Age | Commit message (Collapse) | Author |
|
|
|
Don't mess with conditional compilation with LARGEFILE -- always expose
64-bit file offsets in our public API.
|
|
Update separations interface further to cope with whether spots
should be rendered separately, or as composite colors.
|
|
This commit adds an interface for a color management engine
to MuPDF, together with the internal changes required to
use it. We also add an implementation of this interface using
the LCMS2 library.
This commit is the "lcms" development branch squashed to a single
commit. The individual commits that make it up are as follows:
------
Add LCMS2 submodule.
Add required source files to MSVC libthirdparty project.
Plus changes to the Makefile.
------
Change name of libfont to libresources
The library can hold things other than fonts including
ICC profiles and eventually halftone screens etc.
------
Generate and include icc profiles on windows solution
Makefile for linux needs to be updated
------
Initial cut at interface to little cms
Methods for getting profile handles, link handles
and transform buffers as well as individual colors.
------
Install ICC profiles from source
When the source color space is an ICC profile
install the ICC space. Use alternative color
space if ICC is invalid.
------
Rename ICC resources
The way that fontdump generates the names was
causing some redundant prefix/suffix combinations.
------
Make resource ICC profiles usable
This brings in the resource profiles for use with the
target device. When mudraw is invoked with the icc
color type, it will set up the target pixmap to have an icc
based color space.
------
Make Default ICC color spaces not storable
The ICC color spaces in the color context should not be
put into the store. ICC color spaces that are contained
in the source document are however storable.
------
Add in support for single color conversion
This adds in the selection of the icc single color
converter. Note that we may want to look at using
the float conversion in lcms here since we are going
from and to float. The down side is that the creation
of such converters may be expensive.
------
Pixmap ICC conversion
Also clean up and further simplify some of the code. Use common
dst, src in function parameters instead of src, dst.
------
Add md5 calculation for profiles
Compare md5 of source and destination. If they
are the same set link as identity and do not bother
creating link from cmm.
------
Initial attempt at adding links to store
Next need to make sure I have all the dropping set up
correctly and that the links are getting freed when
we are all done with them.
------
Add drop for link
Make sure that we drop the links when we are done with them.
------
Fix icc link store ref. counting + rendering intent
The key allocations, rc and removal was wrong for the icc links.
Also added in support on the graphic state for rendering intent.
------
Add ICC profiles to Makefile.
------
Move ICC profile loading to colorspace.c
------
Fix build on linux.
------
Use hexdump in creating icc profiles
------
First cut at CalRGB CalGray handling
These color spaces are converted to equivalent ICC profiles
when the drawing operation occurs. ToDo still is avoid
creating brand new CalGray, CalRGB or ICC color spaces when
we encounter the same object in the source file. Instead
we should make use of the store.
------
Adding fz_color_params into device API
Stroke and fill of paths and text which will have color rendering related settings
including overprint, overprint mode, rendering intent and black point settings.
Images have a fz_color_params member added to them.
------
Rendering intent support for graphic and text
Graphic fills and strokes as well as text fill and strokes
handle the rendering intent settings. This works
through the display list. The parameters related to
color rendering, ri, bp, op, opm are packed in the flags
bits through the display list.
To do: Add support for images and shadings.
------
Add support for rendering intent with images
Required change to fill_image api, hence the large
number of files touched.
------
Add support for shadings rendering intent and DeviceN
This adds support for the rendering intent in shading fills.
Also, adds in support for color management of DeviceN and
Separation color spaces where the base space is defined to be
ICC.
------
Add clamping proc to colorspace-imp.h
In the head of mupdf, the index color values were not
being clamped properly. I moved the clamping operation
to be performed by a procedure in of the color space.
The a*, b* values in LAB will need to include a range value
for proper clamping as well as the ICC color space at some
point.
------
Fix assert test with index images
------
Support for DeviceN images color managed
The base space for DeviceN and separation images
is now color managed when the ICC work flow is used.
