Missing text in PDFs starting with nightly 2026-07-23
Categories
(Core :: Graphics: Canvas2D, defect)
Tracking
()
| Tracking | Status | |
|---|---|---|
| firefox-esr140 | --- | unaffected |
| firefox-esr153 | --- | unaffected |
| firefox153 | --- | unaffected |
| firefox154 | --- | unaffected |
| firefox155 | --- | affected |
People
(Reporter: aguertin+bugzilla, Unassigned)
References
(Regression)
Details
(Keywords: regression)
Attachments
(3 files)
Most text is missing from PDFs.
Attach (recommended) or Link to PDF file here: https://www.americanrivers.org/wp-content/uploads/2026/03/DamRemovalCompiledSummaries_2026.pdf
This affects every PDF I've tried - this is just an example.
Steps to reproduce the problem:
- Visit any PDF
- See text is missing
What is the expected behavior? (add screenshot)
Will attach
What went wrong? (add screenshot)
Screenshot attached.
Most text is missing. In many cases the first letter or first few letters of a paragraph are visible. Selecting the text shows it is still there, just not displayed.
2026-07-22-21-40-24 nightly worked fine.
2026-07-23-20-40-11 is the first bad version.
System is Gentoo Linux.
| Reporter | ||
Comment 1•3 months ago
|
||
Comment 2•3 months ago
|
||
Here's the pushlog for build 20260723204011:
https://hg-edge.mozilla.org/mozilla-central/pushloghtml?fromchange=ae4fce337439&tochange=6d21b1fa30fe
and I don't see what could be the regressor.
Myself I don't manage to reproduce your issue on Windows 11 and Ubuntu 26.04.
Lee, would you have any idea ?
| Reporter | ||
Comment 3•3 months ago
|
||
Adding that zooming the PDF affects which text is displayed. Further, text is only missing the first time a given zoom level is displayed; returning to a previously displayed zoom level will show all the text.
In other words:
- Load PDF, defaulting to zoom 100%: some text is missing
- Change zoom to 125%: some text is missing
- Change zoom back to 100%: all text displays
- Change zoom back to 125%: all text displays
- Change zoom to 150%: some text is missing
- etc
From the pushlog, could this be Bug 2056254, Bug 2056251, Bug 2056248, and/or Bug 2056246?
Comment 4•3 months ago
|
||
Can you please attach your about:support information?
Comment 5•3 months ago
|
||
I can't reproduce this on Debian Linux, so there must be something more particular about your configuration that is causing this. The about:support information would potentially give me some clue to what the causative factor might be.
Comment 6•3 months ago
|
||
If I had to guess it is possible either Bug 2056779 or Bug 2055839 could be involved, but as of yet I have no way to connect these bugs to the observed symptoms, other than that they are modifying some sensitive code in DrawTargetWebgl.
| Reporter | ||
Comment 7•3 months ago
|
||
| Reporter | ||
Updated•3 months ago
|
| Reporter | ||
Comment 8•3 months ago
|
||
Got myself set up to build so I could bisect
c547c203c14bc9fcffbdd3efff03b0efea2ec2fd is the first bad commit
commit c547c203c14bc9fcffbdd3efff03b0efea2ec2fd
Author: Jamie Nicol <jnicol@mozilla.com>
Date: Wed Jul 22 21:26:00 2026 +0000
Bug 2055839 - Use shared PBO for DrawTargetWebgl uploads. r=gfx-reviewers,lsalzman
See also Bug 2059433 comment 2, where the same problem seems to have been noted: "the PDF contents don't appear to always get drawn"
Updated•3 months ago
|
Comment 9•3 months ago
|
||
Set release status flags based on info from the regressing bug 2055839
:jnicol, since you are the author of the regressor, bug 2055839, could you take a look?
For more information, please visit BugBot documentation.
Comment 10•3 months ago
|
||
See also Bug 2059433 comment 2, where the same problem seems to have been noted: "the PDF contents don't appear to always get drawn"
What that comment is saying is that without the change in bug 2059433 the test pdf does not get rendered during the test because it's too slow. (Whereas with the change, it's fast enough that we render enough frames to fall back to unaccelerated canvas, which is slower on the main thread, so it does complete rendering before the test finishes). It's unrelated to this bug.
This for sure looks like a driver bug. The likely solution is to add && !mWebgl->GL()->IsMesa() to the condition here so that we avoid the shared PBO upload path on Mesa.
But I see in your about:support you're using a development version of mesa. Can you reproduce with a stable version?
| Reporter | ||
Comment 11•3 months ago
|
||
Good call, it does not reproduce in mesa-26.1.5, and also does not reproduce in the latest mesa development.
Bisecting mesa, I see it was fixed on main a week ago by mesa 479773c7e426: mesa/st: Invalidate FS sampler views after PBO texture transfers, and on the 26.2 branch by mesa 4432d3d3363a
It looks like it was caused by mesa 62efee18607 mesa/st: Convert st_cb_texture to use st_context_invalidate_state in April
4432d3d3363a was included in mesa-26.2.0-rc3, so the only people affected by this should be those using 26.2.0-rc1, 26.2.0-rc2, or out of date git snapshots.
Comment 12•3 months ago
|
||
Great, thanks for testing!
Description
•