Re: Bug#805391: src:rrdtool: FTBFS on mips: `FAIL: tune2`

On 2015-12-19 15:03, Niko Tyni wrote:
> On Tue, Nov 24, 2015 at 04:24:54PM +0000, Jean-Michel Vourgère wrote:
> > Control: tags -1 + help
> > 
> > Antonio Terceiro wrote:
> > > "make  check-TESTS" fails
> > 
> > I tested previous version on mips, and it now FTBFS too. It used to work
> > ok. [1] (thanks Tobias, upstream, for suggesting that.)
> > 
> > So I suppose it is related to another package.
> > 
> > I tried compiling with good old gcc 4.9 and I get the same error in 1.5.4.
> 
> It's a latent bug in rrdtool that triggered with libc6 2.19-20 on mips.
> 
> The test suite failure in tests/tune2 boils down to
> 
>  rrdtool tune --beta 0.4 tests/tune2-testorg.rrd
> 
> zeroing out a big part of the file in addition to changing a single byte
> like it's supposed to. It's easy to see by calling
> 
>  rrdtool dump tests/tune2-testorg.rrd >/dev/null
> 
> which starts to fail with stderr
> 
>  ERROR: unknown consolidation function ''
> 
> on a corrupted file.
> 
> This started to happen with libc6 2.19-20, as tested by compiling rrdtool
> 1.5.5 on jessie and running with LD_LIBRARY_PATH pointing to different
> libc.so.6 versions.
> 
> The problem is that rrd_write() ends up calling memcpy(3) on overlapping
> memory areas, which is explicitly prohibited in its documentation. With
> libc6 2.19-20 on mips, this started zeroing out part of the areas under
> some conditions. The attached testcase triggers it at least on minkus.d.o,
> tested with something like
> 
>   #!/bin/bash
>   dd if=/dev/urandom of=testfile.orig bs=1024 count=1
>   gcc t.c
>   cp testfile.orig testfile
>   ./a.out testfile
>   diff -u <(od -x testfile.orig) <(od -x testfile)
>   
> which normally doesn't show any differences.  I'm cc'ing the libc6
> maintainers, not sure how intentional this change of behaviour is.

The change is likely due to the switch to the MIPS32R2 ISA on mips.
A different code is used, so with a different undefined behaviour.

> Backtrace for the rrd_write() call that triggers it:
> 
>   #0  rrd_write (rrd_file=0x41d038, buf=0x77fef168, count=560) at rrd_open.c:798
>   #1  0x77f65888 in rrd_tune (argc=4, argv=0x7fff6648) at rrd_tune.c:395
>   #2  0x00402f44 in HandleInputLine (argc=5, argv=0x7fff6644, out=0x77f56900 <_IO_2_1_stderr_>) at rrd_tool.c:800
>   #3  0x00400f10 in main (argc=5, argv=0x7fff6644) at rrd_tool.c:545
>   
>   (gdb) print ((rrd_simple_file_t *) rrd_file->pvt)->file_start + rrd_file->pos
>   $6 = 0x77fef168 "HWPREDICT"
>   
>   (gdb) print (char *)buf
>   $8 = 0x77fef168 "HWPREDICT"
> 
> A workaround is to configure with --disable-mmap, except that seems
> to break tests/tune1 on all architectures (at least amd64 and mips).
> I haven't looked into that.

Can't you replace memcpy calls with overlapping areas with a memmove
instead? Even if the testsuite looks fine you will definitely have some
corruptions with memcpy and overlapping area on other architectures too.

Aurelien

-- 
Aurelien Jarno                          GPG: 4096R/1DDD8C9B
[email protected]                 http://www.aurel32.net