You list /dev/video0, start the capture pipeline, and it stops with 'no cameras available'. The dmesg trail shows the real failure: 'no acknowledge from address 0x10', the sensor driver's read probe returning -121, and the probe failing. Nothing above the bus is worth touching yet, because the sensor never answered on its I2C bus. This guide is for engineers integrating camera modules on their own carriers and single-board computers. You will decode the probe log, isolate the device or mux conflict that silences the bus, re-probe the sensor, size the pull-ups that keep it answering, and specify the next camera so the failure mode cannot return.
Why does /dev/video0 exist if the camera sensor probe failed?
The device node is registered by the video-driver layer at boot, while the sensor itself is reached over I2C in a separate probe step; a node with no formats behind it means that I2C probe failed, and the pipeline error 'no cameras available' is its downstream symptom.
Node and chip are different layers. The node appears when the platform binds the video device; the chip has to answer an identification read over I2C before the driver finishes probing it. Two commands separate a live camera from an empty shell:
-
ls /dev/video*lists nodes and says nothing about silicon. -
v4l2-ctl -d /dev/video0 --list-formats-extasks the sensor for its frame formats. An empty reply means the I2C client never answered, so every capture layer above it reports no camera.
When v4l2-ctl lists formats but your vendor's capture element still finds nothing, the I2C stage is healthy and the fault lives in the pipeline; capture the raw stream with v4l2-ctl when the vendor source element finds no camera picks that case apart.
What does 'no acknowledge from address 0x10' with error -121 mean?
The controller clocked the sensor's 7-bit address 0x10 onto the bus and no device pulled SDA low for the acknowledge bit, so the kernel logs 'no acknowledge from address 0x10', the driver's read probe returns -121, and the sensor fails to probe.
Decode the log and the test path follows from it. A client printed as 9-0010 lives on I2C bus 9 at 7-bit address 0x10, which is hexadecimal for 16; i2cdetect -y 9 prints that cell as 10 when a device answers. Error -121 is the kernel's no-acknowledge result: all nine bits of the address byte went out and SDA stayed high through the ninth, so no chip claimed the transfer. The driver records it as 'error during i2c read probe (-121)' followed by 'probe of 9-0010 failed with error -121', with your sensor's driver name in place of the client address.
Treat -121 as 'nothing answered' and -110 as 'the transfer never finished'; the fixes differ. The dmesg trail for a MIPI camera probe failed with error -121 walks the boot log line by line.
Why does a device that is not the camera break the sensor's I2C bus?
Because the sensor's adapter is shared: its I2C controller often feeds a GPIO-selected mux that exposes the camera bus one level down, so a module added on the same adapter — or one whose driver parks the mux on another channel — can leave the camera's address unacknowledged.
An I2C adapter is a party line with one pair of pull-ups, and every client on it, including devices you never think of as camera parts, can hold those two lines. The pattern repeats across builds: the camera works on a clean system, you add a wireless or storage card that shares the adapter, and the next boot ends in -121 at the sensor. Remove that card and the sensor answers. That A/B test — two boots and one log diff — is the fastest confirmation in this whole class of fault, and several camera modules on one host hit shared-bus limits well before the data path does.
The same mechanism reaches one level out from the sensor, into camera peripherals. An integration write-up of a PTZ camera controller's thread and I2C errors shows the same shape of fault one device out from the sensor: bus traffic and errors logged around a controller that shares its I2C.
How do you find what is holding the sensor's I2C bus?
Read the I2C lines in dmesg first: the 'Added multiplexed i2c bus' entries and the mux-driver line tell you which adapter the sensor sits behind, and the failing client's 9-0010 prefix names bus 9 as the one to test.
Then run four checks:
-
dmesg | grep -iE 'i2c|mux'lists every adapter, every multiplexed bus and the mux driver, such asi2c-mux-gpio, that created them. - Read the failing client's prefix:
9-0010means scan bus 9. -
i2cdetect -y 9sweeps it. A cell answering10means the sensor is alive and the boot failure was a conflict or a race; an empty grid means wiring, rails or the mux. - Diff the boot log against the saved working-boot log; the first I2C line that differs names the suspect.
