Repository navigation
[DreamNextGen] AMLogic audio + video backend fixes - #3
Merged
Merged
Conversation
New output adapters that route exteplayer3 directly through the Dreambox-
private DVB-API + ALSA, matching the runtime behaviour of dreamvideosink /
dreamaudiosink as observed by strace. Replaces the BCM-style writer chain
for the AML targets; enabled by --enable-dreamnextgen.
output/dream_video.c
====================
- VIDEO_SET_DEC_SYSINFO (48 B struct) + VIDEO_SET_FRAME (168 B struct,
GstClockTime-style pts) on /dev/dvb/adapter0/video0
- H.264 / H.265 / MPEG1/2 streamtype + dec_format mapping
- Producer/consumer ring (256 slots) between FFMPEGThread and a
dv_consumer thread; consumer burst-submits frames so the kernel
ringbuffer carries 3-4 frames ahead (vstream_cache ~ 250 kB,
buf_avail_num ~ 3) instead of the just-in-time 12 kB / buf=0 we got
from a per-frame pacer. EAGAIN + poll(50 ms, 250 ms giveup) on
/dev/dvb/.../video0 + /dev/amvideo_poll is the back-pressure
- Stream PTS used directly so tsync's pts_video stays in the same
timebase as pts_audio; monotonic 90 kHz fallback only when fr_pts < 0
- VIDEO_GET_EVENT drain before SET_FRAME, VIDEO_GET_PTS after — mirrors
dreamvideosink strace pattern. disable_video=0 on open, no
freerun_mode / show_first_frame_nosync touch (also matches strace)
- OUTPUT_FLUSH / OUTPUT_CLEAR drains the queue, VIDEO_CLEAR_BUFFER,
writes tsync/discontinue=1
- Dual logging (stderr + /tmp/dream_video.log) because serviceapp's
PlayerApp::stderrAvail filters non-JSON stderr out of the e2 debug log
output/dream_audio.c
====================
- ALSA PCM output on the dreamhdmi / dreamspdif / dreambt slave
(selected by /sys/class/amhdmitx/amhdmitx0/audio_source)
- FFmpeg software decoder feeding interleaved S16; rate / channels
from pcmPrivateData_t shipped through extradata by container_ffmpeg.c
- Producer/consumer queue (128 slots, 500 ms bounded push wait)
between DreamAudioWrite (on FFMPEGThread) and a da_consumer thread;
consumer owns the ALSA handle + drift logic so the demuxer never
blocks in snd_pcm_writei. Solves the AAC drain-storm we measured on
AAC live-TV AML mirrors (~30% drops without queue, ~5% with)
- Match dreamaudiosink runtime: tsync ENABLED in PCRMASTER (mode=2),
/proc/stb/pcr_offset = 0x0 + auto_pcr_offset = 0x0, pts_audio /
pts_video wiped on Open, signal tsync/discontinue. AMASTER caused
the kernel to release video at the audio-chunk arrival rate
(~50 fps for our 20 ms chunks) and drop ~33% of 25 fps frames —
visible as the "gefesselt" stutter
- Initial anchor on the first chunk after Open / Flush / Switch:
* |lead| > 2000ms → ALSA flush, drop buffer (1s throttle)
* lead in (+50, +2000]ms → push lead silence (cap 2000)
* lead in [-3000, -50)ms → queue (-lead) bytes of skip budget
Sustained-lag re-arm once anchor disarmed: |av_ms| > 1000 held 2s
re-arms (5s cooldown). 30s heartbeat log; no per-write logging.
