Subject: [PATCH] perf srcline: Implement addr2line using libdw

`perf script` can be quite slow when inlines are enabled, by default it
spawns an `addr2line` process and communicates with it via pipes, that
comes with a certain overhead. The other option is to build perf with
libbfd enabled so that its methods are called directly instead, but due
to its licensing the resulting binary cannot be redistribuited.

This commit adds the ability for perf to use libdw to query the source
lines of binaries from the passed addresses. This avoids the process
spawn overhead and because libdw is licensed dually under
GPL-2.0-or-later and LGPL-3.0-or-later it can be distributed by
downstream packagers. Another thing to consider is that if libdebuginfod
was enabled for libdw then it's used to download debug info, it can be
switched off by unsetting the `DEBUGINFOD_URLS` envvar.

Signed-off-by: Martin Rodriguez Reboredo <[email protected]>
---
tools/perf/Makefile.config | 7 +-
tools/perf/util/srcline.c | 277 ++++++++++++++++++++++++++++++++++++-
2 files changed, 281 insertions(+), 3 deletions(-)

diff --git a/tools/perf/Makefile.config b/tools/perf/Makefile.config
index 1fe8df97fe88..ab6d41e7a6b6 100644
--- a/tools/perf/Makefile.config
+++ b/tools/perf/Makefile.config
@@ -947,13 +947,16 @@ ifdef BUILD_NONDISTRO
$(call feature_check,disassembler-init-styled)
endif

- CFLAGS += -DHAVE_LIBBFD_SUPPORT
- CXXFLAGS += -DHAVE_LIBBFD_SUPPORT
+ CFLAGS += -DHAVE_LIBBFD_SUPPORT -DHAVE_SYMBOLIZER_SUPPORT
+ CXXFLAGS += -DHAVE_LIBBFD_SUPPORT -DHAVE_SYMBOLIZER_SUPPORT
ifeq ($(feature-libbfd-buildid), 1)
CFLAGS += -DHAVE_LIBBFD_BUILDID_SUPPORT
else
$(warning Old version of libbfd/binutils things like PE executable profiling will not be available)
endif
+else ifndef NO_LIBDW_DWARF_UNWIND
+ CFLAGS += -DHAVE_SYMBOLIZER_SUPPORT
+ CXXFLAGS += -DHAVE_SYMBOLIZER_SUPPORT
endif

ifndef NO_DEMANGLE
diff --git a/tools/perf/util/srcline.c b/tools/perf/util/srcline.c
index 7addc34afcf5..8117424137cc 100644
--- a/tools/perf/util/srcline.c
+++ b/tools/perf/util/srcline.c
@@ -130,6 +130,8 @@ static struct symbol *new_inline_sym(struct dso *dso,

#define MAX_INLINE_NEST 1024

+#ifdef HAVE_SYMBOLIZER_SUPPORT
+
#ifdef HAVE_LIBBFD_SUPPORT

/*
@@ -372,6 +374,279 @@ void dso__free_a2l(struct dso *dso)

