Merge commit 'intrigeri/po' into po

1 files changed, 266 insertions, 23 deletions

diff --git a/doc/plugins/po.mdwn b/doc/plugins/po.mdwn
index 923ccd63a..544c7ef62 100644
--- a/doc/plugins/po.mdwn
+++ b/doc/plugins/po.mdwn
@@ -6,6 +6,8 @@ gettext, using [po4a](http://po4a.alioth.debian.org/).
 
 It depends on the Perl `Locale::Po4a::Po` library (`apt-get install po4a`).
 
+[[!toc levels=2]]
+
 Introduction
 ============
 
@@ -127,6 +129,10 @@ Usage
 Templates
 ---------
 
+When `po_link_to` is not set to `negotiated`, one should replace some
+occurrences of `BASEURL` with `HOMEPAGEURL` to get correct links to
+the wiki homepage.
+
 The `ISTRANSLATION` and `ISTRANSLATABLE` variables can be used to
 display things only on translatable or translation pages.
 
@@ -215,10 +221,229 @@ TODO
 Security checks
 ---------------
 
-- Can any sort of directives be put in po files that will
-  cause mischief (ie, include other files, run commands, crash gettext,
-  whatever).
-- Any security issues on running po4a on untrusted content?
+### Security history
+
+The only past security issues I could find in GNU gettext and po4a
+are:
+
+- [CVE-2004-0966](http://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2004-0966),
+  *i.e.* [Debian bug #278283](http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=278283):
+  the autopoint and gettextize scripts in the GNU gettext package
+  1.14 and later versions, as used in Trustix Secure Linux 1.5
+  through 2.1 and other operating systems, allows local users to
+  overwrite files via a symlink attack on temporary files.
+- [CVE-2007-4462](http://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2007-4462):
+  `lib/Locale/Po4a/Po.pm` in po4a before 0.32 allows local users to
+  overwrite arbitrary files via a symlink attack on the
+  gettextization.failed.po temporary file.
+
+**FIXME**: check whether this plugin would have been a possible attack
+vector to exploit these vulnerabilities.
+
+Depending on my mood, the lack of found security issues can either
+indicate that there are none, or reveal that no-one ever bothered to
+find (and publish) them.
+
+### PO file features
+
+Can any sort of directives be put in po files that will cause mischief
+(ie, include other files, run commands, crash gettext, whatever)?
+
+> No [documented](http://www.gnu.org/software/gettext/manual/gettext.html#PO-Files)
+> directive is supposed to do so. [[--intrigeri]]
+
+### Running po4a on untrusted content
+
+Are there any security issues on running po4a on untrusted content?
+
+To say the least, this issue is not well covered, at least publicly:
+
+- the documentation does not talk about it;
+- grep'ing the source code for `security` or `trust` gives no answer.
+
+On the other hand, a po4a developer answered my questions in
+a convincing manner, stating that processing untrusted content was not
+an initial goal, and analysing in detail the possible issues.
+
+#### Already checked
+
+- the core (`Po.pm`, `Transtractor.pm`) should be safe
+- po4a source code was fully checked for other potential symlink
+  attacks, after discovery of one such issue
+- the only external program run by the core is `diff`, in `Po.pm` (in
+  parts of its code we don't use)
+- `Locale::gettext`: only used to display translated error messages
+- Nicolas François "hopes" `DynaLoader` is safe, and has "no reason to
+  think that `Encode` is not safe"
+- Nicolas François has "no reason to think that `Encode::Guess` is not
+  safe". The po plugin nevertheless avoids using it by defining the
+  input charset (`file_in_charset`) before asking `Transtractor` to
+  read any file. NB: this hack depends on po4a internals to stay
+  the same.
+
+#### To be checked
+
+##### Locale::Po4a modules
+
+The modules we want to use have to be checked, as not all are safe
+(e.g. the LaTeX module's behaviour is changed by commands included in
+the content); they may use regexps generated from the content.
+
+`Chooser.pm` only loads the plugin we tell it too: currently, this
+means the `Text` module only.
+
+`Text` module (I checked the CVS version):
+
+- it does not run any external program
+- only `do_paragraph()` builds regexp's that expand untrusted
+  variables; they seem safe to me, but someone more expert than me
+  will need to check. Joey?
+
+##### Text::WrapI18N
+
+`Text::WrapI18N` can cause DoS (see the
+[Debian bug #470250](http://bugs.debian.org/470250)), but it is
+optional and we do not need the features it provides.
+
+It is loaded if available by `Locale::Po4a::Common`; looking at the
+code, I'm not sure we can prevent this at all, but maybe some symbol
+table manipulation tricks could work; overriding
+`Locale::Po4a::Common::wrapi18n` may be easier. I'm no expert at all
+in this field. Joey? [[--intrigeri]]
+
+> Update: Nicolas François suggests we add an option to po4a to
+> disable it. It would do the trick, but only for people running
+> a brand new po4a (probably too late for Lenny). Anyway, this option
+> would have to take effect in a `BEGIN` / `eval` that I'm not
+> familiar with. I can learn and do it, in case no Perl wizard
