[FFmpeg-devel] [PATCH 1/4] lavc/pgs_frame_merge_bsf: add bsf to merge PGS segments

[John Stebbins]

On Sat, 2020-05-02 at 22:30 +0200, Andreas Rheinhardt wrote:
> John Stebbins:
> > Required to remux m2ts to mkv
> > ---
> >  Changelog                        |   1 +
> >  doc/bitstream_filters.texi       |   8 ++
> >  libavcodec/Makefile              |   1 +
> >  libavcodec/bitstream_filters.c   |   1 +
> >  libavcodec/pgs_frame_merge_bsf.c | 168
> > +++++++++++++++++++++++++++++++
> >  5 files changed, 179 insertions(+)
> >  create mode 100644 libavcodec/pgs_frame_merge_bsf.c
> > 
> > diff --git a/Changelog b/Changelog
> > index d9fcd8bb0a..fec4867488 100644
> > --- a/Changelog
> > +++ b/Changelog
> > @@ -59,6 +59,7 @@ version <next>:
> >  - mv30 decoder
> >  - Expanded styling support for 3GPP Timed Text Subtitles (movtext)
> >  - WebP parser
> > +- PGS subtitle frame merge bitstream filter
> >  
> >  
> >  version 4.2:
> > diff --git a/doc/bitstream_filters.texi
> > b/doc/bitstream_filters.texi
> > index 8fe5b3ad75..21ed09986c 100644
> > --- a/doc/bitstream_filters.texi
> > +++ b/doc/bitstream_filters.texi
> > @@ -548,6 +548,14 @@ ffmpeg -i INPUT -c copy -bsf noise[=1]
> > output.mkv
> >  @section null
> >  This bitstream filter passes the packets through unchanged.
> >  
> > + at section pgs_frame_merge
> > +
> > +Merge a sequence of PGS Subtitle segments ending with an "end of
> > display set"
> > +segment into a single packet.
> > +
> > +This is required by some containers that support PGS subtitles
> > +(muxer @code{matroska}).
> > +
> >  @section prores_metadata
> >  
> >  Modify color property metadata embedded in prores stream.
> > diff --git a/libavcodec/Makefile b/libavcodec/Makefile
> > index 88944d9a3a..b630de21bc 100644
> > --- a/libavcodec/Makefile
> > +++ b/libavcodec/Makefile
> > @@ -1115,6 +1115,7 @@ OBJS-$(CONFIG_MP3_HEADER_DECOMPRESS_BSF)  +=
> > mp3_header_decompress_bsf.o \
> >  OBJS-$(CONFIG_MPEG2_METADATA_BSF)         += mpeg2_metadata_bsf.o
> >  OBJS-$(CONFIG_NOISE_BSF)                  += noise_bsf.o
> >  OBJS-$(CONFIG_NULL_BSF)                   += null_bsf.o
> > +OBJS-$(CONFIG_PGS_FRAME_MERGE_BSF)        += pgs_frame_merge_bsf.o
> >  OBJS-$(CONFIG_PRORES_METADATA_BSF)        += prores_metadata_bsf.o
> >  OBJS-$(CONFIG_REMOVE_EXTRADATA_BSF)       +=
> > remove_extradata_bsf.o
> >  OBJS-$(CONFIG_TEXT2MOVSUB_BSF)            += movsub_bsf.o
> > diff --git a/libavcodec/bitstream_filters.c
> > b/libavcodec/bitstream_filters.c
> > index 6b5ffe4d70..92619225f0 100644
> > --- a/libavcodec/bitstream_filters.c
> > +++ b/libavcodec/bitstream_filters.c
> > @@ -49,6 +49,7 @@ extern const AVBitStreamFilter
> > ff_mpeg4_unpack_bframes_bsf;
> >  extern const AVBitStreamFilter ff_mov2textsub_bsf;
> >  extern const AVBitStreamFilter ff_noise_bsf;
> >  extern const AVBitStreamFilter ff_null_bsf;
> > +extern const AVBitStreamFilter ff_pgs_frame_merge_bsf;
> >  extern const AVBitStreamFilter ff_prores_metadata_bsf;
> >  extern const AVBitStreamFilter ff_remove_extradata_bsf;
> >  extern const AVBitStreamFilter ff_text2movsub_bsf;
> > diff --git a/libavcodec/pgs_frame_merge_bsf.c
> > b/libavcodec/pgs_frame_merge_bsf.c
> > new file mode 100644
> > index 0000000000..cae5c75655
> > --- /dev/null
> > +++ b/libavcodec/pgs_frame_merge_bsf.c
> > @@ -0,0 +1,168 @@
> > +/*
> > + * Copyright (c) 2020 John Stebbins <jstebbins.hb at gmail.com>
> > + *
> > + * This file is part of FFmpeg.
> > + *
> > + * FFmpeg is free software; you can redistribute it and/or
> > + * modify it under the terms of the GNU Lesser General Public
> > + * License as published by the Free Software Foundation; either
> > + * version 2.1 of the License, or (at your option) any later
> > version.
> > + *
> > + * FFmpeg is distributed in the hope that it will be useful,
> > + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> > GNU
> > + * Lesser General Public License for more details.
> > + *
> > + * You should have received a copy of the GNU Lesser General
> > Public
> > + * License along with FFmpeg; if not, write to the Free Software
> > + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
> > 02110-1301 USA
> > + */
> > +
> > +/**
> > + * @file
> > + * This bitstream filter merges PGS subtitle packets containing
> > incomplete
> > + * set of segments into a single packet
> > + *
> > + * Packets already containing a complete set of segments will be
> > passed through
> > + * unchanged.
> > + */
> > +
> > +#include "avcodec.h"
> > +#include "bsf.h"
> > +#include "libavutil/intreadwrite.h"
> > +
> > +enum PGSSegmentType {
> > +    PALETTE_SEGMENT      = 0x14,
> > +    OBJECT_SEGMENT       = 0x15,
> > +    PRESENTATION_SEGMENT = 0x16,
> > +    WINDOW_SEGMENT       = 0x17,
> > +    DISPLAY_SEGMENT      = 0x80,
> > +};
> > +
> > +typedef struct PGSMergeContext {
> > +    AVPacket *buffer_pkt, *in;
> > +    int presentation_found;
> > +} PGSMergeContext;
> > +
> > +static void frame_merge_flush(AVBSFContext *bsf)
> > +{
> > +    PGSMergeContext *ctx = bsf->priv_data;
> > +
> > +    av_packet_unref(ctx->in);
> > +    av_packet_unref(ctx->buffer_pkt);
> 
> flush should reset ctx->presentation_found, shouldn't it?

