QPDF Manual

The qpdf package has relatively few external dependencies. In order to build qpdf, the following packages are required:

Part of qpdf's test suite does comparisons of the contents PDF files by converting them images and comparing the images. You can optionally disable this part of the test suite by running configure with the --disable-test-compare-images flag. If you leave this enabled, the following additional requirements are required by the test suite. Note that in no case are these items required to use qpdf.

This option is primarily intended for use by packagers of qpdf so that they can avoid having the qpdf packages depend on tiff and ghostscript software.

If Adobe Reader is installed as acroread, some additional test cases will be enabled. These test cases simply verify that Adobe Reader can open the files that qpdf creates. They require version 8.0 or newer to pass. However, in order to avoid having qpdf depend on non-free (as in liberty) software, the test suite will still pass without Adobe reader, and the test suite still exercises the full functionality of the software.

Pre-built documentation is distributed with qpdf, so you should generally not need to rebuild the documentation. In order to build the documentation from its docbook sources, you need the docbook XML style sheets (http://downloads.sourceforge.net/docbook/). To build the PDF version of the documentation, you need Apache fop (http://xml.apache.org/fop/) version 0.94 of higher.

Building qpdf on UNIX is generally just a matter of running

./configure
make

You can also run make check to run the test suite and make install to install. Please run ./configure --help for options on what can be configured. You can also set the value of DESTDIR during installation to install to a temporary location, as is common with many open source packages. Please see also the README and INSTALL files in the source distribution.

Building on Windows is a little bit more complicated. For details, please see README-windows.txt in the source distribution. You can also download a binary distribution for Windows. There is a port of qpdf in the contrib area generously contributed by Jian Ma. This is also discussed in more detail in README-windows.txt.

There are some other things you can do with the build. Although qpdf uses autoconf, it does not use automake but instead uses a hand-crafted non-recursive Makefile that requires gnu make. If you're really interested, please read the comments in the top-level Makefile.

When running qpdf, the basic invocation is as follows:

qpdf [ options ] infilename [ outfilename ]

This converts PDF file infilename to PDF file outfilename. The output file is functionally identical to the input file but may have been structurally reorganized. Also, orphaned objects will be removed from the file. Many transformations are available as controlled by the options below.

outfilename does not have to be seekable, even when generating linearized files. Specifying “-” as outfilename means to write to standard output. However, you can't specify the same file as both the input and the output because qpdf reads data from the input file as it writes to the output file.

Most options require an output file, but some testing or inspection commands do not. These are specifically noted.

The following options are the most common ones and perform commonly needed transformations.

Password-protected files may be opened by specifying a password. By default, qpdf will preserve any encryption data associated with a file. If --decrypt is specified, qpdf will attempt to remove any encryption information. If --encrypt is specified, qpdf will replace the document's encryption parameters with whatever is specified.

Note that qpdf does not obey encryption restrictions already imposed on the file. Doing so would be meaningless since qpdf can be used to remove encryption from the file entirely. This functionality is not intended to be used for bypassing copyright restrictions or other restrictions placed on files by their producers.

To change the encryption parameters of a file, use the --encrypt flag. The syntax is

--encrypt user-password owner-password key-length [ restrictions ] --

Note that “--” terminates parsing of encryption flags and must be present even if no restrictions are present.

Either or both of the user password and the owner password may be empty strings.

The value for key-length may be 40 or 128. The restriction flags are dependent upon key length. When no additional restrictions are given, the default is to be fully permissive.

If key-length is 40, the following restriction options are available:

If key-length is 128, the following restriction options are available:

The default for each permission option is to be fully permissive.

These transformation options control fine points of how qpdf creates the output file. Mostly these are of use only to people who are very familiar with the PDF file format or who are PDF developers. The following options are available:

By default, when a stream is encoded using non-lossy filters that qpdf understands and is not already compressed using a good compression scheme, qpdf will uncompress and recompress streams. Assuming proper filter implements, this is safe and generally results in smaller files. This behavior may also be explicitly requested with --stream-data=compress.

