Message ID | 20230310091753.10230-2-nicolas.gaullier@cji.paris |
---|---|
State | New |
Headers | show |
Series | avcodec/dolby_e: Add error recovery when parse_mantissas run out of bits | expand |
Context | Check | Description |
---|---|---|
yinshiyou/commit_msg_loongarch64 | warning | Please wrap lines in the body of the commit message between 60 and 72 characters. |
yinshiyou/make_loongarch64 | success | Make finished |
yinshiyou/make_fate_loongarch64 | success | Make fate finished |
andriy/commit_msg_x86 | warning | Please wrap lines in the body of the commit message between 60 and 72 characters. |
andriy/make_x86 | success | Make finished |
andriy/make_fate_x86 | success | Make fate finished |
fre 2023-03-10 klockan 10:17 +0100 skrev Nicolas Gaullier: > + if (i == c->nb_groups - 1 > + && count * size1 > get_bits_left(&s->gb) > + && get_bits_left(&s->gb) >= 0 > + && (int)(mnt - c->mantissas) >= MIN_MANTISSAS) { > + av_log(s->avctx, AV_LOG_WARNING, "Truncated > mantissas @%d, " > + "highest frequencies not recoverable\n", > (int)(mnt - c->mantissas)); > + break; > + } Surely there's a proper way to fix this rather than having an arbitrary threshold. At the very least the get_bits_left() check could be moved to before parse_mantissas(). If get_bits_left() is < 0 after parse_mantissas() then a warning could be issued instead of erroring out, which should have an effect similar to this. Is there a spec saying what to do with truncated packets? Since this is Dolby-E I suspect the answer is "no". /Tomas
>fre 2023-03-10 klockan 10:17 +0100 skrev Nicolas Gaullier: >> + if (i == c->nb_groups - 1 >> + && count * size1 > get_bits_left(&s->gb) >> + && get_bits_left(&s->gb) >= 0 >> + && (int)(mnt - c->mantissas) >= MIN_MANTISSAS) { >> + av_log(s->avctx, AV_LOG_WARNING, "Truncated >> mantissas @%d, " >> + "highest frequencies not recoverable\n", >> (int)(mnt - c->mantissas)); >> + break; >> + } > >Surely there's a proper way to fix this rather than having an arbitrary threshold. At the very least the get_bits_left() check could be moved to before parse_mantissas(). If get_bits_left() is < 0 after >parse_mantissas() then a warning could be issued instead of erroring out, which should have an effect similar to this. > >Is there a spec saying what to do with truncated packets? Since this is Dolby-E I suspect the answer is "no". > >/Tomas Thank you for inspecting the patch. Concerning the arbitrary threshold: it may be understood as a very high level safeguard, but my intend was mostly to make the code more easily understandable. Actually, the fix is restricted to the last audio band (see "i == c->nb_groups - 1") and, at the end, the MIN_MANTISSAS condition could be simply removed with no audible risk, I guess. But to get this precision, the fix cannot be moved elsewhere. The idea here is to still be able to detect harsh packet truncation with full errors and keep current code behaviour in that case. Of course, this patch looks as a work-around, but a work-around that make it possible to decode the official reference bitstream implementation (dp600 at least, but maybe they are no different ones). So, this is a somewhat special scenario. Sadly, I don't see any another way to handle this. To conclude, I don't have any other idea than the complete removal of the arbitrary threshold. Would it make the patch more acceptable ? Nicolas
diff --git a/libavcodec/dolby_e.c b/libavcodec/dolby_e.c index 921c33f3ba..a24edfcc93 100644 --- a/libavcodec/dolby_e.c +++ b/libavcodec/dolby_e.c @@ -845,6 +845,7 @@ static int parse_indices(DBEContext *s, DBEChannel *c) return 0; } +#define MIN_MANTISSAS 800 static int parse_mantissas(DBEContext *s, DBEChannel *c) { DBEGroup *g; @@ -885,6 +886,14 @@ static int parse_mantissas(DBEContext *s, DBEChannel *c) } } } else { + if (i == c->nb_groups - 1 + && count * size1 > get_bits_left(&s->gb) + && get_bits_left(&s->gb) >= 0 + && (int)(mnt - c->mantissas) >= MIN_MANTISSAS) { + av_log(s->avctx, AV_LOG_WARNING, "Truncated mantissas @%d, " + "highest frequencies not recoverable\n", (int)(mnt - c->mantissas)); + break; + } for (k = 0; k < count; k++) mnt[k] = get_sbits(&s->gb, size1) * scale; }
Mantissas are the last data in the channel subsegment and it appears it is sometimes missing a very few bits for the parsing to complete. This should not be confused with data corruption. For 5.1+2@25fps, the occurence of this issue is pretty steady and about once every 2 hours. The truncation is at about 950 out of the 1024 values (never seen below 923 so far). The current code raises a severe 'Read past end' error and all data is lost resulting in 20ms(@25fps) of silence for the affected channel. This patch introduces a tolerance: if 800 out of the 1024 mantissas have been parsed, a simple warning is raised and the data is preserved. Signed-off-by: Nicolas Gaullier <nicolas.gaullier@cji.paris> --- libavcodec/dolby_e.c | 9 +++++++++ 1 file changed, 9 insertions(+)