------
Add DefaultGray etc to names.txt
The DefaultGray, DefaultRGB, and DefaultCMYK settings
in the pdf page resource dictionary need to be parsed
and handled if present.
------
Remove methods for setting color space in context
The methods that existed for setting the color spaces
in the color space context were broken and we will not
be changing them from their initial setting at startup.
------
Add front end support for default color spaces
PDF can specify in the resource dictionary color spaces
that should be used for DeviceGray, DeviceRGB and DeviceCMYK.
This commit handles the extraction of these color spaces
and passes them through the display list and gets them
into a structure on a draw device (if one is used).
Next step is to have the draw device make use of them.
------
Backend of default color spaces
This handles the use of page level definitions for
DefaultGray, DefaultRGB, DefaultCMYK in the
draw device. The interface for the pixmap color
conversion had to be expanded a bit to pass the
default_cs object. This was needed due to the fact
that the image could be DeviceN or Separation and
have a base space that needs to be replaced by
a page level Default color space. Tested this
with a file that had two pages each with an common
reference image object but different DefaultCMYK
definitions in the page resource dictionary. Proper
rendering was obtained.
------
Add icc support to band-writer and png output
For image viewers that support icc profiles, it
is important that we include the icc profile
into those formats that support. Currently the
only output format we have that supports it is
PNG. But if we decide to do tiff, jpeg or psd
those also support the embedding of icc profiles.
------
Work toward multithreading
Each cloned context creates a new cmm context to
reduce contention. This may not be optimal though
as it may create conflicts in any shared store.
------
Add missing files.
Failed to check in icc-imp.h and icc34.h
------
Fix default device color calls
copy and paste errors.
------
Fixes
------
Fix clamping of lab colors
The lab colors when used with the lab resource profile
need to be properly clamped and scaled.
------
Turn off ICC create debug define
------
Fix memory leak issues
------
Set clamp proc in color context color spaces
When NO_ICC is enabled there was an issue as the
default spaces did not have the clamp proc (which was added
in the lcms branch).
------
Fix several issues after rebase
During rebase of the branch there were several
issues that came up. In addition there was
a fix related to the use of the lcms2 context branch
that did not make it in for some reason.
------
Fix for fz_store_type structure changes
------
Fixes for multithreaded
This fixes a few issues. One issue is that we can not be changing the
input and output formatter for the links if they are being shared among
threads. This adds the format as part of the link definition. Next step
is to create a link clone operation in lcms so that we can readily create
a different one with everything the same except the formatter.
Also, this shares the profiles that exist in the color context among the
threads. When a context is cloned, it will use the profiles in the current
context but it will create a new cmm context.c
------
Fix NO_ICC issues
There were a few issues that occurred when we compiled and ran
with the NO_ICC setting.
------
Change CMM to use lower resolution tables in links
The table resolution greatly affects the performance for some files. I don't see any significant color rendering issues going with a lower resolution.
Selecting cmsFLAGS_LOWRESPRECALC uses a MLUT resolution of 17. This is going to be
sufficient color wise and gives a large improvement performance wise especially for files
where the color link creation is significant compared to the rendering.
------
Fix link key creation to populate alpha and bit depth.
The store key for icc links needs to include the bit depth
and if the transform has to handle an alpha value. This
is needed due to the fact that we can't change the formatter when
we are sharing links among different threads.
------
Pull in some MSVC 2005 fixes for lcms2
------
Fix non-prototype prototype.
------
Miscellaneous typos, whitespace fixes.
------
Tidy colorspace creation API.
Rather than pass a magic implementation reference count value
(-1 for static, 1 for normal), pass a boolean "is_static" flag.
This gives a nicer API (IMAO etc).
------
Make all colorspaces use an fz_buffer.
------
Fix internal naming of MSVC project file.
------
Fix some error handling.
------
Add some consts.
------
Tweak fz_color_params and fz_store_hash.
Use uint8_t for fz_color_param entries - no need to use
larger.
Same change in fz_store_hash. This limits the size required
for the hash table too.