When --stream-data=preserve is specified, qpdf will never attempt to change the filtering of any stream data.

When --stream-data=uncompress is specified, qpdf will attempt to remove any non-lossy filters that it supports. This includes /FlateDecode, /LZWDecode, /ASCII85Decode, and /ASCIIHexDecode. This can be very useful for inspecting the contents of various streams.

When --normalize-content=y is specified, qpdf will attempt to normalize whitespace and newlines in page content streams. This is generally safe but could, in some cases, cause damage to the content streams. This option is intended for people who wish to study PDF content streams or to debug PDF content. You should not use this for “production” PDF files.

Ordinarily, qpdf will attempt to recover from certain types of errors in PDF files. These include errors in the cross-reference table, certain types of object numbering errors, and certain types of stream length errors. Sometimes, qpdf may think it has recovered but may not have actually recovered, so care should be taken when using this option as some data loss is possible. The --suppress-recovery option will prevent qpdf from attempting recovery. In this case, it will fail on the first error that it encounters.

Object streams, also known as compressed objects, were introduced into the PDF specification at version 1.5, corresponding to Acrobat 6. Some older PDF viewers may not support files with object streams. qpdf can be used to transform files with object streams to files without object streams or vice versa. As mentioned above, there are three object stream modes: preserve, disable, and generate.

In preserve mode, the relationship to objects and the streams that contain them is preserved from the original file. In disable mode, all objects are written as regular, uncompressed objects. The resulting file should be readable by older PDF viewers. (Of course, the content of the files may include features not supported by older viewers, but at least the structure will be supported.) In generate mode, qpdf will create its own object streams. This will usually result in more compact PDF files, though they may not be readable by older viewers. In this mode, qpdf will also make sure the PDF version number in the header is at least 1.5.

Ordinarily, qpdf reads cross-reference streams when they are present in a PDF file. If --ignore-xref-streams is specified, qpdf will ignore any cross-reference streams for hybrid PDF files. The purpose of hybrid files is to make some content available to viewers that are not aware of cross-reference streams. It is almost never desirable to ignore them. The only time when you might want to use this feature is if you are testing creation of hybrid PDF files and wish to see how a PDF consumer that doesn't understand object and cross-reference streams would interpret such a file.

The --qdf flag turns on QDF mode, which changes some of the defaults described above. Specifically, in QDF mode, by default, stream data is uncompressed, content streams are normalized, and encryption is removed. These defaults can still be overridden by specifying the appropriate options as described above. Additionally, in QDF mode, stream lengths are stored as indirect objects, objects are laid out in a less efficient but more readable fashion, and the documents are interspersed with comments that make it easier for the user to find things and also make it possible for fix-qdf to work properly. QDF mode is intended for people, mostly developers, who wish to inspect or modify PDF files in a text editor. For details, please see Chapter 4, QDF Mode.

These options can be useful for digging into PDF files or for use in automated test suites for software that uses the qpdf library. When any of the options in this section are specified, no output file should be given. The following options are available:

The --raw-stream-data and --filtered-stream-data options are ignored unless --show-object is given. Either of these options will cause the stream data to be written to standard output. In order to avoid commingling of stream data with other output, it is recommend that these objects not be combined with other test/inspection options.

If --filtered-stream-data is given and --normalize-content=y is also given, qpdf will attempt to normalize the stream data as if it is a page content stream. This attempt will be made even if it is not a page content stream, in which case it will produce unusuable results.

This section was written prior to the implementation of the qpdf package and was subsequently modified to reflect the implementation. In some cases, for purposes of explanation, it may differ slightly from the actual implementation. As always, the source code and test suite are authoritative. Even if there are some errors, this document should serve as a road map to understanding how this code works.

