Re: [PATCH v4 10/11] perf env: Set flag for kernel is 64-bit mode

Em Tue, Jul 13, 2021 at 05:31:03PM +0000, Hunter, Adrian escreveu:

quoted

On Mon, Jul 12, 2021 at 03:14:35PM -0300, Arnaldo Carvalho de Melo wrote:

quoted

Em Sun, Jul 11, 2021 at 06:41:04PM +0800, Leo Yan escreveu:

quoted

+++ b/tools/perf/util/env.c

@@ -11,6 +11,7 @@
 #include <stdlib.h>
 #include <string.h>

quoted

quoted

quoted

+int kernel_is_64_bit;
 struct perf_env perf_env;

quoted

quoted

Why can't this be in 'struct perf_env'?

quoted

Good question.  I considered to add it in struct perf_env but finally I used this
way; the reason is this variable "kernel_is_64_bit" is only used during
recording phase for AUX ring buffer, and don't use it for report.  So seems to
me it's over complexity to add a new field and just wander if it's necessary to
save this field as new feature in the perf header.

I think we store the arch, so if the "kernel_is_64_bit" calculation depends only on arch
then I guess we don't need a new feature at the moment.

So, I wasn't suggesting to add this info to the perf.data file header,
just to the in-memory 'struct perf_env'.

And also we should avoid unconditionally initializing things that we may
never need, please structure it as:

static void perf_env__init_kernel_mode(struct perf_env *env)
{
       const char *arch = perf_env__raw_arch(env);

       if (!strncmp(arch, "x86_64", 6)   || !strncmp(arch, "aarch64", 7) ||
           !strncmp(arch, "arm64", 5)    || !strncmp(arch, "mips64", 6) ||
           !strncmp(arch, "parisc64", 8) || !strncmp(arch, "riscv64", 7) ||
           !strncmp(arch, "s390x", 5)    || !strncmp(arch, "sparc64", 7))
               kernel_is_64_bit = 1;
       else
               kernel_is_64_bit = 0;
}


void perf_env__init(struct perf_env *env)
{
	...
	env->kernel_is_64_bit = -1;
	...
}

bool perf_env__kernel_is_64_bit(struct perf_env *env)
{
	if (env->kernel_is_64_bit == -1)
		perf_env__init_kernel_mode(env);

	return env->kernel_is_64_bit;
}

One thing in my TODO is to crack down on the tons of initializations
perf does unconditionally, last time I looked there are lots :-\

- Arnaldo
 

quoted

Combining the comment from Adrian in another email, I think it's good to add
a new field "compat_mode" in the struct perf_env, and this field will be
initialized in build-record.c.  Currently we don't need to save this value into
the perf file, if later we need to use this value for decoding phase, then we
can add a new feature item to save "compat_mode"
into the perf file's header.

quoted

If you have any different idea, please let me know.  Thanks!