no longer stuck after a fault code at the record start. - Decoders: MLT-3 decodes every change between its three levels (half of the ones were lost).
- Phase noise and switching dV/dt, dI/dt no longer take quadratic time on long records; a power-rail clock longer
than the rail no longer crashes. - Tests no longer depend on the speed of the computer (wall-time limits replaced), and the SSH key tests run in a
checkout withoutlocal/. - The power-on picture (
/etc/oos/splash.ppm): one whose header ends inside a comment no longer ends oos-ui at the
start; it is left out like any picture oos-ui cannot read. - The power-on picture: one that claims a size of 2^31 bytes of pixels or more (65536 x 65536) no longer ends
oos-ui; it is refused unless the file holds all its pixels. - Bode with a generator on the USB port: a device that answers with transfers of no data and never ends the answer
no longer holds the generator's thread for ever. After 16 such transfers the read fails, and the Bode page says
that*IDN?failed. - Waveform import: a CSV whose time column gives an infinite time step or start (
infas a time) is refused, as an
OOSWfile with such a time base is. It became a reference with an infinite time step. - Mask test: a mask file whose tolerance
xoryisinforNaNis refused with "Load failed"; such a
tolerance became the mask's margin. - Setup recall: a
bode_sweepline that is refused (a sweep time of 0 orinf, or a later value that is wrong) no
longer leaves its first values in the Bode settings; the line changes nothing. - Decoders: Quad-SPI no longer ends oos-ui on a CS# frame that ends where it starts (both edges between the same two
samples) with a clock edge at that instant; such a frame holds no data. - Update: an image whose device tree nests deeper than 64 levels is refused with a message; oos-update ended on
one (from about 30 000 levels the stack overflowed). - Update: an image whose device tree section (RSCE) names a table longer than the section is refused; one with
entries of no length was read up to 2^32 times before the install went on.
Built by CI from the tag, on arm64, signed with the community key of the repository.