Re: [PATCH] improve performance of std::allocator::deallocate

[Jonathan Wakely <jwakely at redhat dot com>]

On 06/03/19 22:27 +0000, Pádraig Brady wrote:
> On 03/06/2019 01:44 AM, Jonathan Wakely wrote:
> > On 06/03/19 09:20 +0000, Pádraig Brady wrote:
> > > On 03/06/2019 12:50 AM, Jonathan Wakely wrote:
> > > > On 06/03/19 02:43 +0000, Pádraig Brady wrote:
> > > > > On 02/26/2019 04:23 PM, Padraig Brady wrote:
> > > > > > > Note jemalloc >= 5.1 is required to fix a bug with 0 sizes.
> > > > > > >
> > > > > > > How serious is the bug? What are the symptoms?
> > > > > > I've updated the commit summary to say it's a crash.
> > > > > > Arguably that's better than mem corruption.
> > > > > >
> > > > > > > It looks like 5.1.0 is less than a year old, so older versions are
> > > > > > > presumably still in wide use.
> > > > > > >
> > > > > > > We could potentially workaround it by making
> > > > > > > new_allocator::allocate(0) call ::operator new(1) when
> > > > > > > __cpp_sized_deallocation is defined, and deallocate(p, 0) call
> > > > > > > ::operator delete(p, 1). Obviously I'd prefer not to do that,
> > > > > > > because
> > > > > > > the default operator new already has an equivalent check, and only
> > > > > > > programs linking to jemalloc require the workaround.
> > > > > > Right the jemalloc fix was released May 2018.
> > > > > > It would be great to avoid the extra workarounds.
> > > > > > Given this would be released about a year after the jemalloc fix was
> > > > > > released,
> > > > > > and that this would only manifest upon program rebuild,
> > > > > > I would be inclined to not add any workarounds.
> > > > > > For reference tcmalloc and glibc malloc were seen to work fine with
> > > > > > this.
> > > > >
> > > > > Actually the jemalloc issue will only be fixed with the release of 5.2
> > > > > (a few weeks away).
> > > > > I've updated the commit message in the attached accordingly.
> > > >
> > > > Hmm, I'm a bit nervous about making a change that will cause crashes
> > > > unless using an unreleased version (I know it will be released by the
> > > > time GCC 9.1 is released, but some people might upgrade GCC without
> > > > upgrading jemalloc).
> > > Yes it's not ideal. It does make it a lot less risky that one has to
> > > rebuild programs to get the new functionality, so existing programs
> > > will be unaffected. Also -fsized-deallocation is only enabled by
> > > default on gcc with -std >= c++14.
> >
> > The default is -std=gnu++14 so it's enabled unless you explicitly
> > choose an older dialect or add -fno-sized-deallocation.
> Good point :)
> > > > On the other hand, zero sized allocations should be rare in practice.
> > > Yes they were rare in testing here
> > >
> > > So programs _rebuilt_ against the following would need to update
> > > to jemalloc 5.2:
> > >
> > >  zero sized allocs, jemalloc<5.2, c++>=14, GCC>=9.1
> > >
> > > Hopefully that's a small enough set.
> >
> > Which versions of jemalloc replace operator delete(void*, size_t) ?
> >
> > Was that something new in 5.0 or did older versions already provide a
> > replacement for the sized operator delete?
> >
> > If it was introduced in 5.0 then there won't be a problem for 3.x and
> > 4.x because they'll use the default definition from libstdc++ which
> > just calls ::operator delete(p).
> Again very good point. The replacements were only added in 5.0:
> https://github.com/jemalloc/jemalloc/commit/2319152d

OK, that makes me feel better about it. It's presumably much easier to
upgrade to 5.2 from 5.0 or 5.1 than it would be from 4.x.

> > How complicated is the fix to prevent the crashes? Would it be
> > feasible for distros to backport that fix? I see that RHEL8 has
> > jemalloc 5.0.1 for example, but if the fix could be backported to that
> > release then it's less of a problem.
> The patch set is simple enough:
> https://github.com/jemalloc/jemalloc/pull/1341/commits

Thanks. That does seem reasonable for distros and other packagers to
backport, if they want to support 5.0 or 5.1 for their users.

I'm leaning towards accepting the patch for gcc-9 (and if not, we
should do it early in the gcc-10 cycle).