This had been initially dealt with in issue #1699's commits, but one of them
was in the end reverted, which caused this problem (this change should have
been done there).
New FreeBSD versions have introduced RTLD_DEEPBIND (see dlopen(3)), which
triggers a provoked build error to be on the "safe" side.
Indeed, improper use of RTLD_DEEPBIND could cause malloc interposing not to
work on specific libraries. This would happen for libraries that provide their
own malloc/free and exchange pointers with other libraries or the main program,
one allocating an area and another one deallocating it.
It seems it happened for glibc, which is not a concern on FreeBSD, and for the
Flash Player (not a concern either, there is no native Flash Player). Moreover,
there is only a single reference to RTLD_DEEPBIND in the whole platform
currently, namely in the WebRTC code, but the corresponding code chunk is
compiled-in only on Linux (and this library doesn't seem to redefine malloc
functions anyway). So I don't see how a problem could happen. Additionally,
Pale Moon, which is my focus, doesn't use WebRTC.
Corresponding Mozilla bug: https://bugzilla.mozilla.org/show_bug.cgi?id=493541
Coded a simple memory allocator meant to be used during jemalloc bootstrap
(malloc_init_hard()). Although protected by "#ifdef __FreeBSD__", it is not
FreeBSD-specific: Any POSIX platform could use it.
Hook it so that it is used in place of jemalloc's own routines while
malloc_init_hard() is executing or memory for mutexes is being
allocated. Currently, 'malloc', 'calloc', '*memalign' and 'free' are diverted
during init or lock initializations.
Details are quite complex, see the big comment block starting with "There are
several problematic interactions between FreeBSD's libthr and this jemalloc."
for more explanations.
Also replaced ad-hoc BSD code to determine the number of CPUs with 'sysconf',
which is POSIX-compliant (and supported on modern BSDs).
The barriers are here to make sure that setting 'malloc_initialized' at end of
init must be noticed later by any thread running on a different core. They are
in theory necessary in the absence of an explicit pthread lock.
What could happen is that the thread doing the initialization later spawns
other threads that could not have the updated 'malloc_initialized' value,
leading to a second initialization. This is dependent on whether OSes force a
full memory barrier before the new thread is run, which I don't know, and don't
want to bother.
This was done for FreeBSD only, for the sake of robustness. In theory, this
would be needed on Windows too.
Since these are just interpreted comments, there's 0 impact on actual code.
This removes all lines that match /* vim: set(.*)tw=80: */ with S&R -- there are
a few others scattered around which will be removed manually in a second part.
The JEMALLOC_USES_MAP_ALIGN code has turned out to be worse than useless, and in fact it breaks 64-bit SunOS. It's very old and turned out not to be needed anymore because the way the memory allocator works has changed since it was implemented. It prevented a fix I tried for 48-bit addressing from working properly. However, without either this code or the 48-bit addressing fix, the 64-bit version won't even start, which is why I thought the code was still needed.
https://bugzilla.mozilla.org/show_bug.cgi?id=457189https://hg.mozilla.org/mozilla-central/rev/a26c500b98ab
The 48-bit addressing fix is based on code found here:
https://github.com/OpenIndiana/oi-userland/blob/oi/hipster/components/web/firefox/patches/patch-js_src_gc_Memory.cpp.patch
I already applied these changes to js/src/gc/Memory.cpp, but as I was looking through jemalloc.c I saw that there were very similar ifdefs for Linux on SPARC as the ones I'd had to enaable in Memory.cpp, but for whatever reason the patches I found didn't touch them. So I tried doing for jemalloc.c what was already done for Memory.cpp, and it worked (but only after I removed the map align code).
Correct various pre-processor defines for sparc64 and in mozjemalloc use the JS arm64 allocator on Linux/sparc64.
This corrects build problems opn Linux sparc64 and is in line with bugzilla bug #1275204.