------
Throw on errors, rather than returning NULL.
fz_cmm_transform_pixmap should throw if it's fed pixmaps
incompatible with its transforms.
fz_cmm_new_ctx should throw if the cmm fails to initialise.
------
Ensure LCMS2_OBJs are built/linked using Makefile.
Also ensure that ICC_OBJs are included.
------
Fix some unused variable warnings.
------
cs_params fixes.
Use the ones that are passed.
------
fz_color_params tweaks.
Make them const in many places.
Cope with cs_params == NULL meaning default (i.e. fz_cs_param(ctx)).
Minimise the places we call fz_cs_params(ctx).
Consistently use cs_params rather than a mix of cs_param and cs_params.
------
Improve PDF color params handling.
PDF allows different OP settings for strokes and fills.
This means that either we need to keep separate entries in
each fz_color_param structure for the stroke one or the fill
one, OR we need to duplicate the fz_color_param structure.
It seems neatest to do the latter (not least because this
means we don't pass more information to each device function
than it actually needs).
Accordingly, we put an fz_color_param in each pdf_material.
We update the code that reads ExtGStates to set the values
appropriately. We take the opportunity to add support for the
PDF 2.0 UseBlackPtComp option too.
------
Fix colorspace ref count problems.
1) Don't drop colorspace until we've finished using it.
2) Don't drop it twice.
------
Convert NULL deref into a thrown error.
Seen with:
mutool draw -D -o out.png ../tests_private/comparefiles/Bug689760.pdf
We seem to have a pdf-cal space with no profile.
Talk to Michael about this.
------
Avoid using colorspace names to distinguish colorspaces.
Using strcmp is slow.
------
Cope with failure to parse default colorspaces during PDF page load.
------
Avoid SEGV due to to pdf_cal space with no profile.
As seen in:
tests_private/pdf/PDFIA1.7_SUBSET/CATX4879.pdf
------
Handle cases where base space of sep is pdf-cal
We handled this for images, but not for solid fills.
------
Accelerate ICC color conversion.
------
Cope with indexed images in the color management.
Some images (JPX images) come to the cm code still in
the indexed color space. Their base space could be DevN
so we need to cope with multiple base decodes. This
continues the decode until we get to a space for which
we can create an ICC link.
------
Ignore alpha presence in component count check
------
Eliminate recursion in fz_source_colorspace_cm.
------
Cope with bare ICCBased colorspace defs.
Bug692137.pdf has ICCBased colorspaced definitions given
directly as a stream, rather than as
[ /ICCBased <stream reference> ]
Acrobat copes with these, as does gs. We therefore update our
code to cope too.
Also, the PDF spec says that any problems found when reading
the Default spaces should be ignored (or at least not to abort
rendering). Update our code to do that too.
------
Tweak color converter logic for speed.
When we are using an icc profile based conversion, avoid ever
having to lookup the color converter for each conversion call;
do it at the lookup stage.
------
Harden us against failures during ICC link creation.
Seen with corrupted profiles.
------
Fix for handling alpha with lcms.
Note that this currently maintains an alpha when it is present.
We may need to do some work for the gray scale conversion to alpha mask.
------
Delay pdf-cal profile creation
Put the creation of the pdf cal profile into the link
creation function (rather than have it scattered). Also
be robust in the condition of failure to create the profile.
------
Proper clamping with embedded CIELAB ICC profile
If the ICC profiles alternate color space is LAB
then use the LAB clamping proc.
------
Use the color space clamp when converting to base color spaces
In many of the Ghent test files, we have a DeviceN image whose
alternate color space is CIELAB. We need to make sure to use
the CIELAB clamping operation in this case.
------
Make ICC runtime configurable.
Add fz_icc_flowflow, fz_set_icc_workflow functions to read ICC
workflow state, and set the ICC workflow state. The latter will
throw an error if trying to enable ICC workflow in a NO_ICC build.
Add -N flag to mutool to disable ICC.
Incorporates build fixes from Michael.