- da_reset_anchor_locked() called from Stop / Flush / Switch — mirror
of dream_alsa_reset_anchor(). Without it, post-seek the drift code
computed apts_speaker vs a stale pts_video (kernel display still on
the pre-seek frame until the first I-frame decodes)
- Decoded audio rate compared against a separate m_alsa_dec_rate so
the HW-promoted rate from snd_pcm_hw_params_set_rate_near
(44100 → 48000) doesn't trigger a configure-loop on AAC 44.1 kHz
Live-TV streams
- IEC61937 passthrough for AC3 / EAC3 / DTS via byte-swap into a
static SPDIF buffer; TrueHD / DTS-HD MA HBR via a libavformat spdif
muxer producing whole 16-byte ALSA frames at 192 kHz / 8 ch
(per-device capability check, automatic fallback when the receiver
is not HBR-capable)
- Software volume Q15 scaling on the PCM path, bypassed for passthrough
- Dual-log to /tmp/dream_audio.log (same reason as dream_video)
container/container_ffmpeg.c
============================
- HLS-only HTTP read tuning (longer connect / recv windows)
- Local-file probesize + max_analyze_duration caps so MP4 / MKV opens
don't stall scanning thousands of streams
- TrueHD demux packets routed to the existing TrueHD writer
- Video buffers flushed on OUTPUT_FLUSH / OUTPUT_CLEAR
- The seconds-seek branch's output->Command(OUTPUT_CLEAR) stays
commented — that path runs every av_read_frame iteration that
passes the gate, not the seek boundary, and turning it on starved
the dream_audio queue on stitched HLS (PlutoTV vpts stuck for
minutes, av_drift = -425000 ms)
main/exteplayer.c
=================
- Skip noprobe for HTTP(S) streams so the network player gets the
container probe it needs
output/writer/mipsel/h264.c
===========================
- Include config.h on mipsel so the build picks up HAVE_* feature flags
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
When the kernel video decoder stalls (DASH segment underrun on hr-live etc.) pts_video stops advancing. The anchor would otherwise see growing av-drift and fire HUGE-gap → ALSA flush → 2000ms silence push → re-arm in a loop every few seconds, audibly chopping audio while video is stuck anyway. Track pts_video changes per anchor read. When the same value has been seen for >2s, skip the flush / silence-push / sustained-lag re-arm — audio plays through normally. Heartbeat logs " VFROZEN" so this stays visible. When the decoder recovers and pts_video advances again, the anchor resumes normal correction (gentle skip if drift accumulated). Reset state lands in DreamAudioOpen and da_reset_anchor_locked so seeks and channel changes start fresh.
S5443 is a pattern rule on the "/tmp/..." literal. We already opened the file with O_NOFOLLOW | O_CLOEXEC and mode 0600 in 13878c8 — which is the actual mitigation — but the rule does not inspect open flags, so it kept warning. Mark the two opens NOSONAR with a note explaining why the mitigation already covers the threat. [DreamNextGen] harden /tmp log opens against symlink-race SonarRule S5443: fopen("/tmp/dream_*.log", "a") is flagged because /tmp is a publicly writable directory and an attacker could pre-create a symlink at that path. On the receiver /tmp is a single-user tmpfs so the threat is theoretical, but the mitigation is free: fopen("a") -> open(O_WRONLY|O_CREAT|O_APPEND|O_NOFOLLOW|O_CLOEXEC,0600) + fdopen() O_NOFOLLOW refuses to follow a pre-existing symlink, 0600 makes the file private, O_CLOEXEC keeps it out of children spawned by exec(). Both files already include <fcntl.h> and <unistd.h>. Also annotate the http:// scheme prefix-match in exteplayer.c (Sonar S5332): exteplayer3 is the URL consumer, not the transport chooser. The literal "http://" is matched to decide whether to skip ffmpeg's noprobe shortcut; it is not used to open any HTTP socket. // NOSONAR.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
New output adapters that route exteplayer3 directly through the Dreambox- private DVB-API + ALSA, matching the runtime behaviour of dreamvideosink / dreamaudiosink as observed by strace. Replaces the BCM-style writer chain for the AML targets; enabled by --enable-dreamnextgen.
output/dream_video.c
output/dream_audio.c
Sustained-lag re-arm once anchor disarmed: |av_ms| > 1000 held 2s
re-arms (5s cooldown). 30s heartbeat log; no per-write logging.
of dream_alsa_reset_anchor(). Without it, post-seek the drift code
computed apts_speaker vs a stale pts_video (kernel display still on
the pre-seek frame until the first I-frame decodes)
the HW-promoted rate from snd_pcm_hw_params_set_rate_near
(44100 → 48000) doesn't trigger a configure-loop on AAC 44.1 kHz
Live-TV streams
static SPDIF buffer; TrueHD / DTS-HD MA HBR via a libavformat spdif
muxer producing whole 16-byte ALSA frames at 192 kHz / 8 ch
(per-device capability check, automatic fallback when the receiver
is not HBR-capable)
container/container_ffmpeg.c
main/exteplayer.c
output/writer/mipsel/h264.c