Message-ID: <CAP-5=fUg5hoZF8TLK+ivAVzVLKW0FsGaDWS+TN55m72eHLNUHw@mail.gmail.com>
Date: Sun, 14 Jul 2024 11:13:51 -0700
From: Ian Rogers <irogers@...gle.com>
To: Namhyung Kim <namhyung@...nel.org>
Cc: Peter Zijlstra <peterz@...radead.org>, Ingo Molnar <mingo@...hat.com>, 
	Arnaldo Carvalho de Melo <acme@...nel.org>, Mark Rutland <mark.rutland@....com>, 
	Alexander Shishkin <alexander.shishkin@...ux.intel.com>, Jiri Olsa <jolsa@...nel.org>, 
	Adrian Hunter <adrian.hunter@...el.com>, Kan Liang <kan.liang@...ux.intel.com>, 
	James Clark <james.clark@....com>, linux-kernel@...r.kernel.org, 
	linux-perf-users@...r.kernel.org
Subject: Re: [PATCH v2 1/2] tools subcmd: Add non-waitpid check_if_command_finished()

On Sat, Jul 13, 2024 at 8:00 AM Namhyung Kim <namhyung@...nel.org> wrote:
>
> On Fri, Jul 12, 2024 at 02:19:58PM -0700, Ian Rogers wrote:
> > On Fri, Jul 12, 2024 at 1:33 PM Namhyung Kim <namhyung@...nel.org> wrote:
> > >
> > > Hi Ian,
> > >
> > > On Tue, Jul 02, 2024 at 09:24:50PM -0700, Ian Rogers wrote:
> > > > On Tue, Jul 2, 2024 at 8:24 PM Namhyung Kim <namhyung@...nel.org> wrote:
> > > > >
> > > > > Hi Ian,
> > > > >
> > > > > On Sun, Jun 30, 2024 at 09:42:35PM -0700, Ian Rogers wrote:
> > > > > > Using waitpid can cause stdout/stderr of the child process to be
> > > > > > lost. Use Linux's /prod/<pid>/status file to determine if the process
> > > > > > has reached the zombie state. Use the 'status' file rather than 'stat'
> > > > > > to avoid issues around skipping the process name.
> > > > > >
> > > > > > Signed-off-by: Ian Rogers <irogers@...gle.com>
> > > > > > ---
> > > > > >  tools/lib/subcmd/run-command.c | 33 +++++++++++++++++++++++++++++++++
> > > > > >  1 file changed, 33 insertions(+)
> > > > > >
> > > > > > diff --git a/tools/lib/subcmd/run-command.c b/tools/lib/subcmd/run-command.c
> > > > > > index 4e3a557a2f37..ec06683e77a0 100644
> > > > > > --- a/tools/lib/subcmd/run-command.c
> > > > > > +++ b/tools/lib/subcmd/run-command.c
> > > > > > @@ -2,6 +2,7 @@
> > > > > >  #include <unistd.h>
> > > > > >  #include <sys/types.h>
> > > > > >  #include <sys/stat.h>
> > > > > > +#include <ctype.h>
> > > > > >  #include <fcntl.h>
> > > > > >  #include <string.h>
> > > > > >  #include <linux/string.h>
> > > > > > @@ -217,8 +218,40 @@ static int wait_or_whine(struct child_process *cmd, bool block)
> > > > > >
> > > > > >  int check_if_command_finished(struct child_process *cmd)
> > > > > >  {
> > > > > > +#ifdef __linux__
> > > > >
> > > > > Is this really necessary?  I don't think we plan to support other OS..
> > > >
> > > > I don't think it'd be unreasonable to say run "perf report" on
> > > > Windows, or using wasm inside a web browser. Part of the reason for
> > > > doing things this way was to keep the WNOHANG logic although this
> > > > change no longer uses it for __linux__.
> > >
> > > I'm not sure we are ready to run it on other platforms.  So I think
> > > it's better simply remove it for now.
> >
> > So in the office hours there was some discussion with a potential new
> > contributor whose development platform is OS/X. It's fairly obvious
> > this code can't work on anything but Linux and using #error feels
> > annoying. The waitpid code is tested and has a known issue, but I
> > think it is better than just breaking anyone not on Linux.
>
> I feel like it's a potential issue and should be handled by the
> potentiall contributor.  Until that happens, we can assume Linux
> and keep the code minimal.

I'm not clear what the issue is. Arnaldo took the WNOHANG waitpid
contribution but I asked him to drop it due to losing stdout/stderr in
parallel mode due to the waitpid closing these file descriptors early
- we're not running in parallel mode by default any more. Reading
procfs to determine zombie state is clearly a Linux only thing, hence
the ifdef. I'm keen to keep to any extent possible the perf tool
running on non-Linux platforms, for example, gathering a data file on
a server or embedded system then wanting to do perf report on a
different machine which may not be running Linux. libsubcmd is used
throughout the tool and we have many subprocesses for pagers, objdump,
addr2line, etc. I don't agree with making such a core library Linux
only and there's no obligation for the maintainers to take these
patches. I disagree with a minimal patch.

Thanks,
Ian

> Thanks,
> Namhyung
>
> > >
> > > > > > +     char filename[FILENAME_MAX + 12];
> > > > > > +     char status_line[256];
> > > > > > +     FILE *status_file;
> > > > > > +
> > > > > > +     /*
> > > > > > +      * Check by reading /proc/<pid>/status as calling waitpid causes
> > > > > > +      * stdout/stderr to be closed and data lost.
> > > > > > +      */
> > > > > > +     sprintf(filename, "/proc/%d/status", cmd->pid);
> > > > > > +     status_file = fopen(filename, "r");
> > > > > > +     if (status_file == NULL) {
> > > > > > +             /* Open failed assume finish_command was called. */
> > > > > > +             return true;
> > > > > > +     }
> > > > > > +     while (fgets(status_line, sizeof(status_line), status_file) != NULL) {
> > > > > > +             char *p;
> > > > > > +
> > > > > > +             if (strncmp(status_line, "State:", 6))
> > > > > > +                     continue;
> > > > > > +
> > > > > > +             fclose(status_file);
> > > > > > +             p = status_line + 6;
> > > > > > +             while (isspace(*p))
> > > > > > +                     p++;
> > > > > > +             return *p == 'Z';
> > > > > > +     }
> > > > > > +     /* Read failed assume finish_command was called. */
> > > > > > +     fclose(status_file);
> > > > > > +     return true;
> > > > > > +#else
> > > > > >       wait_or_whine(cmd, /*block=*/false);
> > > > > >       return cmd->finished;
> > > > > > +#endif
> > > > > >  }
> > > > > >
> > > > > >  int finish_command(struct child_process *cmd)
> > > > > > --
> > > > > > 2.45.2.803.g4e1b14247a-goog
> > > > > >

Powered by blists - more mailing lists