------
Ensure fz_draw_fill_image uses the pixmap colorspace
It appears that the image colorspace historically has been able
to differ from the pixmap colorspace. While we've fixed the cases
that we know about for this (see the previous commits), tweak
fz_draw_fill_image to work the way it always has in the past.
------
Add support for output intent
PDF documents can have an ICC profile defined in their Catalog
which defines the output color space and the color space to
use for one of the Device color spaces (e.g. DeviceGray,
DeviceRGB or DeviceCMYK).
------
Catch errors in default color spaces
Before setting the default color space contained in
the file, make sure it is the correct type.
Bug692213.pdf
------
Clamp to base space during sep color conversion
This was the source of a problem when the base space
was CIELAB.
------
Rename some functions to be more MuPDFy
drop rather than free etc.
------
Fix "casting away const" warning.
srcv is a const *. Do the clamping operation on src_map
(the same value) before it is assigned into the const
variable.
------
Rejig top level color management interface slightly.
Same code, just change the encapsulation.
------
Remove pre-multiplied alpha prior to color management
The pixmaps in mupdf use a premultiplied alpha format.
Prior to doing any color management we need to undo
the alpha and then reapply after color management.
------
Remove global output intent as unused.
------
Move fz_color_params to be the final arg in dev calls.
Frequently this will be NULL, and it doesn't form part of
the colorspace/color/alpha triple.
------
Rename fz_default_xxx static variables.
Remove fz prefix, to prepare for a later renaming that would conflict.
------
Rename cs_params to color_params.
------
Rename lcms branch identifiers.
------
Return device colorspaces if the default colorspace is NULL.
------
Clean up device call function for set_default_colorspaces.
------
Add missing rethrow.
------
Load page default colorspaces lazily.
------
murun: Add color params device argument.
Stubbed to always be NULL at the moment.
------
Rename extgstate processor ops.
------
Fix a few minor issues from Tor
Removes icc-imp.h
Rename color-icccreate.c
Add context to some methods
------
Fix javalib with recent lcms dev changes.
------
Update lcms2 with sub project fixes.
------
Fix build failure.
------
Add icc profile into other PNG output methods
------
Fix some ints that should be size_t's.
------
Tweak fz_new_icc_data_from_icc_colorspace.
Ensure that the colorspace is const, and that we set size to 0
if we don't find any data.
------
Combine band writer methods for header and ICC writing
For many formats (like PSD), we need to delay writing some of the
header until we know whether we are getting an ICC profile to
write or not. Makes more sense to just write them at the same
time.
------
Miscellaneous tweaks in colorspace.c
Mainly to avoid pointer aliasing, "nicer" whitespace, and a leak
on error.
------
Avoid rightward drift in get_base_icc.
------
Ensure that pdf_load_output_intent copes with exceptions.
If load_icc_based throws and exception, warn and continue. Don't
not render a file just because the inbuilt default profiles
are broken.
------
Revert change in pdf-page.c
The page resources load had been moved so we could get the default
colorspaces, but this has been moved into run_contents, so putting
stuff back as it was before.
------
Revert changes in mudraw.
We had added some colorspages in mudraw, and then later removed them.
Revert the changes to accomodate this to make the overall branch diffs
smaller.
------
Use fz_buffer in color-icc-create.c
------
Force mapping through proof icc profile always.
------
Fix behaviour on fz_cmm_new_profile failure.
If the profile fails to be made, return an error code,
and have the callers take sane steps.
------
Tweak load_icc_based
Cope better with errors in reading the ICC space not stopping
us loading the alternate.
------
Remove unused variable.
------
Get page resource for DefaultCS look-up
res = pdf_dict_get(ctx, PDF_NAME_Resources, page->obj); was not
returning resource.
Replaced with res = pdf_page_resources(ctx, page);
------
Move default color space set up to pdf_run_page_contents_with_usage
------
Review fixes for lcms branch.
------
Fixes for calibrated colorspace loading.
