Filtered inline image data has no length, so its end is an EI chosen
among candidates, and the content parser chose by a different rule
than the renderer's tokenizer (hayro-syntax): the renderer also takes an
EI with no whitespace before it, and checks a candidate by reading what
follows as operators. With image data such as "7FEI 1 0 0 1 200 0 cm",
the renderer ran a cm that redaction read as image data, so text was
drawn 200 pt from where removal and its verify pass looked.
The parser now takes each inline image's end from the renderer's
tokenizer (already in the tree through pdfcraft-cos), so both read the
same operators after the image; a test runs both on the same fixtures.
Only where that tokenizer reads no image (no EI it accepts, or
parameters it can't read) does the parser fall back to its own rules,
counting the image in Parsed::ambiguous_images, and applying redactions
refuses such a page as unreadable.
Structure cleanup cleared alternate text, actual text and expansions
only on the element that directly owned changed marked content, and an
element that lost an annotation reference kept its descriptions unless
it lost every kid. An ancestor's /ActualText covering a redacted span,
or a link element's /Alt repeating the address of a removed link, stayed
in the saved file. Elements that lose an annotation now always lose
their descriptions, and every ancestor of an affected element (found
through both /P and the tree's /K, which may disagree) loses its own.
An element written in place above one with a description to clear
refuses the operation, as other in-place cases already do.
Finding which marked content a redaction changed copied every operator
into the buffer of each MCID enclosing it, so content nested n deep was
held n times: a modest stream of nested BDCs could exhaust memory.
Each operator inside marked content is now serialized once, and each
MCID keeps the ranges its sequences span. The bytes compared are capped;
MCIDs past the cap count as changed, so their descriptions are cleared
rather than kept.
Clearing the pixels of a partly covered image decoded its stream with
the global 1 GiB allowance and only then cut it to the raster length its
Width, Height and bit depth declare. A small Flate stream could make
every such image allocate and inflate far more than it can use. Decoding
now stops at twice the raster plus a byte per row (filter-chain and
predictor overhead) and a little slack; an image past that is removed
whole, as an undecodable one already is.
After applying redactions, removed images and forms were unlinked from
page-tree resource dictionaries but not from the resources of forms the
page still draws. A form sharing the page's original resource dictionary
(a common producer layout) kept the removed image reachable, so a full
save wrote it. The walk over surviving forms now drops XObject entries
for removed objects that the form's own content does not draw, and fails
closed when that cannot be determined.
The redaction interpreter started every form XObject with an empty text
state (no font, size 0, no spacing) although a form inherits its caller's
graphics state, and it ignored the /Font entry of an ExtGState set by gs.
Text shown that way was placed at the wrong positions, so removal and
the verify pass could both miss glyphs under a mark. Forms now start
from the caller's state, and gs sets the font and size as Tf does.
The engine's renderer cross-check test used the gs case as its example
of interpreter/renderer disagreement; it now runs the check directly on
an unredacted page, which also covers an input stream labelled as a
redaction overlay.
Both checks for text left under applied redaction marks (redaction's own
verify pass and the engine's renderer cross-check) skipped any page
content stream whose dictionary said /PCMark /Redaction, so that the
overlay's label would not count. An input file could put that label on
its own content stream and keep it out of both checks. Applying
redactions now records the overlay streams it adds (stamp returns the
new stream), and only those are skipped.
Deleting a field took it off its pages and out of the field tree but left
its entry in the form's /CO array. That reference kept the field
dictionary, its value and its widgets reachable, so a full save after
applying redactions over the field still wrote the redacted value.
A run of images that is not one picture was searched again from each of
its groups, collecting the rest of the run every time, so many images drawn
over one another made a page quadratic. The search reads at most 16 groups
per content operator (at least 65 536) and leaves the rest of the page as
it is once that is spent.
The pixel reader capped the image size the dictionary gives, then decoded
the JPEG at the size its header gives and compared afterwards, so a small
image dictionary over a huge JPEG allocated the huge one first. The reader
takes the size from the JPEG header and refuses a mismatch before decoding.
The body-size tally and the RTF colour table searched every slot seen so
far for each block, and the RTF writer searched the colour table again for
every paragraph and cell, so a document with many distinct sizes or colours
made export quadratic. Both keep their first-seen order and look slots up
in a hash map.
Line wrapping built and measured the whole candidate line for every word,
so a long field value or annotation text at a tiny font size, with many
words on one line, took quadratic time and copying. The wrapper keeps the
line's width as it goes and measures each word, space and broken
character once.
