| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <CACT4Y+YB_Ckfptkyu6yiF5Daa8q4MgGyf_1u7RPjoTsdQ3h=qw@mail.gmail.com>
Date: Tue, 11 Feb 2025 21:23:30 +0100
From: Dmitry Vyukov <dvyukov@...gle.com>
To: Namhyung Kim <namhyung@...nel.org>
Cc: irogers@...gle.com, linux-perf-users@...r.kernel.org,
linux-kernel@...r.kernel.org, Arnaldo Carvalho de Melo <acme@...nel.org>
Subject: Re: [PATCH v5 6/8] perf report: Add --latency flag
On Tue, 11 Feb 2025 at 18:42, Namhyung Kim <namhyung@...nel.org> wrote:
>
> On Tue, Feb 11, 2025 at 09:42:16AM +0100, Dmitry Vyukov wrote:
> > On Tue, 11 Feb 2025 at 02:02, Namhyung Kim <namhyung@...nel.org> wrote:
> > >
> > > On Fri, Feb 07, 2025 at 08:23:58AM +0100, Dmitry Vyukov wrote:
> > > > On Fri, 7 Feb 2025 at 04:44, Namhyung Kim <namhyung@...nel.org> wrote:
> [SNIP]
> > > > > > @@ -3547,10 +3549,15 @@ static int __hpp_dimension__add_output(struct perf_hpp_list *list,
> > > > > > return 0;
> > > > > > }
> > > > > >
> > > > > > -int hpp_dimension__add_output(unsigned col)
> > > > > > +int hpp_dimension__add_output(unsigned col, bool implicit)
> > > > > > {
> > > > > > + struct hpp_dimension *hd;
> > > > > > +
> > > > > > BUG_ON(col >= PERF_HPP__MAX_INDEX);
> > > > > > - return __hpp_dimension__add_output(&perf_hpp_list, &hpp_sort_dimensions[col]);
> > > > > > + hd = &hpp_sort_dimensions[col];
> > > > > > + if (implicit && !hd->was_taken)
> > > > > > + return 0;
> > > > >
> > > > > I don't think you need these implicit and was_taken things.
> > > > > Just removing from the sort list when it's unregistered seems to work.
> > > > >
> > > > > ---8<---
> > > > > @@ -685,6 +685,7 @@ void perf_hpp_list__prepend_sort_field(struct perf_hpp_list *list,
> > > > > static void perf_hpp__column_unregister(struct perf_hpp_fmt *format)
> > > > > {
> > > > > list_del_init(&format->list);
> > > > > + list_del_init(&format->sort_list);
> > > > > fmt_free(format);
> > > > > }
> > > > >
> > > > > ---8<---
> > > >
> > > > It merely suppresses the warning, but does not work the same way. See
> > > > this for details:
> > > > https://lore.kernel.org/all/CACT4Y+ZREdDL7a+DMKGFGae1ZjX1C8uNRwCGF0c8iUJtTTq0Lw@mail.gmail.com/
> > >
> > > But I think it's better to pass --latency option rather than adding it
> > > to -s option. If you really want to have specific output fields, then
> > > please use -F latency,sym instead.
> > >
> > > Also I've realized that it should add one sort key in setup_overhead()
> > > to support hierarchy mode properly. Something like this?
> > >
> > > Thanks,
> > > Namhyung
> > >
> > >
> > > ---8<---
> > > diff --git a/tools/perf/util/sort.c b/tools/perf/util/sort.c
> > > index 2b6023de7a53ae2e..329c2e9bbc69a725 100644
> > > --- a/tools/perf/util/sort.c
> > > +++ b/tools/perf/util/sort.c
> > > @@ -3817,22 +3817,15 @@ static char *setup_overhead(char *keys)
> > > return keys;
> > >
> > > if (symbol_conf.prefer_latency) {
> > > - keys = prefix_if_not_in("overhead", keys);
> > > - keys = prefix_if_not_in("latency", keys);
> > > - if (symbol_conf.cumulate_callchain) {
> > > - keys = prefix_if_not_in("overhead_children", keys);
> > > + if (symbol_conf.cumulate_callchain)
> > > keys = prefix_if_not_in("latency_children", keys);
> > > - }
> > > - } else if (!keys || (!strstr(keys, "overhead") &&
> > > - !strstr(keys, "latency"))) {
> > > - if (symbol_conf.enable_latency)
> > > + else
> > > keys = prefix_if_not_in("latency", keys);
> > > - keys = prefix_if_not_in("overhead", keys);
> > > - if (symbol_conf.cumulate_callchain) {
> > > - if (symbol_conf.enable_latency)
> > > - keys = prefix_if_not_in("latency_children", keys);
> > > + } else {
> > > + if (symbol_conf.cumulate_callchain)
> > > keys = prefix_if_not_in("overhead_children", keys);
> > > - }
> > > + else
> > > + keys = prefix_if_not_in("overhead", keys);
> > > }
> > >
> > > return keys;
> >
> >
> > Have I decoded the patch correctly?
> >
> > if (symbol_conf.prefer_latency) {
> > if (symbol_conf.cumulate_callchain)
> > keys = prefix_if_not_in("latency_children", keys);
> > else
> > keys = prefix_if_not_in("latency", keys);
> > } else {
> > if (symbol_conf.cumulate_callchain)
> > keys = prefix_if_not_in("overhead_children", keys);
> > else
> > keys = prefix_if_not_in("overhead", keys);
> > }
> >
>
> Yep, that's correct.
>
>
> > If I decoded the patch correctly, it's not what we want.
> >
> > For the default prefer_latency case we also want to add overhead, that
> > was intentional for the --latency preset. It does not harm, and allows
> > to see/compare differences in latency and overhead.
> > Again, if a user wants something custom, there is no way to second
> > guess all possible intentions. For non-default cases, we just let the
> > user say what exactly they want, and we will follow that.
> >
> > "latency" should be added even if cumulate_callchain.
>
> Please note that it just sets the sort key - which column you want to
> order the result. The output fields for overhead and children will be
> added in perf_hpp__init() if you remove the 'was_taken' logic. So I
> think this change will have the same output with that.
Yes, but perf_hpp__init() does not have the logic that's currently
contained in setup_overhead().
If the user specified a "latency" field, and we don't want to add
"overhead" in that case, then _both_ setup_overhead() and
perf_hpp__init() must not add "overhead".
If we do what you proposed, then perf_hpp__init() will still add
"overhead" and we go back to square 0.
I used was_taken to not duplicate this non-trivial logic in both
functions. As I mentioned in the previous replies, I tried that but it
was messier/more complex. was_taken is a simple way to not duplicate
logic and keep these functions consistent.
> > For the !prefer_latency case, we don't want to mess with
> > overhead/latency fields if the user specified any of them explicitly.
> > Otherwise this convenience part gets in the user's way and does not
> > allow them to do what they want. User says "I want X" and perf says
> > "screw you, I will give you Y instead, and won't allow you to possibly
> > do X".
>
> That's what -F option does. The -s option used to specify how to group
> the histogram entries and it will add 'overhead' (and/or 'latency') if
> it's not even requested. So I think it's ok to add more output column
> when -s option is used.
>
> But unfortunately, using -F and -s together is confusing and change the
> meaning of -s option - it now says how it sort the result.
>
> >
> > And see above: -F does not work with --hierarchy, so this part is unskippable.
>
> Yep, but I mean it fixes --hierarchy and --latency. I'm thinking of a
> way to support -F and --hierarchy in general.
I don't know why it was disabled. There are likely other things to
improve, but please let's not tie that to this change.
Powered by blists - more mailing lists