#else /* HAVE_LIBBFD_SUPPORT */

+#include <elfutils/libdwfl.h>
+#include <dwarf.h>
+
+static char *debuginfo_path = NULL;
+
+static const Dwfl_Callbacks offline_callbacks = {
+ /* We use this table for core files too. */
+ .find_elf = dwfl_build_id_find_elf,
+ .find_debuginfo = dwfl_standard_find_debuginfo,
+ .section_address = dwfl_offline_section_address,
+ .debuginfo_path = &debuginfo_path,
+
+};
+
+struct a2l_data {
+ const char *input;
+ Dwarf_Addr addr;
+ Dwarf_Addr bias;
+
+ bool found;
+ const char *filename;
+ const char *funcname;
+ int line;
+
+ Dwfl *dwfl;
+ Dwfl_Module *mod;
+ GElf_Sym **syms;
+};
+
+static const char *get_diename(Dwarf_Die *die)
+{
+ Dwarf_Attribute attr;
+ const char *name;
+
+ name = dwarf_formstring(
+ dwarf_attr_integrate(die, DW_AT_MIPS_linkage_name, &attr) ?:
+ dwarf_attr_integrate(die, DW_AT_linkage_name, &attr));
+
+ if (name == NULL)
+ name = dwarf_diename(die) ?: "??";
+
+ return name;
+}
+
+static void find_address_in_section(struct a2l_data *a2l)
+{
+ Dwarf_Addr addr;
+ Dwfl_Line *line;
+
+ if (a2l->found)
+ return;
+
+ a2l->mod = dwfl_addrmodule(a2l->dwfl, a2l->addr);
+ if (!a2l->mod)
+ return;
+
+ dwfl_module_getdwarf(a2l->mod, &a2l->bias);
+ addr = a2l->addr + a2l->bias;
+
+ line = dwfl_module_getsrc(a2l->mod, addr);
+ if (!line)
+ line = dwfl_getsrc(a2l->dwfl, addr);
+ if (!line)
+ return;
+
+ a2l->filename = dwfl_lineinfo(line, NULL, &a2l->line, NULL, NULL, NULL);
+ a2l->funcname = dwfl_module_addrname(a2l->mod, addr);
+
+ if (a2l->filename && !strlen(a2l->filename))
+ a2l->filename = NULL;
+ else
+ a2l->found = true;
+}
+
+static struct a2l_data *addr2line_init(const char *path)
+{
+ Dwfl *dwfl;
+ struct a2l_data *a2l = NULL;
+
+ dwfl = dwfl_begin(&offline_callbacks);
+ if (!dwfl)
+ goto out;
+
+ if (dwfl_report_offline(dwfl, "", path, -1) == NULL)
+ return NULL;
+
+ a2l = zalloc(sizeof(*a2l));
+ if (a2l == NULL)
+ goto out;
+
+ a2l->dwfl = dwfl;
+ a2l->input = strdup(path);
+ if (a2l->input == NULL)
+ goto out;
+
+ if (dwfl_report_end(dwfl, NULL, NULL))
+ goto out;
+
+ return a2l;
+
+out:
+ if (a2l) {
+ zfree((char **)&a2l->input);
+ free(a2l);
+ }
+ dwfl_end(dwfl);
+ return NULL;
+}
+
+static void addr2line_cleanup(struct a2l_data *a2l)
+{
+ if (a2l->dwfl)
+ dwfl_end(a2l->dwfl);
+ zfree((char **)&a2l->input);
+ zfree(&a2l->syms);
+ free(a2l);
+}
+
+static int inline_list__append_dso_a2l(struct dso *dso,
+ struct inline_node *node,
+ struct symbol *sym)
+{
+ struct a2l_data *a2l = dso->a2l;
+ struct symbol *inline_sym = new_inline_sym(dso, sym, a2l->funcname);
+ char *srcline = NULL;
+
+ if (a2l->filename)
+ srcline = srcline_from_fileline(a2l->filename, a2l->line);
+
+ return inline_list__append(inline_sym, srcline, node);
+}
+
+static int get_inline_function(struct dso *dso, struct inline_node *node,
+ struct symbol *sym)
+{
+ struct a2l_data *a2l = dso->a2l;
