Re: [PATCH v2] perf machine: Fix load kernel symbol with '-k' option

From: Leo Yan
Date: Sat Mar 10 2018 - 04:51:19 EST


Hi Jiri,

On Fri, Mar 09, 2018 at 09:56:00PM +0100, Jiri Olsa wrote:
> On Fri, Mar 09, 2018 at 02:05:23PM +0800, Leo Yan wrote:
> > On Hikey arm64 octa A53 platform, when use command './perf report -v
> > -k vmlinux --stdio' it outputs below error info, and it skips to load
> > kernel symbol and doesn't print symbol for event:
> > Failed to open [kernel.kallsyms]_text, continuing without symbols.
> >
> > The regression is introduced by commit ("8c7f1bb37b29 perf machine: Move
> > kernel mmap name into struct machine"), which changes the logic for
> > machine mmap_name by removing function machine__mmap_name() and always
> > use 'machine->mmap_name'. Comparing difference between
> > machine__mmap_name() and 'machine->mmap_name', the later one includes
> > the string for specified kernel vmlinux string with option '-k' in
> > command, but the old function machine__mmap_name() ignores vmlinux
> > path string. As result, event's mmap file name doesn't match with
> > machine mmap file name anymore and it skips to load kernel symbol from
> > vmlinux file.
> >
> > To resolve this issue, this patch adds extra checking for
> > 'symbol_conf.vmlinux_name', when it has been set string so we can know
> > it includes vmlinux path string specified for option '-k'. For this
> > case it sets 'is_kernel_mmap' to true and run into flow to load kernel
> > symbol from vmlinux.
> >
> > This patch has been verified with two commands: './perf report -v
> > -k vmlinux --stdio' and './perf script -v -F cpu,event,ip,sym,symoff
> > -k vmlinux'.
> >
> > Suggested-by: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> > Signed-off-by: Leo Yan <leo.yan@xxxxxxxxxx>
> > ---
> > tools/perf/util/machine.c | 15 ++++++++++++---
> > 1 file changed, 12 insertions(+), 3 deletions(-)
> >
> > diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
> > index 12b7427..3125871 100644
> > --- a/tools/perf/util/machine.c
> > +++ b/tools/perf/util/machine.c
> > @@ -1299,9 +1299,18 @@ static int machine__process_kernel_mmap_event(struct machine *machine,
> > else
> > kernel_type = DSO_TYPE_GUEST_KERNEL;
> >
> > - is_kernel_mmap = memcmp(event->mmap.filename,
> > - machine->mmap_name,
> > - strlen(machine->mmap_name) - 1) == 0;
> > + /*
> > + * When symbol_conf.vmlinux_name is not NULL, it includes the specified
> > + * kernel vmlinux path with option '-k'. So set 'is_kernel_mmap' to
> > + * true for creating machine symbol map.
> > + */
> > + if (symbol_conf.vmlinux_name)
> > + is_kernel_mmap = true;
> > + else
> > + is_kernel_mmap = memcmp(event->mmap.filename,
> > + machine->mmap_name,
> > + strlen(machine->mmap_name) - 1) == 0;
> > +
> > if (event->mmap.filename[0] == '/' ||
> > (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
> > map = machine__findnew_module_map(machine, event->mmap.start,
>
> right, the mmap gets confused with the vmlinux path, but I wonder
> the fix should be not to include symbol_conf.vmlinux_name in the
> mmap_name like below.. untested

I tested below fixing at my side, and confirm your fixing also can
resolve this issue. After reviewing the change, it's more neat than my
own change . So you could add my test tag :)

Tested-by: Leo Yan <leo.yan@xxxxxxxxxx>

BTW, I have a minor comment for code refactoring, it's up to you if
need or not and if need use a new patch for refactoring, please see below
inline comment.

Thanks,
Leo Yan

> ---
> diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
> index 43fbbee409ec..f0cb72022177 100644
> --- a/tools/perf/util/machine.c
> +++ b/tools/perf/util/machine.c
> @@ -51,15 +51,9 @@ static void machine__threads_init(struct machine *machine)
> static int machine__set_mmap_name(struct machine *machine)
> {
> if (machine__is_host(machine)) {
> - if (symbol_conf.vmlinux_name)
> - machine->mmap_name = strdup(symbol_conf.vmlinux_name);
> - else
> - machine->mmap_name = strdup("[kernel.kallsyms]");
> + machine->mmap_name = strdup("[kernel.kallsyms]");
> } else if (machine__is_default_guest(machine)) {
> - if (symbol_conf.default_guest_vmlinux_name)
> - machine->mmap_name = strdup(symbol_conf.default_guest_vmlinux_name);
> - else
> - machine->mmap_name = strdup("[guest.kernel.kallsyms]");
> + machine->mmap_name = strdup("[guest.kernel.kallsyms]");
> } else {
> if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
> machine->pid) < 0)

Do you think below format is better or not?

static int machine__set_mmap_name(struct machine *machine)
{
- if (machine__is_host(machine)) {
- if (symbol_conf.vmlinux_name)
- machine->mmap_name = strdup(symbol_conf.vmlinux_name);
- else
- machine->mmap_name = strdup("[kernel.kallsyms]");
- } else if (machine__is_default_guest(machine)) {
- if (symbol_conf.default_guest_vmlinux_name)
- machine->mmap_name = strdup(symbol_conf.default_guest_vmlinux_name);
- else
- machine->mmap_name = strdup("[guest.kernel.kallsyms]");
- } else {
- if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
- machine->pid) < 0)
- machine->mmap_name = NULL;
- }
+ if (machine__is_host(machine))
+ machine->mmap_name = strdup("[kernel.kallsyms]");
+ else if (machine__is_default_guest(machine))
+ machine->mmap_name = strdup("[guest.kernel.kallsyms]");
+ else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
+ machine->pid) < 0)
+ machine->mmap_name = NULL;

return machine->mmap_name ? 0 : -ENOMEM;
}

> @@ -794,9 +788,15 @@ static struct dso *machine__get_kernel(struct machine *machine)
> struct dso *kernel;
>
> if (machine__is_host(machine)) {
> + if (symbol_conf.vmlinux_name)
> + vmlinux_name = symbol_conf.vmlinux_name;
> +
> kernel = machine__findnew_kernel(machine, vmlinux_name,
> "[kernel]", DSO_TYPE_KERNEL);
> } else {
> + if (symbol_conf.default_guest_vmlinux_name)
> + vmlinux_name = symbol_conf.default_guest_vmlinux_name;
> +
> kernel = machine__findnew_kernel(machine, vmlinux_name,
> "[guest.kernel]",
> DSO_TYPE_GUEST_KERNEL);