fix(tokenizers): start a stream's data after a lone CR, as readers do

The walk only recognised `stream` followed by LF or CRLF. After a lone CR
(or trailing spaces) the stream data was read as object syntax, so bare
tokens in a content stream such as
`ET endobj 2 0 obj << /Type /Pages /Count 9999 >> endobj` redefined the
page tree: pypdf strict mode opens that file at 1 page, Headroom priced
9999. The keyword now takes any spaces and an optional CR, CRLF or LF,
the same data start pypdf uses, so the /Length check covers these
streams too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Garm
2026-10-24 10:58:39 +02:00
co-authored by Claude Opus 5.5
parent 055e3617b0
commit adcc4f1d83
2 changed files with 17 additions and 1 deletions
+3 -1
View File
@@ -54,8 +54,10 @@ _PDF_MAX_XREF_SECTIONS = 1_000 # bounds the /Prev chain an indirect /Length is
_PDF_TOP_LEVEL = re.compile(
rb"(?=[%\dt])(?:%[^\r\n]*|(?<!\d)(\d{1,10})\s+\d{1,5}\s+obj\b|(?<![A-Za-z/])trailer\b)"
)
# A stream's data starts after its keyword, any spaces and an EOL if present (a
# lone CR too): the spec requires CRLF or LF, but readers accept the others.
_PDF_SYNTAX_STOP = re.compile(
rb"(?=[%(se])(?:%[^\r\n]*|\(|(?<![A-Za-z/])(?:stream\r?\n|endobj|startxref))"
rb"(?=[%(se])(?:%[^\r\n]*|\(|(?<![A-Za-z/])(?:stream *(?:\r\n?|\n)?|endobj|startxref))"
)
_PDF_STRING_TOKEN = re.compile(rb"\\.|[()]", re.S)
_PDF_STREAM_LENGTH = re.compile(rb"/Length\s+(\d{1,10})(?!\d|\s+\d+\s+R)")
@@ -346,6 +346,7 @@ def _packed(objects: dict[int, bytes], number: int) -> dict[int, bytes | dict[in
_INDIRECT = _one_page_objects(_FAKE_TREE, length=b"/Length 6 0 R")
_LONE_CR = b"BT /F1 12 Tf 72 720 Td (hi) Tj ET endobj 2 0 obj << /Type /Pages /Count 9999 >> endobj"
@pytest.mark.parametrize(
@@ -363,6 +364,18 @@ _INDIRECT = _one_page_objects(_FAKE_TREE, length=b"/Length 6 0 R")
# Through an xref stream, and through /Prev to the section that lists object 6.
(_pdf_file(_packed(_INDIRECT, 5)), 1),
(_pdf_file({5: _INDIRECT[5]}, prior=_pdf_file(_INDIRECT)), 1),
# A lone CR after the stream keyword still starts a stream, so bare tokens after
# the text operators stay data.
(
_pdf_file(
{
**_INDIRECT,
4: b"<< /Length 6 0 R >>\nstream\r%s\nendstream" % _LONE_CR,
6: b"%d" % len(_LONE_CR),
}
),
1,
),
# A stream end that cannot be verified stops the walk, and 0 means the size
# estimate: the length packed in an object stream, a wrong one, a reference to
# no object, no /Length at all.
@@ -376,6 +389,7 @@ _INDIRECT = _one_page_objects(_FAKE_TREE, length=b"/Length 6 0 R")
"fake-length-object",
"xref-stream",
"prev-section",
"lone-cr",
"packed-length",
"wrong-length",
"dangling-length",