+ Dwarf_Addr addr = a2l->addr + a2l->bias;
+ Dwarf_Addr bias = 0;
+ Dwarf_Die *cudie = dwfl_module_addrdie(a2l->mod, addr, &bias);
+
+ Dwarf_Die *scopes = NULL;
+ int nscopes = dwarf_getscopes(cudie, addr - bias, &scopes);
+ if (nscopes < 0)
+ return 0;
+
+ if (nscopes > 0) {
+ Dwarf_Die subroutine;
+ Dwarf_Off dieoff = dwarf_dieoffset(&scopes[0]);
+ dwarf_offdie(dwfl_module_getdwarf(a2l->mod, &bias), dieoff,
+ &subroutine);
+ free(scopes);
+ scopes = NULL;
+
+ nscopes = dwarf_getscopes_die(&subroutine, &scopes);
+ if (nscopes > 1) {
+ Dwarf_Die cu;
+ Dwarf_Files *files;
+ if (dwarf_diecu(&scopes[0], &cu, NULL, NULL) != NULL &&
+ dwarf_getsrcfiles(cudie, &files, NULL) == 0) {
+ for (int i = 0; i < nscopes - 1; i++) {
+ Dwarf_Word val;
+ Dwarf_Attribute attr;
+ Dwarf_Die *die = &scopes[i];
+ if (dwarf_tag(die) !=
+ DW_TAG_inlined_subroutine)
+ continue;
+
+ for (int j = i + 1; j < nscopes; j++) {
+ Dwarf_Die *parent = &scopes[j];
+ int tag = dwarf_tag(parent);
+ if (tag == DW_TAG_inlined_subroutine ||
+ tag == DW_TAG_entry_point ||
+ tag == DW_TAG_subprogram) {
+ a2l->funcname =
+ get_diename(
+ parent);
+ break;
+ }
+ }
+
+ a2l->filename = NULL;
+ a2l->line = 0;
+ if (dwarf_formudata(
+ dwarf_attr(die,
+ DW_AT_call_file,
+ &attr),
+ &val) == 0)
+ a2l->filename = dwarf_filesrc(
+ files, val, NULL, NULL);
+
+ if (dwarf_formudata(
+ dwarf_attr(die,
+ DW_AT_call_line,
+ &attr),
+ &val) == 0)
+ a2l->line = val;
+
+ if (a2l->filename != NULL)
+ if (inline_list__append_dso_a2l(
+ dso, node, sym))
+ return 0;
+ }
+ }
+ }
+ }
+ free(scopes);
+
+ return 1;
+}
+
+static int addr2line(const char *dso_name, u64 addr, char **file,
+ unsigned int *line, struct dso *dso, bool unwind_inlines,
+ struct inline_node *node, struct symbol *sym)
+{
+ int ret = 0;
+ struct a2l_data *a2l = dso->a2l;
+
+ if (!a2l) {
+ dso->a2l = addr2line_init(dso_name);
+ a2l = dso->a2l;
+ }
+
+ if (a2l == NULL) {
+ if (!symbol_conf.disable_add2line_warn)
+ pr_warning("addr2line_init failed for %s\n", dso_name);
+ return 0;
+ }
+
+ a2l->addr = addr;
+ a2l->found = false;
+
+ find_address_in_section(a2l);
+
+ if (!a2l->found)
+ return 0;
+
+ if (unwind_inlines) {
+ if (node && inline_list__append_dso_a2l(dso, node, sym))
+ return 0;
+
+ if (node && !get_inline_function(dso, node, sym))
+ return 0;
+
+ ret = 1;
+ }
+
+ if (file) {
+ *file = a2l->filename ? strdup(a2l->filename) : NULL;
+ ret = *file ? 1 : 0;
+ }
+
+ if (line)
+ *line = (unsigned int)a2l->line;
+
+ return ret;
+}
+
+void dso__free_a2l(struct dso *dso)
+{
+ struct a2l_data *a2l = dso->a2l;
+
+ if (!a2l)
+ return;
+
+ addr2line_cleanup(a2l);
+
+ dso->a2l = NULL;
+}
+
+#endif /* HAVE_LIBBFD_SUPPORT */
+
+#else /* HAVE_SYMBOLIZER_SUPPORT */
+
static int filename_split(char *filename, unsigned int *line_nr)
{
char *sep;
@@ -788,7 +1063,7 @@ void dso__free_a2l(struct dso *dso)
dso->a2l = NULL;
}