------
Add fz_document_output_intent wrapper. Lazy load intent for PDF.
------
Copy DefaultCS logic into pdf_run_annot_with_usage.
Same code as from pdf_run_page_contents_with_usage.
------
More review fixes.
------
Avoid rightward drift in pdf_load_cal_common
------
Rename color_converter functions to be find/drop.
Better than than lookup/discard. lookup suggests something that
doesn't need dropping, and we use drop rather than discard by
convention.
------
Move cmm from context into colorspace context.
------
Review fixes: Remove recursion and rename functions.
------
Don't access doc->oi directly in pdf_load_default_colorspaces.
------
Rename fz_colorspace_is_pdf_cal to fz_colorspace_is_cal and make it public.
------
Tweak function naming to be more consistent.
------
fz_md5_icc can be implemented using fz_md5_buffer.
------
Print full md5 checksums in link key debug prints.
------
Make fz_md5_buffer NULL safe.
------
Simplify debug saving of ICC profiles.
------
Rename fz_cmm_new/drop_profile to init/fin.
------
Indentation cleanups.
------
Move CMM static inline functions from private header to C file.
------
Tweak fz_icc_data_from_icc_colorspace to return a buffer.
Also, remove the _new_ from the name to reflect the fact that we
are passed a borrowed handle, not given a new reference.
------
java: Add ColorParams.pack() function.
------
Generate one C file for the embedded ICC profiles.
------
Return const pointer from fz_default_color_params.
------
Change misleading argument names to fz_new_colorspace.
------
Rename fz_cmm_new/drop_link to fz_cmm_init/fin_link.
------
Change definition of fz_cmm_instance.
Rather than void, use an undefined struct in keeping with the
rest of the code.
------
Add support for color managed bgr color space
------
Return unsigned char array from fz_lookup_icc.
------
Make default_color_params immutable.
Changing the defaults used by the draw device should happen via a
device call, should we need the functionality in the future.
------
Clean up error handling in color-lcms.c
------
Fix signed/unsigned warning.
|
|
|
|
|
|
This allows for overlaps, merges adjacent (mergeable) ranges
and gets us properly searchable results.
This causes 1 diff in the test suites (Bug694353.pdf), which is
due to the fallback font not having a hypen present at UCS 0x2010.
|
|
Remove AdobeCA.p7c from autogenerated files. Just include the
array in the source.
Simplifies makefile dependencies and makes the sizes of each
bit of data easier to look at.
It also paves the way for eventually using objcopy to create binary
objects for the fonts instead of needing to use hexdump.
|
|
|
|
This has knock on effects in the store.
fix
|
|
This makes it possible to redirect standard out and standard error output
streams to output streams of your liking.
This means that now you can, in gdb, type:
(gdb) call pdf_print_obj(ctx, fz_stdout(ctx), obj, 0)
(gdb) call fflush(0)
or when dealing with an unresolved indirect reference:
(gdb) call pdf_print_obj(ctx, fz_stdout(ctx), pdf_resolve_indirect(ctx, ref), 0)
(gdb) call fflush(0)
|
|
For now, just use it for controlling image decoding and image scaling.
|
|
|
|
Take on a (slightly tweaked) version of Simon Reinhardt's
patch.
The actual logic is left entirely unchanged; minor changes
have been made to the names of functions/types to avoid
clashing in the cmapdump.c repeated inclusion.
Currently this should really only affect xps files, as strtof
is only used as fz_atof, and that's (effectively) all xps for
now.
I will look at updating lex_number to call this in future.
|
|
In general, we should use 'const fz_blah' in device calls whenever
the callee should not alter the fz_blah.
Push this through. This shows up various places where we fz_keep
and fz_drop these const things.
I've updated the fz_keep and fz_drops with appropriate casts
to remove the consts. We may need to do the union dance to avoid
the consts for some compilers, but will only do that if required.
I think this is nicer overall, even allowing for the const<->no const
problems.
|
|
|
|
If FZ_LARGEFILE is defined when building, MuPDF uses 64bit offsets
for files; this allows us to open streams larger than 2Gig.