In general, one should adhere strictly to a specification when writing but be liberal in reading. This way, the product of our software will be accepted by the widest range of other programs, and we will accept the widest range of input files. This library attempts to conform to that philosophy whenever possible but also aims to provide strict checking for people who want to validate PDF files. If you don't want to see warnings and are trying to write something that is tolerant, you can call setSuppressWarnings(true). If you want to fail on the first error, you can call setAttemptRecovery(false). The default behavior is to generating warnings for recoverable problems. Note that recovery will not always produce the desired results even if it is able to get through the file. Unlike most other PDF files that produce generic warnings such as “This file is damaged,”, qpdf generally issues a detailed error message that would be most useful to a PDF developer. This is by design as there seems to be a shortage of PDF validation tools out there. (This was, in fact, one of the major motivations behind the initial creation of qpdf.)

The QPDF package includes support for reading and rewriting PDF files. It aims to hide from the user details involving object locations, modified (appended) PDF files, the directness/indirectness of objects, and stream filters including encryption. It does not aim to hide knowledge of the object hierarchy or content stream contents. Put another way, a user of the qpdf library is expected to have knowledge about how PDF files work, but is not expected to have to keep track of bookkeeping details such as file positions.

A user of the library never has to care whether an object is direct or indirect. All access to objects deals with this transparently. All memory management details are also handled by the library.

The PointerHolder object is used internally by the library to deal with memory management. This is basically a smart pointer object very similar in spirit to the Boost library's shared_ptr object, but predating it by several years. This library also makes use of a technique for giving fine-grained access to methods in one class to other classes by using public subclasses with friends and only private members that in turn call private methods of the containing class. See QPDFObjectHandle::Factory as an example.

The top-level qpdf class is QPDF. A QPDF object represents a PDF file. The library provides methods for both accessing and mutating PDF files.

QPDFObject is the basic PDF Object class. It is an abstract base class from which are derived classes for each type of PDF object. Clients do not interact with Objects directly but instead interact with QPDFObjectHandle.

QPDFObjectHandle contains PointerHolder<QPDFObject> and includes accessor methods that are type-safe proxies to the methods of the derived object classes as well as methods for querying object types. They can be passed around by value, copied, stored in containers, etc. with very low overhead. Instances of QPDFObjectHandle always contain a reference back to the QPDF object from which they were created. A QPDFObjectHandle may be direct or indirect. If indirect, the QPDFObject the PointerHolder initially points to is a null pointer. In this case, the first attempt to access the underlying QPDFObject will result in the QPDFObject being resolved via a call to the referenced QPDF instance. This makes it essentially impossible to make coding errors in which certain things will work for some PDF files and not for others based on which objects are direct and which objects are indirect.

There is no public interface for creating instances of QPDFObjectHandle. They can be created only inside the QPDF library. This is generally done through a call to the private method QPDF::readObject which uses QPDFTokenizer to read an indirect object at a given file position and return a QPDFObjectHandle that encapsulates it. There are also internal methods to create fabricated indirect objects from existing direct objects or to change an indirect object into a direct object, though these steps are not performed except to support rewriting.

When the QPDF class creates a new object, it dynamically allocates the appropriate type of QPDFObject and immediately hands the pointer to an instance of QPDFObjectHandle. The parser reads a token from the current file position. If the token is a not either a dictionary or array opener, an object is immediately constructed from the single token and the parser returns. Otherwise, the parser is invoked recursively in a special mode in which it accumulates objects until it finds a balancing closer. During this process, the “R” keyword is recognized and an indirect QPDFObjectHandle may be constructed.

The QPDF::resolve() method, which is used to resolve an indirect object, may be invoked from the QPDFObjectHandle class. It first checks a cache to see whether this object has already been read. If not, it reads the object from the PDF file and caches it. It the returns the resulting QPDFObjectHandle. The calling object handle then replaces its PointerHolder<QDFObject> with the one from the newly returned QPDFObjectHandle. In this way, only a single copy of any direct object need exist and clients can access objects transparently without knowing caring whether they are direct or indirect objects. Additionally, no object is ever read from the file more than once. That means that only the portions of the PDF file that are actually needed are ever read from the input file, thus allowing the qpdf package to take advantage of this important design goal of PDF files.

If the requested object is inside of an object stream, the object stream itself is first read into memory. Then the tokenizer reads objects from the memory stream based on the offset information stored in the stream. Those individual objects are cached, after which the temporary buffer holding the object stream contents are discarded. In this way, the first time an object in an object stream is requested, all objects in the stream are cached.