-#endif /* HAVE_LIBBFD_SUPPORT */
+#endif /* HAVE_SYMBOLIZER_SUPPORT */

static struct inline_node *addr2inlines(const char *dso_name, u64 addr,
struct dso *dso, struct symbol *sym)
--
2.44.0



2024-04-01 16:56:33

by Ian Rogers

[permalink] [raw]
Subject: Re: [PATCH] perf srcline: Implement addr2line using libdw

On Mon, Apr 1, 2024 at 9:08 AM Martin Rodriguez Reboredo
<[email protected]> wrote:
>
> `perf script` can be quite slow when inlines are enabled, by default it
> spawns an `addr2line` process and communicates with it via pipes, that
> comes with a certain overhead. The other option is to build perf with
> libbfd enabled so that its methods are called directly instead, but due
> to its licensing the resulting binary cannot be redistribuited.
>
> This commit adds the ability for perf to use libdw to query the source
> lines of binaries from the passed addresses. This avoids the process
> spawn overhead and because libdw is licensed dually under
> GPL-2.0-or-later and LGPL-3.0-or-later it can be distributed by
> downstream packagers. Another thing to consider is that if libdebuginfod
> was enabled for libdw then it's used to download debug info, it can be
> switched off by unsetting the `DEBUGINFOD_URLS` envvar.
>
> Signed-off-by: Martin Rodriguez Reboredo <[email protected]>

Awesome sauce! Namhyung was just mentioning the idea to do this to me.
I wonder when this lands we can just work to remove all of the
BUILD_NONDISTRO options, namely libbfd, libiberty, etc. I suspect we
have dead/broken code hiding there.

> ---
> tools/perf/Makefile.config | 7 +-
> tools/perf/util/srcline.c | 277 ++++++++++++++++++++++++++++++++++++-
> 2 files changed, 281 insertions(+), 3 deletions(-)
>
> diff --git a/tools/perf/Makefile.config b/tools/perf/Makefile.config
> index 1fe8df97fe88..ab6d41e7a6b6 100644
> --- a/tools/perf/Makefile.config
> +++ b/tools/perf/Makefile.config
> @@ -947,13 +947,16 @@ ifdef BUILD_NONDISTRO
> $(call feature_check,disassembler-init-styled)
> endif
>
> - CFLAGS += -DHAVE_LIBBFD_SUPPORT
> - CXXFLAGS += -DHAVE_LIBBFD_SUPPORT
> + CFLAGS += -DHAVE_LIBBFD_SUPPORT -DHAVE_SYMBOLIZER_SUPPORT
> + CXXFLAGS += -DHAVE_LIBBFD_SUPPORT -DHAVE_SYMBOLIZER_SUPPORT

What does SYMBOLIZER mean in this context? Shouldn't the code be gated
on say a HAVE_LIBDW?

> ifeq ($(feature-libbfd-buildid), 1)
> CFLAGS += -DHAVE_LIBBFD_BUILDID_SUPPORT
> else
> $(warning Old version of libbfd/binutils things like PE executable profiling will not be available)
> endif
> +else ifndef NO_LIBDW_DWARF_UNWIND
> + CFLAGS += -DHAVE_SYMBOLIZER_SUPPORT
> + CXXFLAGS += -DHAVE_SYMBOLIZER_SUPPORT
> endif
>
> ifndef NO_DEMANGLE
> diff --git a/tools/perf/util/srcline.c b/tools/perf/util/srcline.c
> index 7addc34afcf5..8117424137cc 100644
> --- a/tools/perf/util/srcline.c
> +++ b/tools/perf/util/srcline.c
> @@ -130,6 +130,8 @@ static struct symbol *new_inline_sym(struct dso *dso,
>
> #define MAX_INLINE_NEST 1024
>
> +#ifdef HAVE_SYMBOLIZER_SUPPORT
> +
> #ifdef HAVE_LIBBFD_SUPPORT
>
> /*
> @@ -372,6 +374,279 @@ void dso__free_a2l(struct dso *dso)
>
> #else /* HAVE_LIBBFD_SUPPORT */
>
> +#include <elfutils/libdwfl.h>
> +#include <dwarf.h>
> +
> +static char *debuginfo_path = NULL;
> +
> +static const Dwfl_Callbacks offline_callbacks = {
> + /* We use this table for core files too. */
> + .find_elf = dwfl_build_id_find_elf,
> + .find_debuginfo = dwfl_standard_find_debuginfo,
> + .section_address = dwfl_offline_section_address,
> + .debuginfo_path = &debuginfo_path,
> +
> +};
> +
> +struct a2l_data {

Perhaps libdw_a2l_data to avoid confusion with data used for forked
addr2line. Could you comment the variables? Names like "input" are
fairly generic so you could provide an example of what their values
are. It is also useful to comment when something like a string is
owned by the struct, so that cleaning it up can be checked.

