Bug#1148398: poppler: CVE-2026-93653 - tested candidate patch (bounds tiling-pattern tile count), rebased for trixie and bookworm

Sumit Verma automationcanadian at gmail.com
Fri Nov 6 04:00:22 GMT 2026


Hi,

Here is a candidate fix for CVE-2026-93653 (CPU-exhaustion DoS in tiling
pattern fills). It is still open in all suites and marked no-dsa/minor, so
this is a patch for a future maintainer upload rather than a request for a
DSA.

Root cause: Gfx::doTilingPatternFill() derives the per-axis tile index
ranges (xi0..xi1, yi0..yi1) from the clip region divided by the pattern
step. A pattern whose step is tiny relative to the filled region therefore
yields a tile grid bounded only by the page geometry, up to roughly INT_MAX
on each axis. Both the Splash tiling fast path (through the x1-x0 / y1-y0
arguments of OutputDev::tilingPatternFill) and the drawForm() fallback loop
iterate the whole grid, so a small crafted PDF keeps the renderer at 100%
CPU for an attacker-controlled duration while allocating almost no memory.
The checkedMultiply() guard added for CVE-2026-10118 only rejects integer
overflow inside SplashOutputDev, after which rendering falls back to this
equally unbounded loop.

Fix: cap the number of tiles at 0x800000 (the ceiling already used in
SplashOutputDev::tilingPatternFill), shrinking both axes proportionally.
Legitimate patterns render unchanged; an abusive grid renders a bounded
subset instead of hanging.

Attached (DEP-3):
CVE-2026-93653-poppler-trixie.patch applies to 25.03.0-5+deb13u4
CVE-2026-93653-poppler-bookworm.patch applies to 22.12.0-2+deb12u2

Both are the same one-hunk change to poppler/Gfx.cc and I verified each
applies cleanly to the packaged source. The inserted block uses only
symbols already present in Gfx.cc (error/errSyntaxWarning/getPos and sqrt
from <cmath>).

Testing: I build- and runtime-tested the identical change against current
upstream master. A crafted 707-byte PDF that otherwise pegs a core for
about 82 seconds returns in ~0.012 s with the cap; a normal tiling-pattern
PDF renders byte-for-byte identically (same output hash). I did not rebuild
each packaged version separately, since the change and its surrounding code
are the same.

The same fix was submitted to the poppler mailing list on 2026-11-06
(subject "[PATCH] Bound the tile count in tiling pattern fills"). sid and
forky (26.07.0) can use that upstream patch as-is; the trixie/bookworm
versions attached here differ only because those branches still use the
older (int) cast style around the fix site.

Regards,
Sumit
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://alioth-lists.debian.net/pipermail/pkg-freedesktop-maintainers/attachments/20261105/3d3da2df/attachment.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: CVE-2026-93653-poppler-bookworm.patch
Type: text/x-patch
Size: 2801 bytes
Desc: not available
URL: <http://alioth-lists.debian.net/pipermail/pkg-freedesktop-maintainers/attachments/20261105/3d3da2df/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: CVE-2026-93653-poppler-trixie.patch
Type: text/x-patch
Size: 2801 bytes
Desc: not available
URL: <http://alioth-lists.debian.net/pipermail/pkg-freedesktop-maintainers/attachments/20261105/3d3da2df/attachment-0001.bin>


More information about the Pkg-freedesktop-maintainers mailing list