The downsides to this are that:
* The xref entries are larger.
* All PDF ints are held as 64bit things rather than 32bit things
(to cope with /Prev entries, hint stream offsets etc).
* All file positions are stored as 64bits rather than 32.
The implementation works by detecting FZ_LARGEFILE. Some #ifdeffery
in fitz/system.h sets fz_off_t to either int or int64_t as appropriate,
and sets defines for fz_fopen, fz_fseek, fz_ftell etc as required.
These call the fseeko64 etc functions on linux (and so define
_LARGEFILE64_SOURCE) and the explicit 64bit functions on windows.
|
|
Purge several embedded contexts:
Remove embedded context in fz_output.
Remove embedded context in fz_stream.
Remove embedded context in fz_device.
Remove fz_rebind_stream (since it is no longer necessary).
Remove embedded context in svg_device.
Remove embedded context in XML parser.
Add ctx argument to fz_document functions.
Remove embedded context in fz_document.
Remove embedded context in pdf_document.
Remove embedded context in pdf_obj.
Make fz_page independent of fz_document in the interface.
We shouldn't need to pass the document to all functions handling a page.
If a page is tied to the source document, it's redundant; otherwise it's
just pointless.
Fix reference counting oddity in fz_new_image_from_pixmap.
|
|
Rename fz_close to fz_drop_stream.
Rename fz_close_archive to fz_drop_archive.
Rename fz_close_output to fz_drop_output.
Rename fz_free_* to fz_drop_*.
Rename pdf_free_* to pdf_drop_*.
Rename xps_free_* to xps_drop_*.
|
|
The dtoa function is for doubles (which is what MuJS uses) but for MuPDF
we only need and want float precision in our output formatting.
|
|
|
|
Use intptr_t when casting between a jlong and a pointer to suppress errors
about different size words.
Add a 'u' suffix to unsigned values output by the cmap dump utility.
|
|
Increasing the existing data structure to 32-bit values would bloat the data
tables too much.
Simplify the data structure and use three separate range tables for lookups --
one with small 16-bit to 16-bit range lookups, one with 32-bit range lookups,
and a final one for one-to-many lookups.
This loses the range-to-table optimization we had before, but even with the
extra ranges this necessitates, the total size of the compiled binary CMap data
is smaller than if we were to extend the previous scheme to 32 bits.
|
|
The primary motivator for this is so that we can print floating point
values and get the full accuracy out, without having to print 1.5 as
1.5000000, and without getting 23e24 etc.
We only support %c, %f, %d, %o, %x and %s currently.
We only support the zero padding qualifier, for integers.
We do support some extensions:
%C turns values >=128 into UTF-8.
%M prints a fz_matrix.
%R prints a fz_rect.
%P prints a fz_point.
We also implement a fprintf variant on top of this to allow for
consistent results when using fz_output.
a
|
|
We define a document handler for each file type (2 in the case of PDF, one
to handle files with the ability to 'run' them, and one without).
We then register these handlers with the context at startup, and then
call fz_open_document... as usual. This enables people to select the
document types they want at will (and even to extend the library with more
document types should they wish).
|
|
If MuPDF is used in a project using Subversion or another VCS adding
hidden subfolders to each folder, cmapdump breaks when trying to
load the subfolder as cmap file. This fix is required starting with
643370f04348569b5e5e577660031d638537671c
|
|
|
|
|
|
|
|
|
|
|
|
When running under Windows, replace fopen with our own fopen_utf8
that converts from utf8 to unicode before calling the unicode
version of fopen.
|
|
Remove unused variable, silencing compiler warning.
No need to initialize variables twice.
Remove initialization of unread variable.
Remove unnecessary check for NULL.
Close output file upon error in cmapdump.
|
|
Previously, before interpreting a pages content stream we would
load it entirely into a buffer. Then we would interpret that
buffer. This has a cost in memory use.