> + const char *input;
> + Dwarf_Addr addr;
> + Dwarf_Addr bias;
> +
> + bool found;
> + const char *filename;
> + const char *funcname;
> + int line;

Moving "found" and "line" later will avoid padding. As this data is
attached to a DSO, does there need to be some kind of locking protocol
for >1 symbolizing the same DSO? Perhaps these should be filled in as
out arguments to avoid this kind of complexity.

Also, there is some DSO clean up happening in:
https://lore.kernel.org/lkml/CAM9d7chqnsDBCVFoK2hSs=22QrXBS=13Px5hGA4hM=ho7CZd2g@mail.gmail.com/
where accessor functions are used for the sake of reference count checking:
https://perf.wiki.kernel.org/index.php/Reference_Count_Checking
which may cause some minor conflicts with this patch.

> +
> + Dwfl *dwfl;
> + Dwfl_Module *mod;
> + GElf_Sym **syms;
> +};
> +
> +static const char *get_diename(Dwarf_Die *die)
> +{
> + Dwarf_Attribute attr;
> + const char *name;
> +
> + name = dwarf_formstring(
> + dwarf_attr_integrate(die, DW_AT_MIPS_linkage_name, &attr) ?:
> + dwarf_attr_integrate(die, DW_AT_linkage_name, &attr));
> +
> + if (name == NULL)
> + name = dwarf_diename(die) ?: "??";
> +
> + return name;
> +}
> +
> +static void find_address_in_section(struct a2l_data *a2l)
> +{
> + Dwarf_Addr addr;
> + Dwfl_Line *line;
> +
> + if (a2l->found)
> + return;
> +
> + a2l->mod = dwfl_addrmodule(a2l->dwfl, a2l->addr);
> + if (!a2l->mod)
> + return;
> +
> + dwfl_module_getdwarf(a2l->mod, &a2l->bias);
> + addr = a2l->addr + a2l->bias;
> +
> + line = dwfl_module_getsrc(a2l->mod, addr);
> + if (!line)
> + line = dwfl_getsrc(a2l->dwfl, addr);
> + if (!line)
> + return;
> +
> + a2l->filename = dwfl_lineinfo(line, NULL, &a2l->line, NULL, NULL, NULL);
> + a2l->funcname = dwfl_module_addrname(a2l->mod, addr);
> +
> + if (a2l->filename && !strlen(a2l->filename))
> + a2l->filename = NULL;
> + else
> + a2l->found = true;
> +}
> +
> +static struct a2l_data *addr2line_init(const char *path)
> +{
> + Dwfl *dwfl;
> + struct a2l_data *a2l = NULL;
> +
> + dwfl = dwfl_begin(&offline_callbacks);
> + if (!dwfl)
> + goto out;
> +
> + if (dwfl_report_offline(dwfl, "", path, -1) == NULL)
> + return NULL;
> +
> + a2l = zalloc(sizeof(*a2l));
> + if (a2l == NULL)
> + goto out;
> +
> + a2l->dwfl = dwfl;
> + a2l->input = strdup(path);
> + if (a2l->input == NULL)
> + goto out;
> +
> + if (dwfl_report_end(dwfl, NULL, NULL))
> + goto out;
> +
> + return a2l;
> +
> +out:
> + if (a2l) {
> + zfree((char **)&a2l->input);
> + free(a2l);
> + }
> + dwfl_end(dwfl);
> + return NULL;
> +}
> +
> +static void addr2line_cleanup(struct a2l_data *a2l)
> +{
> + if (a2l->dwfl)
> + dwfl_end(a2l->dwfl);
> + zfree((char **)&a2l->input);
> + zfree(&a2l->syms);
> + free(a2l);
> +}
> +
> +static int inline_list__append_dso_a2l(struct dso *dso,
> + struct inline_node *node,
> + struct symbol *sym)
> +{
> + struct a2l_data *a2l = dso->a2l;
> + struct symbol *inline_sym = new_inline_sym(dso, sym, a2l->funcname);
> + char *srcline = NULL;
> +
> + if (a2l->filename)
> + srcline = srcline_from_fileline(a2l->filename, a2l->line);
> +
> + return inline_list__append(inline_sym, srcline, node);
> +}
> +
> +static int get_inline_function(struct dso *dso, struct inline_node *node,
> + struct symbol *sym)
> +{
> + struct a2l_data *a2l = dso->a2l;
> + Dwarf_Addr addr = a2l->addr + a2l->bias;
> + Dwarf_Addr bias = 0;
> + Dwarf_Die *cudie = dwfl_module_addrdie(a2l->mod, addr, &bias);
> +
> + Dwarf_Die *scopes = NULL;
> + int nscopes = dwarf_getscopes(cudie, addr - bias, &scopes);
> + if (nscopes < 0)
> + return 0;
> +
> + if (nscopes > 0) {