Close with the A/B test: power down, remove the module you suspect, boot, compare again. If bus 9 comes alive only with the other module gone, you have the answer, and the physical-link fixes for an I2C error -121 on a camera sensor probe cover the wiring half of that diff.
How do you re-probe the camera sensor after clearing the conflict?
Power-cycle and re-run the two checks that failed: dmesg must bind the sensor driver without the -121 line, and v4l2-ctl --list-formats-ext must print real frame sizes; when both pass, the capture element finds the camera again.
The sensor driver runs its read probe at boot, so the reliable re-probe is a full power-down and restart; replugging a module while running does nothing for a CSI sensor:
dmesg | grep -iE 'probe|acknowledge'
v4l2-ctl -d /dev/video0 --list-formats-ext
A pass is a driver-bind line naming the sensor with no 'no acknowledge' entry under it, plus a format list with real frame sizes. Only now start the capture pipeline; 'no cameras available' disappears at this point, not before.
If formats list cleanly but frames come out black or striped, the I2C stage is closed and the fault has moved to the data lanes and their device-tree setup; the same device-tree chain that decides node creation also wires those lanes.
Which wiring faults give the same no-acknowledge on a custom carrier?
On a carrier you wired yourself, a NACK at 0x10 is almost always electrical before it is software: swapped SDA and SCL, missing pull-ups, an unpowered sensor rail, or an address strap driven to a different value all leave the chip silent.
Measure before you reflash. Most CSI sensor modules want three rails at the connector — 1.8 V core, 2.8 V analog and 3.3 V I/O is the common set — and a module with only the 3.3 V rail present NACKs every address, because the chip that owns 0x10 is unpowered. Confirm the address straps match the address your driver registers, and scope both lines for a clean rise to the rail.
Pull-ups decide whether the bus works at speed. The I2C-bus specification sets a window: a minimum from the sink current, a maximum from the rise time.
Worked example: a 3.3 V rail, a 400 kHz bus and a 15 cm carrier harness with about 50 pF of bus capacitance.
- Rp(min) = (3.3 V − 0.4 V) / 3 mA = 2.9 V / 3 mA ≈ 1 kΩ
- Rp(max) = 300 ns / (0.8473 × 50 pF) = 300 ns / 42.4 pF ≈ 7.1 kΩ, with 300 ns the fast-mode rise-time limit
- Fit 4.7 kΩ from each line to the 3.3 V I/O rail; it sits mid-window
The margin check explains intermittent boots: 10 kΩ on the same 50 pF gives tr = 0.8473 × 10 kΩ × 50 pF ≈ 424 ns, over the 300 ns limit, which is how a bus that worked yesterday NACKs at speed today. Where the failure pattern is an I2C bus that refuses to start until a probe wakes the line, the I2C bus startup failure cured by a probe on SDA matches that stuck-line behaviour step for step.
When is the I2C mux the cause rather than the device sharing it?
Suspect the mux when dmesg creates the multiplexed buses but the camera bus still never answers even with the other module gone: a GPIO-selected mux parked on the wrong channel, or claimed by another driver, orphans the sensor's bus entirely.
The signature is readable. Lines such as 'Added multiplexed i2c bus 9' and 'Added multiplexed i2c bus 10' under a 2 port mux mean the sensor sits one level down, and its bus number changes with the mux layout. Three checks close it:
- Scan every muxed bus with the suspect module removed —
i2cdetect -y 9andi2cdetect -y 10. The sensor must answer on exactly one of them. - Check which channel the camera's device-tree node names. If it points at the channel another driver parks the mux on, the camera is orphaned although the hardware is perfect.
- Confirm the mux's reset GPIO state selects the camera channel at boot. A mux that powers up on the wrong port produces the works-one-boot, fails-the-next pattern.
Two sensors behind one mux at the same address is legal; two sensors on one flat bus at the same address is not. The second sensor never enumerates and its driver ends in -121; the bus options that work when two image sensors share one I2C address lay out the strap and mux choices.
How do you keep this failure out of the next camera you specify?
Pick a module whose control channel does not share your carrier's general-purpose I2C bus: USB3 and GigE board-level cameras carry control with the data interface, so a busy or muxed I2C adapter cannot block the sensor.
That removes the failure class at the purchasing stage. A board-level camera bolts to your carrier with no I2C wiring of its own, and housed area scan cameras behave the same way, with control riding the GigE or USB3 link.