Code-to-CID and /W range lookups scanned every record in declaration
order, once per glyph and once per explicit ToUnicode entry while a font
was read, so a few thousand records multiplied by as many codes took
billions of steps. Both tables are indexed once into stretches whose
first-declared covering record wins, the precedence the scan had, and the
/W ranges stop at the CMap entry limit.
A sequential bfrange stores one destination and builds every covered
glyph's text from all of it, so one huge destination and many repeats of
its codes made text extraction copy megabytes per glyph. Destinations keep
at most 256 UTF-16 units (the format's 512 bytes), and marked-text
extraction stops asking for a glyph's text once its id has all the text it
keeps.
Figure listing decoded and parsed a page's content once per figure on it,
so many small Figure elements pointing at one page multiplied the work.
The figures are grouped by page and one pass gathers the bounds of every
marked-content id they own; what a sequence draws also counts for the
sequences around it, as before.
Marked-text extraction and figure bounds searched the whole marked-content
stack before every operator, so deeply nested unmarked BMC blocks made a page
quadratic. The text stack now holds each level's effective id (its own, else
the enclosing one's), and the bounds pass counts the open sequences that
match.
page_form listed every page of the document to take the one it
printed, so imposing P pages walked the page tree and copied its leaf
dictionaries P times over: P^2 page copies. impose passes the list it
already holds. A page index outside it is an error, not a panic.
Each unmarked redraw of an /ActualText span's glyphs took its own copy
of the replacement text. The work budget counts comparisons, not
bytes, so a long replacement around one glyph followed by many redraws
at the same place copied it once per redraw: gigabytes from a small
page. Redraws on one page share 4 MiB of replacement text; past it a
redraw keeps its own glyphs, as it does past the work budget.
Each leaf page gets its own copy of every attribute it inherits, so a
large direct /Resources on the root of a page tree with thousands of
leaves was copied thousands of times: a compressed file of a few
kilobytes could need tens of gigabytes. The walk weighs each inherited
value once where it is declared and charges every copy to a 1 GiB
budget; past it the walk is cut off, as an over-large tree already is.
Every string and stream decrypted searched the document's /CF filters
for the method by name. A file controls both the number of filters and
the number of strings, so the search multiplied the two when the
filter in use was listed last. The handler resolves the string, stream
and embedded-file methods when it is built, and looks named filters up
in a map.
The materialized cross-reference reader reread a section's /XRefStm
for every revision that named it, decoding the stream and inserting
its rows each time, although the first read had already merged every
entry. Up to 10,000 revisions naming one large stream repeated that
work 10,000 times. A stream already read is skipped, with one repair
log line.
A stream whose /Length is wrong or missing is ended at the next
endstream, which the parser found by searching the rest of the file.
Many such streams before one late endstream (a damaged file being
reconstructed reads every one) cost the file size per stream. The
document lists where endstream occurs the first time it needs it, and
each stream looks its end up there.
The object caches weigh what each cached object owns, but the walk
stopped four levels down and counted anything deeper as nothing. Many
objects holding a large string inside a few nested arrays stayed cached
far past the byte budget (34 MiB against 32 MiB in the test, and without
limit as objects grow). The walk now goes past the parser's nesting
limit; anything deeper counts as too large to keep.
Stream::filters reads only a direct name or array of names, so an
indirect /Filter (or an array holding a reference) gave an empty chain
and every decode returned the encoded bytes as if they were the data:
a strict decode reported success on content it never decoded, while a
viewer resolving the reference shows something else. Decoding such a
stream is an error, so callers keep it encoded or refuse it.
Every stage of a stream's filter chain could produce up to the output
limit and nothing bounded the number of stages, so a long /Filter array
over data that decodes to itself (FF FF under RunLengthDecode) repeated
the full work once per array entry. Chains of more than eight filters
are unsupported, like an image codec, and stay encoded.
Pieces::splice copied the unedited bytes between edits by visiting
every piece for every edit, so a page of many content pieces and many
replaced operators cost pieces x edits steps. The copy starts at the
piece holding its first byte and stops past its last one.
When an operand failed to parse, the content parser went back to the
byte after its start and tried again, so a run of n bytes that never
reads as a number (a long run of '+', or digits ending in a letter) was
read again from every one of its bytes: n^2/2 bytes from a tiny
compressed stream. A failed number token longer than 64 bytes is now
skipped whole; shorter ones still recover byte by byte.
To read the longest valid prefix of a token like 1.2.3 the parser
retried the conversion on each shorter prefix, so a long digit run
followed by long junk was converted once per junk byte: quadratic in
the token. The prefix is found in one scan and converted once. A
prefix that spells infinity or NaN no longer reads as a number.