An instance of QPDF is constructed by using the class's default constructor. If desired, the QPDF object may be configured with various methods that change its default behavior. Then the QPDF::processFile() method is passed the name of a PDF file, which permanently associates the file with that QPDF object. A password may also be given for access to password-protected files. QPDF does not enforce encryption parameters and will treat user and owner passwords equivalently. Either password may be used to access an encrypted file. ^[1] QPDF will allow recovery of a user password given an owner password. The input PDF file must be seekable. (Output files written by QPDFWriter need not be seekable, even when creating linearized files.) During construction, QPDF validates the PDF file's header, and then reads the cross reference tables and trailer dictionaries. The QPDF class keeps only the first trailer dictionary though it does read all of them so it can check the /Prev key. QPDF class users may request the root object and the trailer dictionary specifically. The cross reference table is kept private. Objects may then be requested by number of by walking the object tree.

When a PDF file has a cross-reference stream instead of a cross-reference table and trailer, requesting the document's trailer dictionary returns the stream dictionary from the cross-reference stream instead.

There are some convenience routines for very common operations such as walking the page tree and returning a vector of all page objects. For full details, please see the header file QPDF.hh.

The following example should clarify how QPDF processes a simple file.

Encryption is supported transparently by qpdf. When opening a PDF file, if an encryption dictionary exists, the QPDF object processes this dictionary using the password (if any) provided. The primary decryption key is computed and cached. No further access is made to the encryption dictionary after that time. When an object is read from a file, the object ID and generation of the object in which it is contained is always known. Using this information along with the stored encryption key, all stream and string objects are transparently decrypted. Raw encrypted objects are never stored in memory. This way, nothing in the library ever has to know or care whether it is reading an encrypted file.

An interface is also provided for writing encrypted streams and strings given an encryption key. This is used by QPDFWriter when it rewrites encrypted files.

The qpdf library supports file writing of QPDF objects to PDF files through the QPDFWriter class. The QPDFWriter class has two writing modes: one for non-linearized files, and one for linearized files. See Chapter 7, Linearization for a description of linearization is implemented. This section describes how we write non-linearized files including the creation of QDF files (see Chapter 4, QDF Mode.

This outline was written prior to implementation and is not exactly accurate, but it provides a correct “notional” idea of how writing works. Look at the code in QPDFWriter for exact details.

Once we have finished the queue, all referenced objects will have been written out and all deleted objects or unreferenced objects will have been skipped. The new cross-reference table will contain an offset for every new object number from 1 up to the number of objects written. This can be used to write out a new xref table. Finally we can write out the trailer dictionary with appropriately computed /ID (see spec, 8.3, File Identifiers), the cross reference table offset, and %%EOF.

Support for streams is implemented through the Pipeline interface which was designed for this package.

When reading streams, create a series of Pipeline objects. The Pipeline abstract base requires implementation write() and finish() and provides an implementation of getNext(). Each pipeline object, upon receiving data, does whatever it is going to do and then writes the data (possibly modified) to its successor. Alternatively, a pipeline may be an end-of-the-line pipeline that does something like store its output to a file or a memory buffer ignoring a successor. For additional details, look at Pipeline.hh.

QPDF can read raw or filtered streams. When reading a filtered stream, the QPDF class creates a Pipeline object for one of each appropriate filter object and chains them together. The last filter should write to whatever type of output is required. The QPDF class has an interface to write raw or filtered stream contents to a given pipeline.

To avoid the incestuous problem of having the qpdf library validate its own linearized files, we have a special linearized file checking mode which can be invoked via qpdf --check-linearization (or qpdf --check). This mode reads the linearization parameter dictionary and the hint streams and validates that object ordering, parameters, and hint stream contents are correct. The validation code was first tested against linearized files created by external tools (Acrobat and pdlin) and then used to validate files created by QPDFWriter itself.

