ersatztv/docs/decisions/records/ffmpeg/qsv-extra-hw-frames-floor.md

key, title, status, since, supersedes, superseded-by, rule, signals, mechanics

key	title	status	since	supersedes	superseded-by	rule	signals	mechanics
ffmpeg.qsv-extra-hw-frames-floor	2026-07-21 — QSV hardware-frame headroom is a floor, not an operator preference (#529)	active	2026-07-21	none	none	a QSV upload never emits `extra_hw_frames` below `FFmpegState.MinimumQsvExtraHardwareFrames` (64); a stored `0` or negative value is treated as "no pool configured" rather than honored literally, because with no headroom any unthrottled read exhausts the pool and the transcode writes nothing at all.	QSV, extra_hw_frames, hwupload, hardware frame pool, ENOMEM, "Could not open encoder before EOF", readrate as an incidental allocation bound · paths: `FFmpegState`, `QsvPipelineBuilder`, `HardwareUploadFilter`, `ScaleQsvFilter`, `DeinterlaceQsvFilter` · issues: #529, #350, #516, #519	`FFmpegState.MinimumQsvExtraHardwareFrames`; `FFmpegState.QsvExtraHardwareFrames`; `QsvPipelineBuilder` logs once when a configured value is raised

• extra_hw_frames=0 is not a valid pool size; it is a dead channel waiting for an unthrottled read. FFmpegState.QsvExtraHardwareFrames honored a stored 0 literally, so hwupload=extra_hw_frames=0 reached FFmpeg with no headroom for frames in flight through the filter graph. Measured against the deployed FFmpeg 8.1.2 on one real logged command (software mpeg4 decode → hwupload → vpp_qsv → h264_qsv): the graph fails with -12 (Cannot allocate memory), h264_qsv reports "Could not open encoder before EOF", and zero segments are written. It now clamps to MinimumQsvExtraHardwareFrames (64), which is also the value IfNone already used for an unset profile and the seeded profile default.

• Input throttling was the only thing hiding it, which is why this looked like a #350 regression. Truth table, same command, same binary, only the marked tokens differing:

  readrate	extra_hw_frames	result
  1.05, no burst	0	14 segments, exit 0
  1.05 + burst 2 / 4 / 8	0	ENOMEM, 0 segments
  no readrate at all	0	ENOMEM, 0 segments
  1.05 + burst 8	64	14 segments, exit 0
  no readrate at all	64	14 segments, exit 0

  So the defect predates #350's burst: any work-ahead start (which takes no -readrate) on a pipeline that uploads to QSV was already failing on a profile with 0. The burst did not introduce it — it removed the throttle on every realtime session, converting an intermittent failure into a near-deterministic one, which is how it finally got noticed.

• -readrate was doing load-bearing work nobody had written down. Its stated job is live-TV pacing; it was also incidentally bounding how fast decoded frames enter the filter graph. This is why #350's FFmpeg-level benchmark and #516's argument-generation tests were both green and neither could see it: the burst is bounded in seconds of input, which is not a bound on memory or hardware surfaces. Corrects the #350 entry above, which records the burst as bounded and safe and does not mention hardware frame pools.

• A floor, not a clamp-to-default — and the floor is wider than the evidence. Values above 64 are honored unchanged; values below it are raised. We measured only 0 (fails) and 64 (works), so 1..63 are untested, not known-bad: we raise them rather than trust them, because the failure they risk is a channel that serves nothing at all. That is a deliberate over-reach, and it is not free — extra_hw_frames allocates additional surfaces (64 NV12 1080p surfaces ≈ 190 MiB, ≈760 MiB at 4K), so an operator who deliberately set a small pool on a memory-constrained iGPU silently gets a larger one. QsvPipelineBuilder therefore logs a warning naming both the configured and the applied value, so the override is discoverable rather than silent. If a smaller pool is ever measured safe, lower the floor rather than removing it.

• Fixed at FFmpegState.QsvExtraHardwareFrames, the single point every QSV upload site reads. HardwareUploadFilter, HardwareUploadQsvFilter and WatermarkHardwareUploadFilter read it directly; ScaleQsvFilter and DeinterlaceQsvFilter take values QsvPipelineBuilder passes down from it. One guard covers them all rather than five call sites that can drift apart. (A sixth formatter, SubtitleScaleQsvFilter, also emits extra_hw_frames but is currently dead code — no construction site exists in the solution — so it is not covered by this guard and would need the same value threading if it is ever revived.)

Accepted residual: the floor is applied at render time only, so a stored 0 keeps displaying as 0 in the SPA and over GET /api/v1/ffmpeg/profiles/{id} while FFmpeg receives 64 — the config no longer literally describes the behavior. Render-time was chosen deliberately: it fixes every existing deployment with no DB edit and no migration, which matters because this bug is already failing transcodes in production. New and updated profiles are normalized on save so stored rows converge on the truth, and the warning log closes the discoverability gap for rows that predate it; a backfill migration for old rows was judged not worth the dual-provider cost. Two consequences to know about: the save-time normalization is unconditional on hardwareAcceleration, so a non-QSV profile's stored value moves too (harmless — only the QSV path ever reads it — but it is a stored-state change on a field the user didn't touch); and a machine client that PUTs 0 gets a 200 and then reads back 64, which is a silent transform of a submitted value that the OpenAPI description does not advertise.