Here, we update the code to read from a stream on the fly.
This has required changes in various different parts of the code.
Firstly, we have removed all use of the FILE lock - as stream
reads can now safely be interrupted by resource (or object) reads
from elsewhere in the file, the file lock becomes a very hard
thing to maintain, and doesn't actually benefit us at all. The
choices were to either use a recursive lock, or to remove it
entirely; I opted for the latter.
The file lock enum value remains as a placeholder for future use in
extendable data streams.
Secondly, we add a new 'concat' filter that concatenates a series of
streams together into one, optionally putting whitespace between each
stream (as the pdf parser requires this).
Finally, we change page/xobject/pattern content streams to work
on the fly, but we leave type3 glyphs using buffers (as presumably
these will be run repeatedly).
|
|
If compiled with -DDEBUG, cmapdump throws a large number of warnings
regarding thread locking. These are harmless and can be ignored, but
are, nonetheless, not pretty. Fixed here.
Thanks to Bas Weelinck for the report.
|
|
Make fz_clone_context copy existing AA settings.
Add accessor function for fz_bitmap.
Add more documentation for various functions/types.
|
|
Attempt to separate public API from internal functions.
|
|
Add docs for fz_store, fz_image, fz_halftones.
Move fz_item definition into res_store.c as it does not need to be
external.
Rename fz_store_context to fz_keep_store_context to be consistent.
|
|
We only open one instance of freetype per document. We therefore
have to ensure that only 1 call to it takes place at a time. We
introduce a lock for this purpose (FZ_LOCK_FREETYPE), and arrange
to take/release it as required.
We also update the font context so it is properly shared.
|
|
This is a significant change to the use of locks in MuPDF.
Previously, the user had the option of passing us lock/unlock
functions for a single mutex as part of the allocation struct.
Now we remove these entries from the allocation struct, and
make a separate 'locks' struct. This enables people to use
fz_alloc_default with locking.
If multithreaded operation is required, then the user is
required to create FZ_LOCK_MAX mutexes, which will be locked
or unlocked by MuPDF calling the lock/unlock functions within
the new fz_locks_context structure passed in at context creation.
These mutexes are not required to be recursive (they may be, but
MuPDF should never call them in this way). MuPDF avoids deadlocks
by imposing a locking ordering on itself; a thread will never take
lock n, if it already holds any lock i for which 0 <= i <= n.
Currently, there are 4 locks used within MuPDF.
Lock 0: The alloc lock; taken around all calls to user supplied
(or default) allocation functions. Also taken around all accesses
to the refs field of storable items.
Lock 1: The store lock; taken whenever the store data structures
(specifically the linked list pointers) are accessed.
Lock 2: The file lock; taken whenever a thread is accessing the raw
file. We use the debugging macros to insist that this is held
whenever we do a file based seek or read. We also insist that this
is never held when we resolve an indirect reference, as this can
have the effect of moving the file pointer.
Lock 3: The glyphcache lock; taken whenever a thread calls freetype,
or accesses the glyphcache data structures. This introduces some
complexities w.r.t type3 fonts.
Locking can be hugely problematic, so to ease our minds as to
the correctness of this code, we introduce some debugging macros.
These compile away to nothing unless FITZ_DEBUG_LOCKING is defined.
fz_assert_lock_held(ctx, lock) checks that we hold lock.
fz_assert_lock_not_held(ctx, lock) checks that we do not hold lock.
In addition fz_lock_debug_lock and fz_lock_debug_unlock are used
on every fz_lock/fz_unlock to check the validity of the operation
we are performing - in particular it checks that we do/do not already
hold the lock we are trying to take/drop, and that by taking this
lock we are not violating our defined locking order.
The RESOLVE macro (used throughout the code to check whether we need
to resolve an indirect reference) calls fz_assert_lock_not_held to
ensure that we aren't about to resolve an indirect reference (and
hence move the stream pointer) when the file is locked.