Before creating a linearized PDF file from any other PDF file, the PDF file must be altered such that all page attributes are propagated down to the page level (and not inherited from parents in the /Pages tree). We also have to know which objects refer to which other objects, being concerned with page boundaries and a few other cases. We refer to this part of preparing the PDF file as optimization, discussed in Section 7.3, “Optimization”. Note the, in this context, the term optimization is a qpdf term, and the term linearization is a term from the PDF specification. Do not be confused by the fact that many applications refer to linearization as optimization or web optimization.

When creating linearized PDF files from optimized PDF files, there are really only a few issues that need to be dealt with:

In order to perform various operations such as linearization and splitting files into pages, it is necessary to know which objects are referenced by which pages, page thumbnails, and root and trailer dictionary keys. It is also necessary to ensure that all page-level attributes appear directly at the page level and are not inherited from parents in the pages tree.

We refer to the process of enforcing these constraints as optimization. As mentioned above, note that some applications refer to linearization as optimization. Although this optimization was initially motivated by the need to create linearized files, we are using these terms separately.

PDF file optimization is implemented in the QPDF_optimization.cc source file. That file is richly commented and serves as the primary reference for the optimization process.

After optimization has been completed, the private member variables obj_user_to_objects and object_to_obj_users in QPDF have been populated. Any object that has more than one value in the object_to_obj_users table is shared. Any object that has exactly one value in the object_to_obj_users table is private. To find all the private objects in a page or a trailer or root dictionary key, one merely has make this determination for each element in the obj_user_to_objects table for the given page or key.

Note that pages and thumbnails have different object user types, so the above test on a page will not include objects referenced by the page's thumbnail dictionary and nothing else.

We will create files with only primary hint streams. We will never write overflow hint streams. (As of PDF version 1.4, Acrobat doesn't either, and they are never necessary.) The hint streams contain offset information to objects that point to where they would be if the hint stream were not present. This means that we have to calculate all object positions before we can generate and write the hint table. This means that we have to generate the file in two passes. To make this reliable, QPDFWriter in linearization mode invokes exactly the same code twice to write the file to a pipeline.

In the first pass, the target pipeline is a count pipeline chained to a discard pipeline. The count pipeline simply passes its data through to the next pipeline in the chain but can return the number of bytes passed through it at any intermediate point. The discard pipeline is an end of line pipeline that just throws its data away. The hint stream is not written and dummy values with adequate padding are stored in the first cross reference table, linearization parameter dictionary, and /Prev key of the first trailer dictionary. All the offset, length, object renumbering information, and anything else we need for the second pass is stored.

At the end of the first pass, this information is passed to the QPDF class which constructs a compressed hint stream in a memory buffer and returns it. QPDFWriter uses this information to write a complete hint stream object into a memory buffer. At this point, the length of the hint stream is known.

In the second pass, the end of the pipeline chain is a regular file instead of a discard pipeline, and we have known values for all the offsets and lengths that we didn't have in the first pass. We have to adjust offsets that appear after the start of the hint stream by the length of the hint stream, which is known. Anything that is of variable length is padded, with the padding code surrounding any writing code that differs in the two passes. This ensures that changes to the way things are represented never results in offsets that were gathered during the first pass becoming incorrect for the second pass.

Using this strategy, we can write linearized files to a non-seekable output stream with only a single pass to disk or wherever the output is going.

Once a file is optimized, we have information about which objects access which other objects. We can then process these tables to decide which part (as described in “Linearized PDF Document Structure” in the PDF specification) each object is contained within. This tells us the exact order in which objects are written. The QPDFWriter class asks for this information and enqueues objects for writing in the proper order. It also turns on a check that causes an exception to be thrown if an object is encountered that has not already been queued. (This could happen only if there were a bug in the traversal code used to calculate the linearization data.)

There are a handful of known issues with this linearization code. These issues do not appear to impact the behavior of linearized files which still work as intended: it is possible for a web browser to begin to display them before they are fully downloaded. In fact, it seems that various other programs that create linearized files have many of these same issues. These items make reference to terminology used in the linearization appendix of the PDF specification.

