Date:   Sun, 25 Jun 2017 23:43:02 +0900
From:   Namhyung Kim <namhyung@...nel.org>
To:     Arnaldo Carvalho de Melo <acme@...nel.org>
Cc:     Ingo Molnar <mingo@...nel.org>,
        Peter Zijlstra <a.p.zijlstra@...llo.nl>,
        Jiri Olsa <jolsa@...nel.org>,
        LKML <linux-kernel@...r.kernel.org>, kernel-team@....com,
        Masami Hiramatsu <mhiramat@...nel.org>,
        Andi Kleen <andi@...stfloor.org>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Wang Nan <wangnan0@...wei.com>
Subject: Re: [PATCH/RFC 9/9] perf record: Add --module-dir option

On Fri, Jun 23, 2017 at 11:45:24AM -0300, Arnaldo Carvalho de Melo wrote:
> Em Fri, Jun 23, 2017 at 02:48:27PM +0900, Namhyung Kim escreveu:
> > Currently perf only searches module binaries on the canonical
> > directory (/lib/modules/`uname -r`).  But sometimes user needs to load
> > local modules.  These cannot be copied to the build-id cache since long
> > name (i.e. real path) of DSOs was not set.
>  
> > This patch fixes the problem by adding a new --module-dir option so that
> > perf can record modules in the directory.
>  
> > +++ b/tools/perf/Documentation/perf-record.txt
> > @@ -480,6 +480,9 @@ Implies --tail-synthesize.
> > +--module-dir=PATH::
> > +Directory name where extra modules are located.
> 
> > +++ b/tools/perf/builtin-record.c
> > @@ -1671,6 +1671,8 @@ static struct option __record_options[] = {
> >  		    "Parse options then exit"),
> >  	OPT_BOOLEAN(0, "use-kcore", &symbol_conf.use_kcore,
> >  		    "Use /proc/kcore for object code"),
> > +	OPT_STRING(0, "module-dir", &symbol_conf.extra_module_path, "path",
> > +		   "directory name where extra modules are located"),
> >  	OPT_END()
> >  };
> >  
> > +++ b/tools/perf/util/machine.c
> > @@ -1117,7 +1118,19 @@ static int machine__set_modules_path(struct machine *machine)
> >  		 machine->root_dir, version);
> >  	free(version);
> >  
> > -	return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
> > +	ret = map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
> > +	if (ret < 0)
> > +		return ret;
> > +
> > +	if (symbol_conf.extra_module_path) {
> > +		snprintf(modules_path, sizeof(modules_path), "%s/%s",
> > +			 machine->root_dir, symbol_conf.extra_module_path);
> > +
> > +		ret = map_groups__set_modules_path_dir(&machine->kmaps,
> > +						       modules_path, 0);
> 
> What if we have samples in a module that is in the canonical dir _and_
> samples in a module in this extra_module_path? Shouldn't we have
> something like vmlinux_path, but for modules? Where you can add entries
> in that path?
> 
> I'm ok with adding entries to where module files are looked up, with
> semantics similar to $PATH in a shell, i.e. entries added via
> --module-path (rename of your --module-dir) will take precedence over
> the canonical dir (/lib/modules/`uname -r`).

I think we can check the build-id in sysfs and binaries in both
directories.  If the sysfs is not avaiable (i.e. the module is
unloaded in the meantime), the precedence can be used IMHO.

Thanks,
Namhyung


> 
> But for the future I think we should try to get a PERF_RECORD_MMAP that
> will allow reloading modules during a 'perf record' session when a
> module gets reloaded, I thought about using existing tracepoints, but...
> 
> [root@...et ~]# cat /sys/kernel/debug/tracing/events/module/module_load/format 
> name: module_load
> ID: 370
> format:
> 	field:unsigned short common_type;	offset:0;	size:2;	signed:0;
> 	field:unsigned char common_flags;	offset:2;	size:1;	signed:0;
> 	field:unsigned char common_preempt_count;	offset:3;	size:1;	signed:0;
> 	field:int common_pid;	offset:4;	size:4;	signed:1;
> 
> 	field:unsigned int taints;	offset:8;	size:4;	signed:0;
> 	field:__data_loc char[] name;	offset:12;	size:4;	signed:1;
> 
> print fmt: "%s %s", __get_str(name), __print_flags(REC->taints, "", { (1UL << 0), "P" }, { (1UL << 12), "O" }, { (1UL << 1), "F" }, { (1UL << 10), "C" }, { (1UL << 13), "E" })
> [root@...et ~]# perf trace --no-syscalls --event module:module_load modprobe ppp
> modprobe: FATAL: Module ppp not found in directory /lib/modules/4.12.0-rc4+
> [root@...et ~]# perf trace --no-syscalls --event module:module_load modprobe e1000
>      0.000 module:module_load:e1000 )
> [root@...et ~]#
> 
> It just gets us the module name, not the file from what it is loaded, bummer :-\
> 
> Something like:
> 
> diff --git a/kernel/module.c b/kernel/module.c
> index 4a3665f8f837..50e9718a141d 100644
> --- a/kernel/module.c
> +++ b/kernel/module.c
> @@ -3739,6 +3739,7 @@ static int load_module(struct load_info *info, const char __user *uargs,
>  
>  	/* Done! */
>  	trace_module_load(mod);
> +	perf_event_mmap_mod(mod);
>  
>  	return do_init_module(mod);
>  
>  ----------
> 
> struct module *mod has:
> 
>   mod->name:		"e1000"
>   mod->debug->filename: /lib/modules/4.11.0-rc8+/kernel/drivers/net/ethernet/intel/e1000/e1000.ko
> 
> and at that point we also have load_info, htat has ELF headers where we could extract the build-id
> and insert it in a field in a new PERF_RECORD_MMAP3 record :-)
> 
>   perf_event_mmap_mod() would be modelled after perf_event_mmap(vma), but getting what
>   it needs not from the vma present in a sys_mmap() but from struct module and load_info.
> 
> - Arnaldo
> 
> > +	}
> > +
> > +	return ret;
> >  }
> >  int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
> >  				const char *name __maybe_unused)
> > diff --git a/tools/perf/util/symbol.h b/tools/perf/util/symbol.h
> > index 88361eeae813..59370ceb87c4 100644
> > --- a/tools/perf/util/symbol.h
> > +++ b/tools/perf/util/symbol.h
> > @@ -124,7 +124,8 @@ struct symbol_conf {
> >  	const char	*vmlinux_name,
> >  			*kallsyms_name,
> >  			*source_prefix,
> > -			*field_sep;
> > +			*field_sep,
> > +			*extra_module_path;
> >  	const char	*default_guest_vmlinux_name,
> >  			*default_guest_kallsyms,
> >  			*default_guest_modules;
> > -- 
> > 2.13.1

Powered by blists - more mailing lists