Re: [PATCH] [v6] malloc: Consistently apply trim_threshold to all heaps [BZ #17195]

[Mel Gorman <mgorman at suse dot de>]

On Tue, Feb 24, 2015 at 12:31:18PM +0530, Siddhesh Poyarekar wrote:
> On Fri, Feb 20, 2015 at 01:27:47PM +0000, Mel Gorman wrote:
> > 2015-02-10  Mel Gorman  <mgorman@suse.de>
> > 
> >   [BZ #17195]
> >   * malloc/arena.c (free): Apply trim threshold to per-thread heaps
> >     as well as the main arena.
> > ---
> >  malloc/arena.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/malloc/arena.c b/malloc/arena.c
> > index 886defb074a2..fec339819309 100644
> > --- a/malloc/arena.c
> > +++ b/malloc/arena.c
> > @@ -696,7 +696,7 @@ heap_trim (heap_info *heap, size_t pad)
> >      }
> >    top_size = chunksize (top_chunk);
> >    extra = (top_size - pad - MINSIZE - 1) & ~(pagesz - 1);
> > -  if (extra < (long) pagesz)
> > +  if ((unsigned long) extra < mp_.trim_threshold)
> >      return 0;
> >  
> >    /* Try to shrink. */
> 
> Looks like this exposes some kind of a race.  I ran the testsuite with
> this patch and intl/tst-gettext4 and intl/tst-gettext5 fail with a
> segfault.  From a quick peek it looks like a malloc trying to tinker
> with the part of an arena that has not yet been allocated, i.e. does
> not have any permissions.  Can you please look into it?  I can't push
> this till that problem is solved.
> 

Ok, I'll start taking a look. I have yet to properly investigate but I
do think this is an existing problem that the patch makes worse because
I find this

# This is a test run using the glibc installed on the system
mel@stampy:~/git-public/glibc/obj-baseline > ./intl/tst-gettext4
beauty
thread 1 call 1 returned: beauty
beauty
thread 2 call 1 returned: beauty
beauty
thread 1 call 2 returned: beauty
beauty
thread 2 call 2 returned: beauty

# This is a test run using glibc 2.21
mel@stampy:~/git-public/glibc/obj-baseline > ./testrun.sh ./intl/tst-gettext4
Schï¿œnheit
beautￜ
Schï¿œnheit
beautￜ

# This is a test run using glibc 2.21 + the patch
mel@stampy:~/git-public/glibc/obj-madvise-reduce-v6r1 > ./testrun.sh ./intl/tst-gettext4
Segmentation fault (core dumped)

glibc 2.21 does not segfault but the output from glibc 2.21 looks pretty
screwed up relative to glibc 2.19 that is installed on the system.

-- 
Mel Gorman
SUSE Labs