Ok

> 
> > +}
> > +
> > +static int frame_merge_filter(AVBSFContext *bsf, AVPacket *out)
> > +{
> > +    PGSMergeContext *ctx = bsf->priv_data;
> > +    AVPacket *in = ctx->in, *pkt = ctx->buffer_pkt;
> > +    int ret, i, size, pos, display = 0;
> > +
> > +    if (!in->data) {
> > +        ret = ff_bsf_get_packet_ref(bsf, in);
> > +        if (ret < 0)
> > +            return ret;
> > +    }
> > +    if (!in->size) {
> > +        av_packet_unref(in);
> > +        return AVERROR(EAGAIN);
> > +    }
> > +
> > +    // Validate packet data and find display_end segment
> > +    size = in->size;
> > +    i = 0;
> > +    while (i + 3 <= in->size) {
> > +        uint8_t segment_type;
> > +        int segment_len;
> > +
> > +        segment_type = in->data[i];
> > +        segment_len  = AV_RB16(in->data + i + 1) + 3;
> > +        if (i + segment_len > in->size) // Invalid data
> 
> Possible overflow. Use an unsigned as segment_len to avoid it.

Not trying to be pithy, just making sure I understand exactly what you
mean.

If I'm understanding right, you wish to protect against the case that
in->size is > 2^31 - 2^16 and there are more than 32768 segments found
and the last segment length is > 2^31 - in->size? 

Those are the only circumstances I see an overflow happening, so am I
understanding correctly?

> 
> > +            break;
> > +        if (segment_type == PRESENTATION_SEGMENT && i + 10 < in-
> > >size &&
> > +            !ctx->presentation_found) {
> > +            uint8_t state;
> > +            ctx->presentation_found = 1;
> > +            ret = av_packet_copy_props(pkt, in);
> 
> This will copy the properties even if it eventually turns out that
> one
> can just pass through the packet. Even worse: Imagine the input
> packet
> having side-data. Then pkt contains side-data, yet the code thinks
> that
> this packet is blank. This side data will then leak when the
> properties
> of the next presentation segment are copied.
> 

Right, thanks.

