[PATCH v2] perf callchain: Return directly when use '--call-graph dwarf' under !HAVE_DWARF_SUPPORT

[PATCH v2] perf callchain: Return directly when use '--call-graph dwarf' under !HAVE_DWARF_SUPPORT

[Tiezhu Yang]

DWARF register mappings have not been defined for some architectures,
at least for mips, so we can print an error message and then return
directly when use '--call-graph dwarf'.

E.g. without this patch:

[root@linux perf]# ./perf record --call-graph dwarf cd
Error:
The sys_perf_event_open() syscall returned with 89 (Function not implemented) for event (cycles).
/bin/dmesg | grep -i perf may provide additional information.

With this patch:

[root@linux perf]# ./perf record --call-graph dwarf cd
DWARF is not supported for architecture mips64

 Usage: perf record [<options>] [<command>]
    or: perf record [<options>] -- <command> [<options>]

        --call-graph <record_mode[,record_size]>
                          setup and enables call-graph (stack chain/backtrace):

				record_mode:	call graph recording mode (fp|dwarf|lbr)
				record_size:	if record_mode is 'dwarf', max size of stack recording (<bytes>)
						default: 8192 (bytes)

				Default: fp

Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
---

v2: Use HAVE_DWARF_SUPPORT to check

 tools/perf/util/callchain.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
index 1b60985..ad08554 100644
--- a/tools/perf/util/callchain.c
+++ b/tools/perf/util/callchain.c
@@ -18,6 +18,7 @@
 #include <math.h>
 #include <linux/string.h>
 #include <linux/zalloc.h>
+#include <sys/utsname.h>
 
 #include "asm/bug.h"
 
@@ -276,6 +277,7 @@ int parse_callchain_record(const char *arg, struct callchain_param *param)
 
 		/* Dwarf style */
 		} else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
+#ifdef HAVE_DWARF_SUPPORT
 			const unsigned long default_stack_dump_size = 8192;
 
 			ret = 0;
@@ -290,6 +292,15 @@ int parse_callchain_record(const char *arg, struct callchain_param *param)
 				ret = get_stack_size(tok, &size);
 				param->dump_size = size;
 			}
+#else
+			struct utsname uts;
+
+			ret = uname(&uts);
+			pr_err("DWARF is not supported for architecture %s\n",
+				ret ? "unknown" : uts.machine);
+
+			return -ENOTSUP;
+#endif
 		} else if (!strncmp(name, "lbr", sizeof("lbr"))) {
 			if (!strtok_r(NULL, ",", &saveptr)) {
 				param->record_mode = CALLCHAIN_LBR;
-- 
2.1.0

Re: [PATCH v2] perf callchain: Return directly when use '--call-graph dwarf' under !HAVE_DWARF_SUPPORT

[Arnaldo Carvalho de Melo]

Em Tue, Dec 15, 2020 at 09:35:31AM +0800, Tiezhu Yang escreveu:
> DWARF register mappings have not been defined for some architectures,
> at least for mips, so we can print an error message and then return
> directly when use '--call-graph dwarf'.
> 
> E.g. without this patch:


Thanks, applied.

- Arnaldo

Re: [PATCH v2] perf callchain: Return directly when use '--call-graph dwarf' under !HAVE_DWARF_SUPPORT

[Namhyung Kim]

Hello,

On Tue, Dec 15, 2020 at 10:35 AM Tiezhu Yang <yangtiezhu@loongson.cn> wrote:
>
> DWARF register mappings have not been defined for some architectures,
> at least for mips, so we can print an error message and then return
> directly when use '--call-graph dwarf'.
>
> E.g. without this patch:
>
> [root@linux perf]# ./perf record --call-graph dwarf cd
> Error:
> The sys_perf_event_open() syscall returned with 89 (Function not implemented) for event (cycles).
> /bin/dmesg | grep -i perf may provide additional information.
>
> With this patch:
>
> [root@linux perf]# ./perf record --call-graph dwarf cd
> DWARF is not supported for architecture mips64
>
>  Usage: perf record [<options>] [<command>]
>     or: perf record [<options>] -- <command> [<options>]
>
>         --call-graph <record_mode[,record_size]>
>                           setup and enables call-graph (stack chain/backtrace):
>
>                                 record_mode:    call graph recording mode (fp|dwarf|lbr)
>                                 record_size:    if record_mode is 'dwarf', max size of stack recording (<bytes>)
>                                                 default: 8192 (bytes)
>
>                                 Default: fp
>
> Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
> ---
>
> v2: Use HAVE_DWARF_SUPPORT to check

I'm not sure whether this is because of lack of dwarf library or kernel support.
Based on the error message, I guess it's from the kernel.  Then I think this
patch won't be sufficient..

