Re: Large memory usage by gdb

[Alex Lindsay <alexlindsay239 at gmail dot com>]

So I wanted to share what I've been doing to make sure that I'm not 
wasting my time (or the list's). Since this is my first deep dive into 
valgrind, I started "simple".  I've been running more or less:

valgrind --xtree-leak=yes gdb --args ./hello

Where `hello` is just a hello world program. My initial leak report was:

==5923== LEAK SUMMARY:
==5923==    definitely lost: 734,882 bytes in 6,231 blocks
==5923==    indirectly lost: 42,581 bytes in 6 blocks
==5923==      possibly lost: 112,422 bytes in 327 blocks
==5923==    still reachable: 8,514,469 bytes in 16,790 blocks
==5923==         suppressed: 0 bytes in 0 blocks

with the initial call-graph (call-graphs viewable at 
https://github.com/lindsayad/markdown-notebooks/blob/master/gdb/images-for-list.md). 
Consequently, I made some changes in the `cp_canonicalize_string` 
function of `cp-support.c` (viewable at 
https://github.com/lindsayad/gdb/pull/1/files). Subsequent running of 
the same valgrind command resulted in the new summary:

==1748== LEAK SUMMARY:
==1748==    definitely lost: 74,226 bytes in 21 blocks
==1748==    indirectly lost: 42,581 bytes in 6 blocks
==1748==      possibly lost: 111,142 bytes in 324 blocks
==1748==    still reachable: 8,515,463 bytes in 16,791 blocks
==1748==         suppressed: 0 bytes in 0 blocks

and a new call-graph. I was fairly pleased that I reduced the number of 
definitely lost bytes by an order of magnitude. So iterating again, I 
made some changes to `elfread.c` and generated the new summary:

==30129== LEAK SUMMARY:
==30129==    definitely lost: 37,538 bytes in 15 blocks
==30129==    indirectly lost: 0 bytes in 0 blocks
==30129==      possibly lost: 111,142 bytes in 324 blocks
==30129==    still reachable: 8,512,473 bytes in 16,788 blocks
==30129==         suppressed: 0 bytes in 0 blocks

and new call-graph. So my question is, is what I'm doing valuable? I 
haven't done any profiling yet to see how these changes affect my real 
use case where I'm debugging an executable with lots of shared 
libraries. Nevertheless, these leaks do seem to be very real. I know 
that GDB developers are way better programmers than I am, so the fact 
that these leaks haven't been found yet makes me wonder whether they 
matter in real use cases or not. I am using a gdb built from the git 
repository (GNU gdb (GDB) 8.0.50.20170803-git).

Thanks for your time,
Alex



On 08/01/2017 02:11 PM, Philippe Waroquiers wrote:
> On Mon, 2017-07-31 at 17:11 -0500, Alex Lindsay wrote:
> > Philippe,
> >
> > Is memcheck a better tool to use here compared to massif?
> In valgrind 3.13, memcheck provides a quite detailed/precise
> way to see delta memory increase/decrease.
> Typically, you will give --xtree-memory=full argument,
> and then e.g. use vgdb to launch a (delta) leak search
> after each run.
> You can then use kcachegrind to visualise the resulting
> memory increase.
>
> Massif is IMO less precise, but automatically produces
> memory status at regular interval.
>
> So, in summary, I would use valgrind 3.13 and memcheck
> (with an additional benefit that if ever your use case
> causes real memory leaks, memcheck will detect them).
>
> Philippe
>
> > Alex
> >
> > On 07/25/2017 03:28 PM, Philippe Waroquiers wrote:
> > > Run gdb under Valgrind, and make some heap profiling dump at regular
> > > interval, (e.g. after each run).
> > >
> > > With valgrind 3.12 or before, you can do a leak report to show
> > > the delta (increase or decrease) compared to the previous leak search,
> > > including the reachable blocks. So, you will be able to see what
> > > increases the memory.
> > >
> > > If you compile the latest Valgrind (3.13), you can e.g. use memcheck
> > > and produce heap profiling reports readable with kcachegrind.
> > >
> > > You will need a gdb compiled with debug or install the debug info
> > > of gdb to have understandable stack traces.
> > >
> > > Philippe
> > >
> > > On Tue, 2017-07-25 at 15:20 -0500, Alex Lindsay wrote:
> > > > My OS is Ubuntu 17.04. Using both gdb 7.12 and 8.0, I experience large
> > > > memory usage when debugging my executable. As I add breakpoints and run
> > > > the executable multiple times in a single session, memory usage grows
> > > > continuously, regularly hitting 10s of GBs. I don't recall experiencing
> > > > this issue with earlier Ubuntu versions (and also likely earlier
> > > > versions of gdb). When I debug the same executable with `lldb`, memory
> > > > usage is pretty much constant at around 2 GB. Does anyone have any
> > > > suggestions?
> > > >
> > > > Alex