| 1 |
Hello, |
| 2 |
|
| 3 |
So I provided a patch for a sandbox bug hitting bigger projects using |
| 4 |
-export-symbols-regex with a long list of object files. 3 months ago. |
| 5 |
Bug has been there since forever, reported 15 months ago, with some |
| 6 |
good clues to what's up since 9 months. |
| 7 |
It has been sitting there, collecting dust, with no action from |
| 8 |
sandbox@ whatsoever. As such, I plan to finally non-maintainer push |
| 9 |
this fix straight to ~arch as a sandbox-2.10 revision bump once I have |
| 10 |
my months old GPG machine tree and system updated (this week or early |
| 11 |
next week). And 2.11, but because that is still p.masked due to it |
| 12 |
causing issues for XUL stuff (with analysis of what's going on also |
| 13 |
available since a while now), that's going to be a p.masked revbump |
| 14 |
alongside the 2.11 masks. |
| 15 |
If I can't do my gnome-builder bumps that depends on this right away, I |
| 16 |
might let it simmer in p.mask for some hours or days too, especially if |
| 17 |
I see some sort of sandbox@ action appearing or some valid objections |
| 18 |
by the time I get to it. |
| 19 |
|
| 20 |
This is the bug I have fix for: |
| 21 |
https://bugs.gentoo.org/show_bug.cgi?id=553092 |
| 22 |
|
| 23 |
libtool ends up running "nm -B" with the long list of object files as |
| 24 |
arguments and saves the result in a temporary file (which it'll apply |
| 25 |
the regex to then), but various shells in some environments (including |
| 26 |
bash-4.3 and dash) end up trying to glob it and check if it's a dir, |
| 27 |
calling opendir with the whole commandline as argument. If that is |
| 28 |
longer than 8196 characters, sandbox gets confused because it |
| 29 |
internally uses PATH_MAX*2 buffers, it gets cut and things fall over in |
| 30 |
ways I'm not interested in finding out deeper. |
| 31 |
|
| 32 |
At least gnome-builder-3.20+ and graphicsmagick are affected for some |
| 33 |
(might depend on what their shell is doing). |
| 34 |
|
| 35 |
Because of this, gnome-builder hasn't seen version bumps, while the |
| 36 |
existing version in tree (3.18, it didn't use so many object files in |
| 37 |
the linker line quite yet back then to trigger the bug) are completely |
| 38 |
unusable with current stable gtksourceview and co. |
| 39 |
|
| 40 |
So, any objections with me pushing in the sandbox revbumps? |
| 41 |
|
| 42 |
|
| 43 |
PS: I'm sure our mozilla team would appreciate also help with sandbox |
| 44 |
bug 580726, which is a bug in the ptrace fallback, which now gets |
| 45 |
triggered with the p.masked sandbox 2.11 due to some inherent issues |
| 46 |
with the default non-ptrace code that were hit in Chrome OS project |
| 47 |
thing doing some own memory management (and so it fallbacks more often, |
| 48 |
when it finds custom memory allocation stuff based on some heuristics). |
| 49 |
The ptrace fallback gets now used with 2.11 for firefox and co as well |
| 50 |
(probably due to jemalloc usage), and that fallback sandbox codepath is |
| 51 |
apparently buggy for its more complex case. Alternatively maybe these |
| 52 |
heuristics could be less triggerhappy to fallback to ptrace. |
| 53 |
|
| 54 |
|
| 55 |
Mart |
Archives