Thanks,
Namhyung


>
>  tools/perf/util/callchain.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
>
> diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
> index 1b60985..ad08554 100644
> --- a/tools/perf/util/callchain.c
> +++ b/tools/perf/util/callchain.c
> @@ -18,6 +18,7 @@
>  #include <math.h>
>  #include <linux/string.h>
>  #include <linux/zalloc.h>
> +#include <sys/utsname.h>
>
>  #include "asm/bug.h"
>
> @@ -276,6 +277,7 @@ int parse_callchain_record(const char *arg, struct callchain_param *param)
>
>                 /* Dwarf style */
>                 } else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
> +#ifdef HAVE_DWARF_SUPPORT
>                         const unsigned long default_stack_dump_size = 8192;
>
>                         ret = 0;
> @@ -290,6 +292,15 @@ int parse_callchain_record(const char *arg, struct callchain_param *param)
>                                 ret = get_stack_size(tok, &size);
>                                 param->dump_size = size;
>                         }
> +#else
> +                       struct utsname uts;
> +
> +                       ret = uname(&uts);
> +                       pr_err("DWARF is not supported for architecture %s\n",
> +                               ret ? "unknown" : uts.machine);
> +
> +                       return -ENOTSUP;
> +#endif
>                 } else if (!strncmp(name, "lbr", sizeof("lbr"))) {
>                         if (!strtok_r(NULL, ",", &saveptr)) {
>                                 param->record_mode = CALLCHAIN_LBR;
> --
> 2.1.0
>

Re: [PATCH v2] perf callchain: Return directly when use '--call-graph dwarf' under !HAVE_DWARF_SUPPORT

[Arnaldo Carvalho de Melo]

Em Tue, Dec 15, 2020 at 10:49:59PM +0900, Namhyung Kim escreveu:
> Hello,
> 
> On Tue, Dec 15, 2020 at 10:35 AM Tiezhu Yang <yangtiezhu@loongson.cn> wrote:
> >
> > DWARF register mappings have not been defined for some architectures,
> > at least for mips, so we can print an error message and then return
> > directly when use '--call-graph dwarf'.
> >
> > E.g. without this patch:
> >
> > [root@linux perf]# ./perf record --call-graph dwarf cd
> > Error:
> > The sys_perf_event_open() syscall returned with 89 (Function not implemented) for event (cycles).
> > /bin/dmesg | grep -i perf may provide additional information.
> >
> > With this patch:
> >
> > [root@linux perf]# ./perf record --call-graph dwarf cd
> > DWARF is not supported for architecture mips64
> >
> >  Usage: perf record [<options>] [<command>]
> >     or: perf record [<options>] -- <command> [<options>]
> >
> >         --call-graph <record_mode[,record_size]>
> >                           setup and enables call-graph (stack chain/backtrace):
> >
> >                                 record_mode:    call graph recording mode (fp|dwarf|lbr)
> >                                 record_size:    if record_mode is 'dwarf', max size of stack recording (<bytes>)
> >                                                 default: 8192 (bytes)
> >
> >                                 Default: fp
> >
> > Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
> > ---
> >
> > v2: Use HAVE_DWARF_SUPPORT to check
> 
> I'm not sure whether this is because of lack of dwarf library or kernel support.
> Based on the error message, I guess it's from the kernel.  Then I think this
> patch won't be sufficient..

tools/perf/Makefile.config

  ifndef NO_DWARF
    ifeq ($(origin PERF_HAVE_DWARF_REGS), undefined)
      msg := $(warning DWARF register mappings have not been defined for architecture $(SRCARCH), DWARF support disabled);
      NO_DWARF := 1
    else
      CFLAGS += -DHAVE_DWARF_SUPPORT $(LIBDW_CFLAGS)
      LDFLAGS += $(LIBDW_LDFLAGS)
      EXTLIBS += ${DWARFLIBS}
      $(call detected,CONFIG_DWARF)
    endif # PERF_HAVE_DWARF_REGS
  endif # NO_DWARF

[acme@five perf]$ find tools/perf -type f | xargs grep PERF_HAVE_DWARF_REGS
tools/perf/arch/xtensa/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/x86/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/arm64/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/arm/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/s390/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/powerpc/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/sh/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/riscv/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/sparc/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/arch/csky/Makefile:PERF_HAVE_DWARF_REGS := 1
tools/perf/Makefile.config:    ifeq ($(origin PERF_HAVE_DWARF_REGS), undefined)
tools/perf/Makefile.config:    endif # PERF_HAVE_DWARF_REGS
[acme@five perf]$

Ouch:

[acme@five perf]$ cat tools/perf/arch/xtensa/Makefile
# SPDX-License-Identifier: GPL-2.0-only
ifndef NO_DWARF
PERF_HAVE_DWARF_REGS := 1
endif
[acme@five perf]$

So you have a point, only if NO_DWARF is not defined, then
PERF_HAVE_DWARF_REGS is can be used to define HAVE_DWARF_SUPPORT, too
clumsy.

So probably my hunch that this should be done at
evsel__open_strerror() and use perf_missing_features.dwarf_regs is
what we should go...

I'm removing this patch from my current tree, please take a look at:

https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/commit/?h=tmp.perf/core&id=f55c66234c1967f6e56e56c3e084f80b417c124b

For how I did this for another feature that the kernel may or not
support, perf_event_open() fails, it notices that in
perf_missing_features and then later evsel__open_strerror() returns a
sensible warning, reacting to something the running kernel returned.

- Arnaldo