trace-cmd: fix extract output option

Commit Message

Vincent Donnefort Sept. 29, 2020, 5:49 p.m. UTC

From: Vincent Donnefort <[e-mail]>

During the introduction of instance's output_file copy:

  3a206ca ("trace-cmd: Have instances include a copy of its output file")

The extract path has been omitted, leading to a broken output option:

  $ trace-cmd extract -o /foo/bar.dat # Will fallback to ./trace.dat

Signed-off-by: Vincent Donnefort <[e-mail]>

Comments

Steven Rostedt Sept. 29, 2020, 8:31 p.m. UTC | #1

On Tue, 29 Sep 2020 18:49:26 +0100
[e-mail] wrote:

> From: Vincent Donnefort <[e-mail]>
> 
> During the introduction of instance's output_file copy:
> 
>   3a206ca ("trace-cmd: Have instances include a copy of its output file")
> 
> The extract path has been omitted, leading to a broken output option:
> 
>   $ trace-cmd extract -o /foo/bar.dat # Will fallback to ./trace.dat

When I tried this it worked fine to me. But then I walked through the logic
via gdb and found that the intermediate step (the one that writes the
individual buffers directly), which can be an issue if you happen to
execute this in a directory that you can not write to, or doesn't have
enough space to hold all the data. Thus your patch is correct, but the
change log is not.

Do you really see "trace.dat" at the end of that command? Because I
see /foo/bar.dat.

But if I try to run the extract in /sys/kernel/tracing, it will fail
because it will try to write "(null).cpuX" where X is the CPU number.

But the creation of the actual file uses ctx->output, which is what we want.

Anyway, I'll update the change log to this:

    During the introduction of instance's output_file copy:

      3a206ca ("trace-cmd: Have instances include a copy of its output file")

    The extract path has been omitted, causing the temp files created to be
    written in the same directory using the null "output_file" of the
    instance, to create "(null).cpuX" files. If this is executed in a
    directory that is not writable (like /sys/kernel/tracing) or does not
    have enough space to hold the temp files, then it will fail to write.

Fair?

-- Steve


> 
> Signed-off-by: Vincent Donnefort <[e-mail]>
> 
> diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
> index bd00457..72a5c8c 100644
> --- a/tracecmd/trace-record.c
> +++ b/tracecmd/trace-record.c
> @@ -6622,6 +6622,9 @@ void trace_extract(int argc, char **argv)
>  
>  	/* Save the state of tracing_on before starting */
>  	for_all_instances(instance) {
> +		instance->output_file = strdup(ctx.output);
> +		if (!instance->output_file)
> +			die("Failed to allocate output file name for instance");
>  
>  		if (!ctx.manual && instance->flags & BUFFER_FL_PROFILE)
>  			enable_profile(ctx.instance);

Vincent Donnefort Sept. 30, 2020, 8:31 a.m. UTC | #2

On Tue, Sep 29, 2020 at 04:31:52PM -0400, Steven Rostedt wrote:
> On Tue, 29 Sep 2020 18:49:26 +0100
> [e-mail] wrote:
> 
> > From: Vincent Donnefort <[e-mail]>
> > 
> > During the introduction of instance's output_file copy:
> > 
> >   3a206ca ("trace-cmd: Have instances include a copy of its output file")
> > 
> > The extract path has been omitted, leading to a broken output option:
> > 
> >   $ trace-cmd extract -o /foo/bar.dat # Will fallback to ./trace.dat
> 
> When I tried this it worked fine to me. But then I walked through the logic
> via gdb and found that the intermediate step (the one that writes the
> individual buffers directly), which can be an issue if you happen to
> execute this in a directory that you can not write to, or doesn't have
> enough space to hold all the data. Thus your patch is correct, but the
> change log is not.
> 
> Do you really see "trace.dat" at the end of that command? Because I
> see /foo/bar.dat.
> 
> But if I try to run the extract in /sys/kernel/tracing, it will fail
> because it will try to write "(null).cpuX" where X is the CPU number.
> 
> But the creation of the actual file uses ctx->output, which is what we want.
> 
> Anyway, I'll update the change log to this:
> 
>     During the introduction of instance's output_file copy:
> 
>       3a206ca ("trace-cmd: Have instances include a copy of its output file")
> 
>     The extract path has been omitted, causing the temp files created to be
>     written in the same directory using the null "output_file" of the
>     instance, to create "(null).cpuX" files. If this is executed in a
>     directory that is not writable (like /sys/kernel/tracing) or does not
>     have enough space to hold the temp files, then it will fail to write.
> 
> Fair?
> 
> -- Steve

Apologies, the description (... and my testing) was indeed incorrect. The
output is working fine, the problem being only the temporary files.

No problem with the updated commit description. We probably need to update
the short description as well:

  "trace-cmd: Fix extract temporary files path" ?

Patch

diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
index bd00457..72a5c8c 100644
--- a/tracecmd/trace-record.c
+++ b/tracecmd/trace-record.c
@@ -6622,6 +6622,9 @@  void trace_extract(int argc, char **argv)
 
 	/* Save the state of tracing_on before starting */
 	for_all_instances(instance) {
+		instance->output_file = strdup(ctx.output);
+		if (!instance->output_file)
+			die("Failed to allocate output file name for instance");
 
 		if (!ctx.manual && instance->flags & BUFFER_FL_PROFILE)
 			enable_profile(ctx.instance);