> Furthermore, I would actually have expected the pts of the display
> segment to be used. Or are they guaranteed to be identical?

pgssubdec gave the "end display set segment" a poor name (i'm guessing
for the sake of brevity) and I've gone an propagated it :(  I'll rename
that enum to make it more clear.

Display time is actually defined by the presentation segment's pts. 

> 
> > +            if (ret < 0)
> > +                goto fail;
> > +            state = in->data[i + 10] & 0xc0;
> 
> This amounts to the composition_state of the
> composition_descriptor(),
> doesn't it?

Yes, it turns out any non-zero value for the composition state says
this in the spec:

"The Display Set contains all elements needed to display the next 
composition." 

Different non-zero states define how the "Epoch" is handled which
basically defines when things come into and go out of scope.

> 
> > +            if (state)
> > +                pkt->flags |= AV_PKT_FLAG_KEY;
> 
> The keyframe flag will be ignored on passthrough. 

Will fix

> Also, should the flag
> be cleared if state == 0?
> 

Yes. Will do.

> > +        }
> > +        i += segment_len;
> > +        if (segment_type == DISPLAY_SEGMENT) {
> > +            size = display = i;
> > +            break;
> > +        }
> > +    }
> > +    if (display && pkt->size == 0 && size == in->size) { //
> > passthrough
> > +        ctx->presentation_found = 0;
> > +        av_packet_move_ref(out, in);
> > +        return 0;
> > +    }
> > +    if ((!display && i != in->size) || size > in->size) {
> 
> I don't see how size > in->size can ever be true.
> 

It could in an earlier iteration.  Will fix.

> > +        av_log(bsf, AV_LOG_WARNING, "Failed to parse PGS
> > segments.\n");
> > +        // force output what we have
> > +        display = size = in->size;
> 
> If you output a corrupt packet, you should set the
> AV_PKT_FLAG_CORRUPT flag.
> 

OK

> > +    }
> > +
> > +    pos = pkt->size;
> > +    ret = av_grow_packet(pkt, size);
> > +    if (ret < 0)
> > +        goto fail;
> > +    memcpy(pkt->data + pos, in->data, size);
> > +
> > +    if (size == in->size)
> > +        av_packet_unref(in);
> > +    else {
> > +        in->data += size;
> > +        in->size -= size;
> > +    }
> > +
> > +    if (display) {
> > +        ctx->presentation_found = 0;
> 
> I don't see anything in the code that actually enforces that a
> presentation segment has been found at this point; the properties of
> the
> output packet would then still be blank. Is this intended? If it is
> spec-incompliant, then one should add the AV_PKT_FLAG_CORRUPT flag.
> 

Good point.  Will do.

> > +        av_packet_move_ref(out, pkt);
> > +        return 0;
> > +    }
> > +    return AVERROR(EAGAIN);
> > +
> > +fail:
> > +    ctx->presentation_found = 0;
> > +    frame_merge_flush(bsf);
> > +    return ret;
> > +}
> > +
> > +static int frame_merge_init(AVBSFContext *bsf)
> > +{
> > +    PGSMergeContext *ctx = bsf->priv_data;
> > +
> > +    ctx->in  = av_packet_alloc();
> > +    ctx->buffer_pkt = av_packet_alloc();
> > +    if (!ctx->in || !ctx->buffer_pkt)
> > +        return AVERROR(ENOMEM);
> > +
> > +    return 0;
> > +}
> > +
> > +static void frame_merge_close(AVBSFContext *bsf)
> > +{
> > +    PGSMergeContext *ctx = bsf->priv_data;
> > +
> > +    av_packet_free(&ctx->in);
> > +    av_packet_free(&ctx->buffer_pkt);
> > +}
> > +
> > +static const enum AVCodecID frame_merge_codec_ids[] = {
> > +    AV_CODEC_ID_HDMV_PGS_SUBTITLE, AV_CODEC_ID_NONE,
> > +};
> > +
> > +const AVBitStreamFilter ff_pgs_frame_merge_bsf = {
> > +    .name           = "pgs_frame_merge",
> > +    .priv_data_size = sizeof(PGSMergeContext),
> > +    .init           = frame_merge_init,
> > +    .flush          = frame_merge_flush,
> > +    .close          = frame_merge_close,
> > +    .filter         = frame_merge_filter,
> > +    .codec_ids      = frame_merge_codec_ids,
> > +};
> > 
> 
> 


Also, will add av_log header you comment on in your next email.