nvidia-smi on a busy card can take past the 3 s deadline, which latched the card
as stalled until restart. A late read now carries the last readings and is used
when it lands; no second read starts while one is out; only one out for 15 s
(past both nvidia-smi timeouts) marks the card stalled, and its return clears it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Graphs stepped once a history point (1 s at a 5 min window) while the numbers changed every
tick; they now slide a fraction of a column each tick. The process table (and the GPU boxes,
MHz and watts it carries) was polled on its own timer; it is now read per sample and committed
with it. GPU clients are walked every sample, and the 1–2 s floors on the gallery, Network,
Docker and ZFS ARC live views are gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The proprietary driver's DRM card says DRIVER=nvidia, which the driver list did not
know, so the card was never enumerated; and it exposes no sysfs busy counter or
fdinfo engine totals. Read it with nvidia-smi instead (--query-gpu per sample,
pmon -c 1 -s um for clients), by PCI address. Parsers unit-tested, including an
older pmon column layout and [N/A] fields.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The viewer tries the file as it is first. When the browser gives up — Safari and an MKV,
HEVC without a hardware decoder — or on a tap, a picker: codec (H.264, HEVC, AV1, or the
original codec repackaged) and size (original, 1080p, 720p, 480p), each codec labelled
with the engine this machine would use. POST /api/files/play runs ffmpeg into an HLS
event playlist of fragmented-MP4 segments under <data dir>/transcode/<id>/; Safari plays
HLS itself and everything else goes through hls.js, loaded only then. The recipes are
tried best first: a card on the bus before the integrated GPU, decoding on the same card
before decoding on the CPU and uploading, the software encoders last; a recipe that
produces no first segment is dropped for the next. The seekable range grows as ffmpeg
gets ahead; status says how far. Streams are reused for the same request, closed when
the player goes, swept after ten idle minutes, and the folder is emptied at start. Root
mode runs ffmpeg as root through a helper that ends when the daemon lets go of it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A previews button on the Files tab turns the icon column into a 256 px thumbnail for
every image and video row, loaded as rows scroll near. internal/thumbs probes what the
machine can do rather than assuming it: ffmpeg and ffprobe, every /dev/dri/renderD*
opened and exercised, vainfo's decode profiles per node where it exists. A video goes to
the first GPU that lists its codec, with ffmpeg's own software fallback behind it and a
CPU retry behind that; images decode on the CPU, JPEGs at a reduced DCT size. Previews
are kept under <data dir>/thumbs/ keyed by path, size and mtime, with undecodable files
remembered too. Root mode makes them through the same root helper the file reads use.
Where ffmpeg is missing the tab offers to install it; where a render node exists but is
closed to the daemon's user, it offers a drop-in adding the group and a restart.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The router has no lspci, so its GeminiLake chip came up as 'Intel GPU
3185' beside the identical chip's 'Intel UHD Graphics 600' on hamtv.
A trimmed pci.ids — the graphics devices of Intel, AMD and NVIDIA, and the
virtual GPUs, 115 KB — rides in the binary and is consulted first; lspci
is the fallback for a card newer than the table. scripts/pci-ids-gpu.sh
regenerates it from the hub's copy.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
On grid the sampler sat 42 minutes in runtime_pollWait on the Arc A380's
hwmon fan1_input: kernfs accepts epoll, so os.ReadFile made the fd
non-blocking, i915 answered EAGAIN once, and the poller waited for a
readiness kernfs only signals on sysfs_notify. Every graph on the machine
froze while the API carried on.
readStr now reads with plain blocking syscalls. Each GPU's sysfs reads run
under a three-second deadline; a device that does not answer is marked
stalled, logged, and left alone until restart, with the box saying so.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Every DRM device with a real GPU driver gets a box under the cpu box, laid out
like it: the busy curve, a meter for the device and one per engine, temperature,
clock, power and fan where the card has them, and the processes holding it open
with their share. The process table gains a Gpu% column and sort.
Device busy is Intel's RC6 idle residency inverted (what intel_gpu_top shows,
and readable without root) or amdgpu's gpu_busy_percent. Per-process shares are
DRM fdinfo deltas, as nvtop reads them; other users' fdinfo needs root, so a
daemon running as a user keeps one root child (hamlab drm-clients) over a pipe
rather than paying a sudo a second. Everything is also a gpu.* sensor.
What cannot be measured is left out, not shown as zero: an Arc A380 on i915 6.12
has no temperature sensor, an integrated GPU wears the package temperature
marked *, and Asahi has neither a busy counter nor fdinfo yet.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
