On Wed, Mar 21, 2018 at 03:23:51PM +0100, Milian Wolff wrote:
> Here's one way to investigate where perf and dwfl disagree on the file
> mappings:
>
> diff --git a/tools/perf/util/unwind-libdw.c b/tools/perf/util/unwind-libdw.c
> index 7bdd239c795c..739c68b73772 100644
> --- a/tools/perf/util/unwind-libdw.c
> +++ b/tools/perf/util/unwind-libdw.c
> @@ -48,11 +48,17 @@ static int __report_module(struct addr_location *al, u64
> ip,
> mod = 0;
> }
>
> - if (!mod)
> + if (!mod) {
> + Dwarf_Addr s, e;
> mod = dwfl_report_elf(ui->dwfl, dso->short_name,
> (dso->symsrc_filename ?
> dso->symsrc_filename : dso-
> >long_name), -1, al->map->start,
> false);
>
> + dwfl_module_info(mod, NULL, &s, &e, NULL, NULL, NULL, NULL);
> + if (al->map->start != s || al->map->end != e)
> + pr_warning("MEH: %s | mmap: %zx %zx | dwfl: %zx %zx\n",
> dso-
> >short_name, al->map->start, al->map->end, s, e);
> + }
> +
> return mod && dwfl_addrmodule(ui->dwfl, ip) == mod ? 0 : -1;
> }
>
> On one of my perf.data files with many broken backtraces this gives me:
>
> MEH: heaptrack_print | mmap: 56166e9d4000 56166ea39000 | dwfl: 56166e9d4000
> 56166ea38880
> MEH: ld-2.26.so | mmap: 7fd0afc6c000 7fd0afe93000 | dwfl: 7fd0afc6c000
> 7fd0afe920f8
> MEH: libc-2.26.so | mmap: 7fd0ae16a000 7fd0ae521000 | dwfl: 7fd0ae16a000
> 7fd0ae5208f0
> MEH: libstdc++.so.6.0.24 | mmap: 7fd0aea84000 7fd0aee0b000 | dwfl:
> 7fd0aea84000 7fd0aee0a640
> MEH: libzstd.so.1.3.3 | mmap: 7fd0aee0b000 7fd0af087000 | dwfl: 7fd0aee0b000
> 7fd0af086030
>
> Interestingly, here the mmap events observed by perf are actually always
> *larger* than what dwfl sees...
I think that is simply the end address being extended to the next page.
Where the page size seems to be 4K (0x1000). I assume if you look at
the actual LOAD segment sizes of the files (eu-readelf -l) they match
with what dwfl reports, and if it extends to the nearest 4k page it
matches whatp mmap reports.
Cheers,
Mark