In order to implement the file locking properly, pdf_open_stream
(and friends) now lock the file as a side effect (because they
fz_seek to the start of the stream). The lock is automatically
dropped on an fz_close of such streams.
Previously, the glyph cache was created in a context when it was first
required; this presents problems as it can be shared between several
contexts or not, depending on whether it is created before the
contexts are cloned. We now always create it at startup, so it is
always shared.
This means that we need reference counting for the glyph caches.
Added here.
In fz_render_glyph, we take the glyph cache lock, and check to see
whether the glyph is in the cache. If it is, we bump the refcount,
drop the lock and returned the cached character. If it is not, we
need to render the character.
For freetype based fonts we keep the lock throughout the rendering
process, thus ensuring that freetype is only called in a single
threaded manner.
For type3 fonts, however, we need to invoke the interpreter again
to render the glyph streams. This can require reentrance to this
routine. We therefore drop the glyph cache lock, call the
interpreter to render us our pixmap, and take the lock again.
This dropping and retaking of the lock introduces a possible race
condition; 2 threads may try to render the same character at the
same time. We therefore modify our hash table insert routines to
behave differently if it comes to insert an entry only to find
that an entry with the same key is already there.
We spot this case; if we have just rendered a type3 glyph and when
we try to insert it into the cache discover that someone has beaten
us to it, we just discard our entry and use the cached one.
Hopefully this will seldom be a problem in practise; to solve it
properly would require greater complexity (probably involving
spotting that another thread is already working on the desired
rendering, and sleeping on a semaphore until it completes).
|
|
|
|
When we moved over to a context based system, we laid the foundation
for a thread-safe mupdf. This commit should complete that process.
Firstly, fz_clone_context is properly implemented so that it
makes a new context, but shares certain sections (currently
just the allocator, and the store).
Secondly, we add locking (to parts of the code that have
previously just had placeholder LOCK/UNLOCK comments). Functions
to lock and unlock a mutex are added to the allocator structure;
omit these (as is the case today) and no multithreading is
(safely) possible. The context will refuse to clone if these are
not provided.
Finally we flesh out the LOCK/UNLOCK comments to be real calls of
the functions - unfortunately this requires us to plumb fz_context
into the fz_keep_storable function (and all the fz_keep_xxx
functions that call it). This is the largest section of the patch.
No changes expected to any test files.
|
|
|
|
|
|
When fz_malloc (etc) are about to fail, we try to scavenge memory
from the store and then retry. We repeatedly try to bin objects
from the store until the malloc succeeds, or until we have nothing
else to bin.
This means we no longer need the 'aging' of the store, so this is
removed.
|
|
Firstly, we rename pdf_store to fz_store, reflecting the fact that
there are no pdf specific dependencies on it.
Next, we rework it so that all the objects that can be stored in
the store start with an fz_storable structure. This consists of
a reference count, and a function used to free the object when
the reference count reaches zero.
All the keep/drop functions are then reimplemented by calling
fz_keep_sharable/fz_drop_sharable. The 'drop' functions as supplied
by the callers are thus now 'free' functions, only called if
the reference count drops to 0.
The store changes to keep all the items in the store in the linked
list (which becomes a doubly linked one). We still make use of
the hashtable to index into this list quickly, but we now have
the objects in an LRU ordering within the list.
Every object is put into the store, with a size record; this is
an estimate of how much memory would be freed by freeing that
object.
The store is moved into the context and given a maximum size;
when new things are inserted into the store, care is taken to
ensure that we do not expand beyond this size. We evict any
stored items (that are not in use) starting from the least
recently used.
Finding an object in the store now takes a reference to it already.
LOCK and UNLOCK comments are used to indicate where locks need to
be taken and released to ensure thread safety.
|
|
Also: use 'cannot' instead of 'failed to' in error messages.
|
|
|
|
In builds that support configurable layers of antialiasing, move
the variables that control this into the context. This makes it
possible to safely use different levels of antialiasing in different
threads.
|
|
Freetype globals are not shared between threads currently - to do that
we'll need to introduce a lock.
|