+> volunteers to provide the po4a patch. [[--intrigeri]]
+
+##### Term::ReadKey
+
+`Term::ReadKey` is not a hard dependency in our case, *i.e.* po4a
+works nicely without it. But the po4a Debian package recommends
+`libterm-readkey-perl`, so it will probably be installed on most
+systems using the po plugin.
+
+If `$ENV{COLUMNS}` is not set, `Locale::Po4a::Common` uses
+`Term::ReadKey::GetTerminalSize()` to get the terminal size. How safe
+is this?
+
+Part of `Term::ReadKey` is written in C. Depending on the runtime
+platform, this function use ioctl, environment, or C library function
+calls, and may end up running the `resize` command (without
+arguments).
+
+IMHO, using Term::ReadKey has too far reaching implications for us to
+be able to guarantee anything wrt. security. Since it is anyway of no
+use in our case, I suggest we define `ENV{COLUMNS}` before loading
+`Locale::Po4a::Common`, just to be on the safe side. Joey?
+[[--intrigeri]]
+
+> Update: adding an option to disable `Text::WrapI18N`, as Nicolas
+> François suggested, would as a bonus disable `Term::ReadKey`
+> as well. [[--intrigeri]]
+
+### msgmerge
+
+`refreshpofiles()` runs this external program. A po4a developer
+answered he does "not expect any security issues from it".
+
+### Fuzzing input
+
+I was not able to find any public information about gettext or po4a
+having been tested with a fuzzing program, such as `zzuf` or `fusil`.
+Moreover, some gettext parsers seem to be quite
+[easy to crash](http://fusil.hachoir.org/trac/browser/trunk/fuzzers/fusil-gettext),
+so it might be useful to bang msgmerge/po4a's heads against such
+a program in order to easily detect some of the most obvious DoS.
+[[--intrigeri]]
+
+> po4a was not fuzzy-tested, but according to one of its developers,
+> "it would be really appreciated". [[--intrigeri]]
+
+Test conditions:
+
+- a 21M file containing 100 concatenated copies of all the files in my
+  `/usr/share/common-licenses/`; I had no existing PO file or
+  translated versions at hand, which renders these tests
+  quite incomplete.
+- po4a was the Debian 0.34-2 package; the same tests were also run
+  after replacing the `Text` module with the CVS one (the core was not
+  changed in CVS since 0.34-2 was released), without any significant
+  difference in the results.
+- Perl 5.10.0-16
+
+#### po4a-gettextize
+
+`po4a-gettextize` uses more or less the same po4a features as our
+`refreshpot` function.
+
+Without specifying an input charset, zzuf'ed `po4a-gettextize` quickly
+errors out, complaining it was not able to detect the input charset;
+it leaves no incomplete file on disk.
+
+So I had to pretend the input was in UTF-8, as does the po plugin.
+
+Two ways of crashing were revealed by this command-line:
+
+        zzuf -vc -s 0:100 -r 0.1:0.5 \
+          po4a-gettextize -f text -o markdown -M utf-8 -L utf-8 \
+            -m LICENSES >/dev/null
+
+They are:
+
+        Malformed UTF-8 character (UTF-16 surrogate 0xdcc9) in substitution iterator at /usr/share/perl5/Locale/Po4a/Po.pm line 1443.
+        Malformed UTF-8 character (fatal) at /usr/share/perl5/Locale/Po4a/Po.pm line 1443.
+
+and
+
+        Malformed UTF-8 character (UTF-16 surrogate 0xdcec) in substitution (s///) at /usr/share/perl5/Locale/Po4a/Po.pm line 1443.
+        Malformed UTF-8 character (fatal) at /usr/share/perl5/Locale/Po4a/Po.pm line 1443.
+
+Perl seems to exit cleanly, and an incomplete PO file is written on
+disk. I not sure whether if this is a bug in Perl or in `Po.pm`.
+
+#### po4a-translate
+
+`po4a-translate` uses more or less the same po4a features as our
+`filter` function.
+
+Without specifying an input charset, same behaviour as
+`po4a-gettextize`, so let's specify UTF-8 as input charset as of now.
+
+        zzuf -cv \
+          po4a-translate -d -f text -o markdown -M utf-8 -L utf-8 \
+            -k 0 -m LICENSES -p LICENSES.fr.po -l test.fr
+
+... prints tons of occurences of the following error, but a complete
+translated document is written (obviously with some weird chars
+inside):
+
+        Use of uninitialized value in string ne at /usr/share/perl5/Locale/Po4a/TransTractor.pm line 854.
+        Use of uninitialized value in string ne at /usr/share/perl5/Locale/Po4a/TransTractor.pm line 840.
+        Use of uninitialized value in pattern match (m//) at /usr/share/perl5/Locale/Po4a/Po.pm line 1002.
+
+While:
+
+        zzuf -cv -s 0:10 -r 0.001:0.3 \
+          po4a-translate -d -f text -o markdown -M utf-8 -L utf-8 \
+            -k 0 -m LICENSES -p LICENSES.fr.po -l test.fr
+
+... seems to lose the fight, at the `readpo(LICENSES.fr.po)` step,
+against some kind of infinite loop, deadlock, or any similar beast.
+It does not seem to eat memory, though.
+
+Whatever format module is used does not change anything. This is thus
+probably a bug in po4a's core or in a lib it depends on.
+
+The sub `read`, in `TransTractor.pm`, seems to be a good debugging
+starting point.
+
+#### msgmerge
+
+`msgmerge` is run in our `refreshpofiles` function. I did not manage
+to crash it with `zzuf`.
 
