This patch is analogous to 20f972701806be20a77f808db332d9489343bb78:
It hides the internal part of AVBitStreamFilter by adding a new
internal structure FFBitStreamFilter (declared in bsf_internal.h)
that has an AVBitStreamFilter as its first member; the internal
part of AVBitStreamFilter is moved to this new structure.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Mainly reindentation, but some variables were also put into a smaller
scope.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
1. Left shifts of signed values are undefined as soon as the result is
no longer representable in the target type. Therefore make nal_size
an uint32_t and drop the check for whether it is < 0.
2. The two checks for overreads (whether the length field is contained
in the packet and whether the actual unit is contained in the packet)
can be combined into one because the packet is padded, i.e. a potential
overread caused by reading the length field without checking whether
said length field is actually part of the packet's buffer is allowed
as one always stays within the padding. But one has to be aware of
a pitfall: The comparison must be performed in (at least) int64_t as
otherwise buf_end - buf might be promoted to uint32_t in which case
an already occured overread would appear as a very large number.
A comment explaining this has been added, too.
3. Units of size zero are now silently dropped; the earlier code would
instead read the first byte of the next length field (or the first byte
of padding) to infer the type of the current unit.
4. Futhermore, the earlier code returned the wrong error code. This has
been fixed, too.
Fixes#8290.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Up until now, h264_mp4toannexb would grow the output packet's buffer by
the desired amount every time another NAL unit of the input packet has
been read; this commit changes this: The input buffer is now essentially
parsed twice, once to determine the final size of the output packet and
once to write the output packet's data.
Fixes: Timeout
Fixes: 19322/clusterfuzz-testcase-minimized-ffmpeg_BSF_H264_MP4TOANNEXB_fuzzer-5688407821123584
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
h264_mp4toannexb_filter currently uses both indices/offsets as well as
direct pointers comparisons for the checks whether one has reached or
even surpassed the end. This commit removes the offsets.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
If processing an input NAL unit triggers the insertion of data from
extradata in front of said NAL unit, the output packet is grown (i.e.
reallocated) once to accomodate both the new extradata as well as the
input NAL unit itself; this has been changed: In such a situation, the
packet is now grown twice. While this is bad for performance, it allows
to simplify the code and ultimately to stop reallocating the packet
altogether.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
According to the H.264 specifications, the only NAL units that need to
have four byte startcodes in H.264 Annex B format are SPS/PPS units and
units that start a new access unit. Before af7e953a, the first of these
conditions wasn't upheld as already existing in-band parameter sets
would not automatically be written with a four byte startcode, but only
when they already were at the beginning of their input packets. But it
made four byte startcodes be used too often as every unit that is written
together with a parameter set that is inserted from extradata received a
four byte startcode although a three byte start code would suffice
unless the unit itself were a parameter set.
FATE has been updated to reflect the changes. Although the patch leaves
the extradata unchanged, the size of the extradata according to the FATE
reports changes. This is due to a quirk in ff_h2645_packet_split which
is used by extract_extradata: If the input is Annex B, the first zero of
a four byte startcode is considered a part of the last unit (if any).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Up until now, h264_mp4toannexb stored the offset of the first SPS and
the first PPS in the (output) extradata in its context and used these
two numbers together with the size of the extradata and the pointer to
the extradata to determine what to insert when inserting extradata. This
led to some very long lines like "s->pps_offset != -1 ? s->pps_offset :
ctx->par_out->extradata_size - s->sps_offset". Therefore now pointers to
SPS and PPS are stored along with their respective sizes, so that e.g.
the above line can be changed to "s->sps_size".
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The format of an AVCDecoderConfigurationRecord, the out-of-band
extradata of H.264 in mp4, is as follows: First four bytes containing
version, profile and level, one byte for the length size and one byte
each for the number of SPS, followed by the SPS (each with its own size
field), followed by a byte containing the number of PPS followed by the
PPS with their size fields. While the number of SPS/PPS may be zero, the
bytes containing these numbers are mandatory. Yet the byte containing
the number of PPS has been ignored in two places:
1. In the initial check for whether the extradata can contain an
AVCDecoderConfigurationRecord. The minimum size is 7, not 6.
2. No check is made for whether the extradata ended right after the last
byte of the last SPS of the SPS array. Instead the first byte of the
padding is read as if it were part of the extradata and contained the
number of PPS (namely zero, given that the padding is zeroed). No error
or warning was ever raised.
This has been changed. Such truncated extradata is now considered
invalid; the check for 2. has been incorporated into the general size
check.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Before reading a 16bit size field during parsing of extradata, no check
is performed to make sure that said length field is actually contained
in the extradata. Given that this overread is not dangerous (the extradata
is supposed to be padded), only a comment for it has been added; the error
itself will be detected as part of the normal check for overreads.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Currently during parsing the extradata, h264_mp4toannexb checks for
overreads by adding the size of the current unit to the current position
pointer and comparing this to the end position of the extradata. But
pointer comparisons and pointer arithmetic are only defined if it does not
exceed the object it is used on (one past the last element of an array
is allowed, too). In practice, this might lead to overflows. Therefore
the check has been changed to use bytestream2_get_bytes_left() which
means that the pointers get subtracted and the result gets compared to
the available size.
Furthermore, the error code has been fixed.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This is done in order to improve readability. No functional change is
intended with this commit at all; in particular, the unsafe read
functions are used throughout as h264_extradata_to_annexb already
performs its own checks. (These checks will nevertheless be improved
in further commits.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
There can be at most 31 SPS and 255 PPS in the mp4/Matroska extradata.
Given that each has a size of at most 2^16-1, the length of the output
derived from these parameter sets can never overflow an ordinary 32 bit
integer. So use a simple uint32_t instead of uint64_t and replace the
unnecessary check with an av_assert1.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
the pps offset is used to locate pps in the spspps_buf; however, the
current calc method is wrong because it is the offset of the original
avctx->extradata;
when there is only one sps in the avcc; the value is correct by
coincidence, however, it will fail in avcc with multi sps
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
When only SPS or PPS is present in the stream, copy the missing one from
AVCC before insertion to the output stream.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
In order not to break a sequence like "SPS IDR SPS IDR", the boolean
telling that the SPS/PPS has been seen should always be set.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
If there are consecutive IDR pictures, then SPS/PPS should be prepended
to all of them, not only the first one.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Andreas Rheinhardt
Anton Khirnov
Andreas Rheinhardt
James Almer
Yusuke Nakamura
赵宇龙
Michael Niedermayer
Vittorio Giovara
Ivan Uskov
Benoit Fouet
![](/avatar/69a57eb249ecabf44270db708635b3d8?size=84 "Chris \\"Koying\\" Browet"){kind=small}
Chris \"Koying\" Browet
Michael Niedermayer
Alexandra Khirnova
Paul B Mahol
Stefano Sabatini
Luca Barbato
Martin Storsjö
gs_gail
Alex Converse
Mans Rullgard
Philip Langdale
Ronald S. Bultje