| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
| |
This reverts commit 19d6c3854e96d89bf4dc2d874df433beac27ee8b.
|
| |
|
|
|
| |
These were discovered while testing the Pod::Checker that is intended to
be used in 5.25.
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
| |
This avoids the following error, reported by Jitka Plesníková
<jplesnik@redhat.com> in a test of RC1:
op.c: In function 'Perl_ck_ftst':
op.c:9754:58: error: format not a string literal and no format arguments [-Werror=format-security]
Perl_warner(aTHX_ packWARN(WARN_SYNTAX), array_passed_to_stat);
In addition, the string in question is now made const.
|
| |
|
|
|
|
|
| |
Give people a heads up that in the next major perl release, subroutine
signatures may no longer populate @_ by default.
See http://nntp.perl.org/group/perl.perl5.porters/235332.
|
| |
|
|
|
| |
Only the text of these should be present in a delta, with a link to the
full explanation. Most of perldelta conformed to this, but not all.
|
| | |
|
| |
|
|
|
| |
I also removed one instance of a 5.23-only bug, which isn't relevant to
a dot zero delta
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
| |
* behaviour -> behavior
* =over -> =over 4
* Add =items for consistency
* Capitalization fixes
|
| | |
|
| |
|
|
| |
Some of this was suggested by Jarkko Hietaniemi
|
| |
|
|
| |
This fixes the 5.24 part of [perl #127899].
|
| |
|
|
| |
(Thanks, Ed Avis!)
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
| |
We say something here that is no longer true; update it.
|
| |
|
|
|
| |
This includes deleting some bug fixes for things introduced in the 5.23
series, and so not relevant to a dot zero perldelta
|
| | |
|
| | |
|
| |\
| |
| |
| |
| | |
This commit is just prep, turning blead into the branch from which
v5.24.0 will be released, after suitable testing.
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | | |
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
This was showing up on Darwin because its setlocale is particularly not
thread safe. But the problem is more generic. Using locales just
isn't a good idea in a threaded application, as locales are
process-wide, not thread-specific. Calling setlocale() changes the
locale of all threads at the same time. Further the return of
setlocale() is a pointer to internal storage. If you call setlocale()
just to learn what it currently is without actually changing the locale,
there is no guarantee that another thread won't interrupt your thread,
switching the locale to something else before you've had a chance to
copy it somewhere else for safekeeping, and the internal storage may
have been freed during that interruption, leading to things like
segfaults.
This is a problem that has been around in the locale handling code for a
long time. I don't know why it hasn't shown up before, or maybe it has
and is not reproducible because it's timing dependent, and so any
problems didn't have tickets written for them, or were rejected as not
reproducible.
But the problem has been made worse in recent releases. Only fairly
recently has perl changed so this problem can occur in programs that
don't use locale explicitly: ones that don't 'use locale' nor call
setlocale(). This ticket is for such a program that gets a
locale-related segfault, without ever touching locales itself.
I have done an audit of the perl source, looking for all such
occurrences, and this patch fixes all of them that I found. The only
other ones, besides "$!", is in converting to/from UTF-8 in cygwin.c.
In all such cases, perl briefly switches the locale, does an operation,
then switches back. The solution here is to add mutexes to make these
areas of code uninterruptible critical sections, so that they can rely
on having the locale be what they expect it to be during the entirety of
the operation, and can't have a setlocale() from another thread free
internal storage. But this is not a general solution. A thread
executing these sections can interrupt some other thread doing a
setlocale() and zap that. However, we have long cautioned against doing
setlocales() in a thread, and that caution was strengthened in a commit
made yesterday, fc82b82ef4784a38877f35f56ee16f14934460ce.
The current commit should make safe all threaded programs that don't use
locales explicitly.
It's too close to the 5.24 release to do the rearchitecting required for
a general solution. That would involve adding more critical sections.
POSIX 2008 introduced new locale handling functions that are
thread-safe, and affect only a single thread, and don't require mutexes.
The ultimate solution would be to use those tools where available, and
to hide from the outer code which set is being used. Thus, perl would
be thread-safe on such platforms, while remaining problematic on older
ones, though fixed so segfaults wouldn't occur. Tony Cook believes we
could emulate the newer behavior on all platforms at a significant
performance penalty. I think this would require a lot of code, and
suspect there would be glitches in it for XS code. But he may have
some ideas about how to do it simply. In any case, this has to wait
until post 5.24.
Three other notes:
It seems to me that the cygwin code could be replaced by equivalent code
that doesn't use locales at all. The comments in the source seem to
even indicate that. I'll look into doing this in 5.25.
Another possible reason that this hasn't shown up in earlier perls is
that the problems may have been entirely affecting I/O operations and
there are already mutexes involving I/O, and so those could be
inadvertently protecting from, or at least minimizing, the problems
found here. I haven't investigated to verify this.
This commit doesn't add a test. I am asking on p5p for assistance in
writing one
|