 gettext/po4a rough corners
 --------------------------
@@ -227,37 +452,43 @@ gettext/po4a rough corners
   live in different directories): say bla.fr.po has been updated in
   repo2; pulling repo2 from repo1 seems to trigger a PO update, that
   changes bla.fr.po in repo1; then pushing repo1 to repo2 triggers
-  a PO update, that changes bla.fr.po in repo2; etc.; fixed in
-  `629968fc89bced6727981c0a1138072631751fee`?
+  a PO update, that changes bla.fr.po in repo2; etc.; quickly fixed in
+  `629968fc89bced6727981c0a1138072631751fee`, by disabling references
+  in Pot files. Using `Locale::Po4a::write_if_needed` might be
+  a cleaner solution. (warning: this function runs the external
+  `diff` program, have to check security)
 - new translations created in the web interface must get proper
   charset/encoding gettext metadata, else the next automatic PO update
   removes any non-ascii chars; possible solution: put such metadata
   into the Pot file, and let it propagate; should be fixed in
   `773de05a7a1ee68d2bed173367cf5e716884945a`, time will tell.
 
-Misc. improvements
-------------------
-
-### page titles
+Page titles in links
+--------------------
 
-Use nice page titles from meta plugin in links, as inline already
-does. This is actually a duplicate for
-[[bugs/pagetitle_function_does_not_respect_meta_titles]], which might
-be fixed by something like [[todo/using_meta_titles_for_parentlinks]].
+To use the page titles set with the [meta](plugins/meta) plugin when
+rendering links would be very much nicer, than the current
+"filename.LL" format. This is actually a duplicate for
+[[bugs/pagetitle_function_does_not_respect_meta_titles]].
 
-### backlinks
+Page formats
+------------
 
-`po_link_to = negotiated`: if a given translatable `sourcepage.mdwn`
-links to \[[destpage]], `sourcepage.LL.po` also link to \[[destpage]],
-and the latter has the master page *and* all its translations listed
-in the backlinks.
+Markdown is well supported, great, but what about others?
 
-`po_link_to = current`: seems to work nicely
+The [po](plugins/po) uses `Locale::Po4a::Text` for every page format;
+this can be expected to work out of the box with most other wiki-like
+formats supported by ikiwiki. Some of their ad-hoc syntax might be
+parsed in a strange way, but the worst problems I can imagine would be
+wrapping issues; e.g. there is code in po4a dedicated to prevent
+re-wrapping the underlined Markdown headers.
 
-### parentlinks
+While it would be easy to better support formats such as [[html]] or
+LaTeX, by using for each one the dedicated po4a module, this can be
+problematic from a security point of view.
 
-When `usedirs` is disabled and the home page is translatable, the
-parent link to the wiki home page is broken (`/index.html`).
+**TODO**: test the more popular formats and write proper documentation
+about it.
 
 Translation quality assurance
 -----------------------------
@@ -270,3 +501,15 @@ A new `cansave` type of hook would be needed to implement this.
 
 Note: committing to the underlying repository is a way to bypass
 this check.
+
+Broken links
+------------
+
+See [[contrib/po]].
+
+Documentation
+-------------
+
+Maybe write separate documentation depending on the people it targets:
+translators, wiki administrators, hackers. This plugin is maybe
+complex enough to deserve this.