_TIFFCheckMalloc()/_TIFFCheckRealloc() used a unsafe way to detect overflow
in the multiplication of nmemb and elem_size (which are of type tmsize_t, thus
signed), which was especially easily triggered on 32-bit builds (with recent
enough compilers that assume that signed multiplication cannot overflow, since
this is undefined behaviour by the C standard). The original issue which lead to
this fix was trigged from tif_fax3.c
There were also unsafe (implementation defied), and broken in practice on 64bit
builds, ways of checking that a uint64 fits of a (signed) tmsize_t by doing
(uint64)(tmsize_t)uint64_var != uint64_var comparisons. Those have no known
at that time exploits, but are better to fix in a more bullet-proof way.
Or similarly use of (int64)uint64_var <= 0.
IGNORE placeholder in tif_dirread.c is now replaced by a field dir_ignore in the TIFFDirEntry structure
Currently, in tif_dirread.c a special IGNORE value for the tif tags is defined
in order to flag status preventing already processed tags from further processing.
This irrational behaviour prevents reading of custom tags with id code 0 - like tag GPSVERSIONID from EXIF 2.31 definition.
An additional field 'tdir_ignore' is now added to the TIFFDirEntry structure and code is changed
to allow tags with id code 0 to be read correctly.
This change was already proposed as pending improvement in tif_dirread.c around line 32.
Reference is also made to:
- Discussion in https://gitlab.com/libtiff/libtiff/merge_requests/39
- http://bugzilla.maptools.org/show_bug.cgi?id=2540
Comments and indention adapted.
Preparation to rebase onto master
in configure/CMakeList.txt :
- TIFF_INT8_T/TIFF_UINT8_T is signed/unsigned char
sizeof(char)==1 in C standard
- TIFF_INT16_T/TIFF_UINT16_T is signed/unsigned short
sizeof(short)>=2 in C standard
- TIFF_INT32_T/TIFF_UINT32_T is defined so its sizeof() is 4
- TIFF_INT64_T/TIFF_UINT64_T is defined so its sizeof() is 8
- TIFF_SIZE_T is defined so it has same sizeof() than size_t
- TIFF_SSIZE_T is defined so it has same sizeof() than unsigned char *
Those advanced writing functions must be used in a particular sequence
to make their intended effect. Their aim is to control when/where
the [Strip/Tile][Offsets/ByteCounts] arrays are written into the file.
The purpose of this is to generate 'cloud-optimized geotiff' files where
the first KB of the file only contain the IFD entries without the potentially
large strile arrays. Those are written afterwards.
The typical sequence of calls is:
TIFFOpen()
[ TIFFCreateDirectory(tif) ]
Set fields with calls to TIFFSetField(tif, ...)
TIFFDeferStrileArrayWriting(tif)
TIFFWriteCheck(tif, ...)
TIFFWriteDirectory(tif)
... potentially create other directories and come back to the above directory
TIFFForceStrileArrayWriting(tif): emit the arrays at the end of file
See test/defer_strile_writing.c for a practical example.
Found on GDAL with https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=14894
Disabling the TIFF_DEFERSTRILELOAD bit in ChopupStripArray() was a
bad idea since when using TIFFReadDirectory() to reload the directory again
would lead to a different value of td_rowsperstrip, which could confuse
readers if they relied on the value found initially.
This function replaces the use of TIFFReadEncodedStrip()/TIFFReadEncodedTile()
when the user can provide the buffer for the input data, for example when
he wants to avoid libtiff to read the strile offset/count values from the
[Strip|Tile][Offsets/ByteCounts] array.
... and add per-strile offset/bytecount loading capabilities.
Part of this commit makes the behaviour that was previously met when
libtiff was compiled with -DDEFER_STRILE_LOAD available for default builds
when specifying the new 'D' (Deferred) TIFFOpen() flag. In that mode, the [Tile/Strip][ByteCounts/Offsets]
arrays are only loaded when first accessed. This can speed-up the opening
of files stored on the network when just metadata retrieval is needed.
This mode has been used for years by the GDAL library when compiled with
its embeded libtiff copy.
To avoid potential out-of-tree code (typically codecs) that would use
the td_stripbytecount and td_stripoffset array inconditionnaly assuming they
have been loaded, those have been suffixed with _p (for protected). The
use of the new functions mentionned below is then recommended.
Another addition of this commit is the capability of loading only the
values of the offset/bytecount of the strile of interest instead of the
whole array. This is enabled with the new 'O' (Ondemand) flag of TIFFOpen()
(which implies 'D'). That behaviour has also been used by GDAL, which hacked
into the td_stripoffset/td_stripbytecount arrays directly. The new code
added in the _TIFFFetchStrileValue() and _TIFFPartialReadStripArray() internal
functions is mostly a port of what was in GDAL GTiff driver previously.
Related to that, the public TIFFGetStrileOffset[WithErr]() and TIFFGetStrileByteCount[WithErr]()
functions have been added to API. They are of particular interest when
using sparse files (with offset == bytecount == 0) and you want to detect
if a strile is present or not without decompressing the data, or updating
an existing sparse file.
They will also be used to enable a future enhancement where client code can entirely
skip bytecount loading in some situtations
A new test/defer_strile_loading.c test has been added to test the above
capabilities.
In most situations of BigTIFF file, the tile/strip sizes are of reasonable size,
that is they fit on a 4-byte LONG. So in that case, use LONG instead of LONG8
to save some space. For uncompressed file, it is easy to detect such situations
by checking at the TIFFTileSize64()/TIFFStripSize64() return. For compressed file,
we must take into account the fact that compression may sometimes result in
larger compressed data. So we allow this optimization only for a few select
compression times, and take a huge security margin (10x factor). We also only
apply this optimization on multi-strip files, so as to allow easy on-the-fly
growing of single-strip files whose strip size could grow above the 4GB threshold.
This change is compatible with the BigTIFF specification. According to
https://www.awaresystems.be/imaging/tiff/bigtiff.html:
"The StripOffsets, StripByteCounts, TileOffsets, and TileByteCounts tags are
allowed to have the datatype TIFF_LONG8 in BigTIFF. Old datatypes TIFF_LONG,
and TIFF_SHORT where allowed in the TIFF 6.0 specification, are still valid in BigTIFF, too. "
On a practical point of view, this is also compatible on reading/writing of
older libtiff 4.X versions.
The only glitch I found, which is rather minor, is when using such a BigTIFF
file with TileByteCounts/StripByteCounts written with TIFF_LONG, and updating
it with an older libtiff 4.X version with a change in the
[Tile/Strip][ByteCounts/Offsets] array. In that case the _TIFFRewriteField()
function will rewrite the directory and array with TIFF_LONG8, instead of updating
the existing array (this is an issue fixed by this commit). The file will
still be valid however, hence the minor severity of this.