Dict::set searched every existing entry before inserting a key, so a
parsed dictionary of n distinct keys cost n^2/2 key comparisons: a
few megabytes of keys stalled parsing for minutes. From 32 entries a
dictionary keeps a key index, so each insert and lookup is one hash
probe. Equality and debug output still compare and show the entries
alone.
The name-tree count cloned each node's /Kids array and skipped repeated
references without charging the budget, so a node listing the same kid
many times cost work proportional to every entry while the budget saw one
step. The walk now borrows the array, charges every entry it does not
descend into, and stops once the budget is spent.
Metadata, private application data, associated files and additional
actions were removed from the catalog and pages only. The same entries on
an image or font (XMP), a Form XObject (PieceInfo), a page or annotation
(/AF with its embedded file) or a form field (/AA scripts) survived the
full rewrite, and the scan did not count them, so removing only such a
category reported nothing to remove. A walk of every object reachable from
the catalog now counts and strips these entries wherever they are, leaving
name maps such as resource dictionaries alone (their keys are names a
producer chose), and refuses a graph nested too deeply to walk in full.
Removing hidden text or hidden layers rewrote page and Form content but
left the structure tree alone, unlike applying redactions. An element's
ActualText, Alt or expansion text still described the removed content,
and a marked-content reference into a Form kept the original Form, with
the removed text, in the full rewrite. The hidden-content pass now cleans
the tags of what each page no longer shows and moves Form references to
the rewritten copy, refusing when the tree can't be cleaned completely.
Removing form fields flattens them and then drops the interactive form,
but flattening keeps hidden and no-view widgets (they draw nothing) and
widgets with no appearance. The pages still listed them, so their values
and their parent fields' values survived the full rewrite. Every widget
left after flattening now leaves the pages and the structure tree, inline
annotations included.
Page and Form content was decoded tolerantly, keeping what came before a
corrupt spot and dropping the rest without saying so. A page whose content
broke at its first bytes therefore read as empty: the hidden-text scan
counted nothing, the removal was skipped and the original stream, which
other viewers still draw, stayed in the file. Content is now decoded
strictly; an unreadable page counts as something to remove, so the removal
runs and refuses the page, and a Form read only in part is treated like an
undecodable one.
The hidden-information count followed every /Kids entry of the attachment
and document-script name trees with only a depth limit, so a node listed
twice by each level was read once per path: a few hundred bytes made the
scan take 2^32 steps. Indirect nodes are now read once each, under a total
node budget.
The /Rect fallback dropped non-numbers before checking the length, so a
malformed /Rect such as [0 0 /Bad 100 100] read as a valid rectangle and
gave the mark quads. The array itself must now hold exactly four finite
numbers.
A quad whose corners all lie on one line (or on one point) encloses
nothing, yet such /QuadPoints counted as drawable, replaced the mark's
/Rect and left an imported redaction that covers nothing. Any quad without
area now makes the mark's quads unusable, so its /Rect is used instead.
Merge Data Files into Spreadsheet writes field names and values from other
people's files as CSV cells, escaping only the delimiters. A cell starting
with =, +, - or @ is evaluated as a formula when the CSV is opened in a
spreadsheet program. Such cells, after any leading whitespace, are now
written with a leading apostrophe; finite numbers such as -12.5 are left
as they are.
A redaction mark's outline is drawn from its /QuadPoints, but applying the
mark falls back to its /Rect when it has none. An imported mark with only a
/Rect, or with quads the outline can't draw, therefore had no visible
outline while still removing content when applied, and an imported opacity
of 0 hid the outline too. Imported marks now get quads from their /Rect
when their own are unusable (or none, so they remove nothing), and their
/CA and /ca are dropped.
Comment import replaces an existing annotation with the same /NM in place.
The check for annotations it must keep covered locked comments and
redaction marks but not widgets or field dictionaries, so an imported
comment could overwrite a field's widget object, setting its value and
validation entries without the form value setter while the AcroForm field
list still pointed at it. Widgets and field dictionaries are now kept, and
the imported comment is added beside them under no name.
The fragment merge multiplied the merged width by a height cast from a
float without checking either fit. A tiny fragment rectangle gives a huge
height: the multiplication overflowed (a panic in debug builds, a wrap past
the pixel limit in release builds) and the u32 cast saturated, sizing an
enormous allocation. The height must now fit u32 before the pixel limit is
checked, and the buffer size uses checked multiplication.
aiongg