All three modules below carry their control over the data interface:
- Sensor format
- 1/2.9"
- Lens mount
- M12
- Sensor format
- 1/1.8"
- Lens mount
- C-mount
- Sensor format
- 1/2.7"
- Lens mount
- M12
The MV-CB016-10GM-M-W is a 1.6 MP GigE board camera with an M12 mount and a global shutter. The LS-BCM-0003 is a 3.1 MP USB3 board camera rated at 120 fps behind a C-mount with a global shutter. The LS-BCM-0061 is a 1.3 MP GigE module with an M12 mount for tight spaces.
If the task genuinely needs a raw CSI sensor, design the bus instead of discovering it: get the documented sensor address, rail list and connector pinout before layout, strap the address away from anything else on the adapter, place test points on SDA and SCL, and work through the board camera evaluation checklist for inspection systems before committing a carrier revision.
What order of checks clears a sensor probe failure fastest?
Work the six checks in order and most cases close in one boot cycle: read dmesg, remove the suspect bus device, power down fully, scan the sensor's bus, re-probe the format list, then start the capture pipeline — address 0x10 answering is the gate for everything after it.
- Read the I2C and mux lines in dmesg; take the sensor's bus number from the failing client's prefix.
- Power down fully and remove or reseat the module that shares or gates the adapter.
- Boot and confirm the driver binds with no 'no acknowledge' line under it.
- Scan the sensor's bus; the documented address —
10for 0x10 on bus 9 — must answer. - List formats with v4l2-ctl; real frame sizes close the I2C stage.
- Start the capture pipeline only now. If it still reports no camera, the fault is in the pipeline, not the bus.
The decision, per what you saw:
- NACK at boot, gone once another module is removed: a bus conflict. Keep that module off the camera's adapter or give the camera its own mux channel.
- NACK on every boot and an empty scan grid: wiring, rails or address straps. Fix the electrical layer before touching software.
- NACK that comes and goes between boots: the mux's default state, or pull-ups at the rise-time margin. Pin the mux default and refit 4.7 kΩ pull-ups.
- Formats list but frames are black: the I2C stage is closed; work the data lanes next.
If you would rather not own an I2C bus at all, the camera finder and the camera step of the configurator filter modules by interface, mount and resolution; tell us the carrier and the task and we will spec it with you. More fault isolation of this kind sits under the systems topic.
Questions
Is error -121 the same fault as error -110 on a camera's I2C bus?
No. -121 is the no-acknowledge result: the controller clocked the address and no device claimed it, so the fault is an absent, unpowered or conflicting device at that address. -110 is a timeout: a device or a wiring short holds SDA or SCL low so the transfer cannot complete at all. Fix -121 at the device level and -110 at the line level — pull-ups, shorts, or a chip clamping the bus.
Do all CSI camera sensors sit at I2C address 0x10?
No, and the log tells you yours. 0x10 is the address many sensor drivers register first, but the true address is whatever the module's documentation and the driver's board setup declare. Read it from the failing client's prefix — 9-0010 means bus 9, address 0x10 — or sweep the bus after a clean boot and use the answered value in every later test.
Can I remove the interfering module while the system is running?
For a pluggable wireless or storage card, yes: shut the interface down, remove the card, and reboot — the camera driver probes at boot, so the re-probe needs a restart anyway. Never do it to the camera ribbon itself. CSI modules have no hot-plug state machine, and mating or unmatting a powered ribbon can back-power the sensor through its I/O and destroy it.
Why does the camera work on one boot and fail on the next?
Intermittent probes come from three places: a mux whose reset GPIO state parks it on the wrong channel until something selects the camera port, pull-ups sitting at the rise-time margin so a warm bus NACKs at speed, and enumeration races where another client claims the adapter first. Diff the dmesg from a good boot against a bad one; the first differing I2C line names which of the three you have.
Does adding a second camera to the same adapter cause this failure?
Only through its address or its wiring. Two sensors with distinct addresses can share one adapter, though bus capacitance and trace length set the practical limit. Two sensors at the same address on one flat bus cannot: the probe reaches whichever chip answers first and the second never enumerates, which shows up as -121 on one of them. Give same-address sensors separate mux channels or strap one to a different address.