The qpdf --show-linearization command can show the complete contents of linearization hint streams. To look at the raw data, you can extract the filtered contents of the linearization hint tables using qpdf --show-object=n --filtered-stream-data. Then, to convert this into a bit stream (since linearization tables are bit streams written without regard to byte boundaries), you can pipe the resulting data through the following perl code:

use bytes;
binmode STDIN;
undef $/;
my $a = <STDIN>;
my @ch = split(//, $a);
map { printf("%08b", ord($_)) } @ch;
print "\n";

Object streams can contain any regular object except the following:

In addition, Adobe reader (at least as of version 8.0.0) appears to not be able to handle having the document catalog appear in an object stream if the file is encrypted, though this is not specifically disallowed by the specification.

There are additional restrictions for linearized files. See Section 8.3, “Implications for Linearized Files”for details.

The PDF specification refers to objects in object streams as “compressed objects” regardless of whether the object stream is compressed.

The generation number of every object in an object stream must be zero. It is possible to delete and replace an object in an object stream with a regular object.

The object stream dictionary has the following keys:

Stream collections are formed with /Extends. They must form a directed acyclic graph. These can be used for semantic information and are not meaningful to the PDF document's syntactic structure. Although qpdf preserves stream collections, it never generates them and doesn't make use of this information in any way.

The specification recommends limiting the number of objects in object stream for efficiency in reading and decoding. Acrobat 6 uses no more than objects per object stream for linearized files and no more 200 objects per stream for non-linearized files. QPDFWriter, in object stream generation mode, never puts more than 100 objects in an object stream.

Object stream contents consists of N pairs of integers, each of which is the object number and the byte offset of the object relative to the first object in the stream, followed by the objects themselves, concatenated.

For non-hybrid files, the value following startxref is the byte offset to the xref stream rather than the word xref.

For hybrid files (files containing both xref tables and cross-reference streams), the xref table's trailer dictionary contains the key /XRefStm whose value is the byte offset to a cross-reference stream that supplements the xref table. A PDF 1.5-compliant application should read the xref table first. Then it should replace any object that it has already seen with any defined in the xref stream. Then it should follow any /Prev pointer in the original xref table's trailer dictionary. The specification is not clear about what should be done, if anything, with a /Prev pointer in the xref stream referenced by an xref table. The QPDF class ignores it, which is probably reasonable since, if this case were to appear for any sensible PDF file, the previous xref table would probably have a corresponding /XRefStm pointer of its own. For example, if a hybrid file were appended, the appended section would have its own xref table and /XRefStm. The appended xref table would point to the previous xref table which would point the /XRefStm, meaning that the new /XRefStm doesn't have to point to it.

Since xref streams must be read very early, they may not be encrypted, and the may not contain indirect objects for keys required to read them, which are these:

The other fields in the xref stream, which may be indirect if desired, are the union of those from the xref table's trailer dictionary.

For linearized files, the linearization dictionary, document catalog, and page objects may not be contained in object streams.

Objects stored within object streams are given the highest range of object numbers within the main and first-page cross-reference sections.

It is okay to use cross-reference streams in place of regular xref tables. There are on special considerations.

Hint data refers to object streams themselves, not the objects in the streams. Shared object references should also be made to the object streams. There are no reference in any hint tables to the object numbers of compressed objects (objects within object streams).

When numbering objects, all shared objects within both the first and second halves of the linearized files must be numbered consecutively after all normal uncompressed objects in that half.

There are three modes for writing object streams: disable, preserve, and generate. In disable mode, we do not generate any object streams, and we also generate an xref table rather than xref streams. This can be used to generate PDF files that are viewable with older readers. In preserve mode, we write object streams such that written object streams contain the same objects and /Extends relationships as in the original file. This is equal to disable if the file has no object streams. In generate, we create object streams ourselves by grouping objects that are allowed in object streams together in sets of no more than 100 objects. We also ensure that the PDF version is at least 1.5 in generate mode, but we preserve the version header in the other modes. The default is preserve.

We do not support creation of hybrid files. When we write files, even in preserve mode, we will lose any xref tables and merge any appended sections.