"USING VLA'S IS ACTIVELY STUPID! It generates much more code, and much
slower code (and more fragile code), than just using a fixed key size
would have done." -- Linus Torvalds
We never use jemalloc without MOZ_MEMORY surrounding code (mostly a
FreeBSD concession which doesn't work anyway). This allows us to get rid
of the "poor man's mutex" (CPU-based spinlock) in mozjemalloc.
This basic validation is hard-set to always-on, so no point in keeping
it conditional since we basically always want this rudimentary pointer
validation to be done.
https://bugzilla.mozilla.org/show_bug.cgi?id=1375467
I would ifdef this, but it's been in Firefox since version of 56, and Petr Sumbara's explanation as to why it's wrong in the first place is so detailed that it's pretty obvious the code wasn't technically doing things properly to begin with.
Basically, they tried to redefine a system function after including the header file that declares it, and it caused problems on Solaris because libc functions are imported into the C++ std namespace in a different way that also complies with standards. So the existing implementation is technically bad code on all platforms, the Solaris implementation just uncovered the lack of standards compliance in the Mozilla code.
Compared with what Pale Moon had for Solaris originally, this is mostly the same zero point I started patching from, but I've made the following changes here after reviewing all this initial code I never looked at closely before.
1. In package-manifest.in for both Basilisk and Pale Moon, I've made the SPARC code for libfreebl not interefere with the x86 code, use the proper build flags, and also updated it to allow a SPARC64 build which is more likely to be used than the 32-bit SPARC code we had there.
2. See Mozilla bug #832272 and the old rules.mk patch from around Firefox 30 in oracle/solaris-userland. I believe they screwed up NSINSTALL on Solaris when they were trying to streamline the NSS buildsystem, because they started having unexplained issues with it around that time after Firefox 22 that they never properly resolved until Mozilla began building NSS with gyp files. I'm actually not even sure how relevant the thing they broke actually is to Solaris at this point, bug 665509 is so old it predates Firefox itself and goes back to the Mozilla suite days. I believe $(INSTALL) -t was wrong, and they meant $(NSINSTALL) -t because that makes more sense and is closer to what was there originally. It's what they have for WINNT, and it's possible a fix more like that could serve for Solaris as well. Alternatively, we could get rid of all these half-broken Makefiles and start building NSS with gyp files like Mozilla did.
3. I've completely cut out support for the Sun compiler and taken into account the reality that everyone builds Firefox (and therefore its forks) with GCC now on Solaris. This alone helped clean up a lot of the uglier parts of the code.
4. I've updated all remaining SOLARIS build flags to the newer XP_SOLARIS, because the SOLARIS flag is no longer set when building Solaris.
5. I've confirmed the workaround in gtxFontconfigFonts.cpp is no longer necessary. The Solaris people got impatient about implementing a half-baked patch for a fontconfig feature that wasn't ready yet back in 2009, and somehow convinced Mozilla to patch their software to work around it when really they should have just fixed or removed their broken fontconfig patch. The feature they wanted has since been implemented properly, and no version of Solaris still uses the broken patch that required this fix. If anyone had ever properly audited this code, it would have been removed a long time ago.
We're not actually using it, and it messes up with the zone allocator in
mozjemalloc after fork(). See the lengthy analysis in
https://bugzilla.mozilla.org/show_bug.cgi?id=1424709#c34 and following.
--HG--
extra : rebase_source : b191048290a907cc7668ad7ab6369ef8661f31dc
extra : intermediate-source : 45c5d467a46077dcc3ccd59feafd0c2784432fef
extra : source : bf1efa161edb20a83fe8db2f96c51f4e66880153
We are using |throw(std::bad_alloc)|, but dynamic exception specifications have been deprecated in C++11. The C++11 equivalent is |noexcept(false)|. This causes build warning spam when using newer compilers such as GCC 7.x.
Moonchild
Olivier Certner
Brian Smith
athenian200
JMadgwick
Matt A. Tobin
wolfbeast
trav90
wolfbeast
Mike Hommey
trav90
Matt A. Tobin