Try to prefer early return to avoid indenting the code. So I think the
above can just be:

if (nscopes <= 0)
return 0;

and then you don't need this "if (nscopes > 0) {"

> + Dwarf_Die subroutine;
> + Dwarf_Off dieoff = dwarf_dieoffset(&scopes[0]);
> + dwarf_offdie(dwfl_module_getdwarf(a2l->mod, &bias), dieoff,
> + &subroutine);
> + free(scopes);
> + scopes = NULL;

Is this dead code?

> +
> + nscopes = dwarf_getscopes_die(&subroutine, &scopes);
> + if (nscopes > 1) {

Similar early return comment to above to avoid indentation.

Thanks,
Ian

> + Dwarf_Die cu;
> + Dwarf_Files *files;
> + if (dwarf_diecu(&scopes[0], &cu, NULL, NULL) != NULL &&
> + dwarf_getsrcfiles(cudie, &files, NULL) == 0) {
> + for (int i = 0; i < nscopes - 1; i++) {
> + Dwarf_Word val;
> + Dwarf_Attribute attr;
> + Dwarf_Die *die = &scopes[i];
> + if (dwarf_tag(die) !=
> + DW_TAG_inlined_subroutine)
> + continue;
> +
> + for (int j = i + 1; j < nscopes; j++) {
> + Dwarf_Die *parent = &scopes[j];
> + int tag = dwarf_tag(parent);
> + if (tag == DW_TAG_inlined_subroutine ||
> + tag == DW_TAG_entry_point ||
> + tag == DW_TAG_subprogram) {
> + a2l->funcname =
> + get_diename(
> + parent);
> + break;
> + }
> + }
> +
> + a2l->filename = NULL;
> + a2l->line = 0;
> + if (dwarf_formudata(
> + dwarf_attr(die,
> + DW_AT_call_file,
> + &attr),
> + &val) == 0)
> + a2l->filename = dwarf_filesrc(
> + files, val, NULL, NULL);
> +
> + if (dwarf_formudata(
> + dwarf_attr(die,
> + DW_AT_call_line,
> + &attr),
> + &val) == 0)
> + a2l->line = val;
> +
> + if (a2l->filename != NULL)
> + if (inline_list__append_dso_a2l(
> + dso, node, sym))
> + return 0;
> + }
> + }
> + }
> + }
> + free(scopes);
> +
> + return 1;
> +}
> +
> +static int addr2line(const char *dso_name, u64 addr, char **file,
> + unsigned int *line, struct dso *dso, bool unwind_inlines,
> + struct inline_node *node, struct symbol *sym)
> +{
> + int ret = 0;
> + struct a2l_data *a2l = dso->a2l;
> +
> + if (!a2l) {
> + dso->a2l = addr2line_init(dso_name);
> + a2l = dso->a2l;
> + }
> +
> + if (a2l == NULL) {
> + if (!symbol_conf.disable_add2line_warn)
> + pr_warning("addr2line_init failed for %s\n", dso_name);
> + return 0;
> + }
> +
> + a2l->addr = addr;
> + a2l->found = false;
> +
> + find_address_in_section(a2l);
> +
> + if (!a2l->found)
> + return 0;
> +
> + if (unwind_inlines) {
> + if (node && inline_list__append_dso_a2l(dso, node, sym))
> + return 0;
> +
> + if (node && !get_inline_function(dso, node, sym))
> + return 0;
> +
> + ret = 1;
> + }
> +
> + if (file) {
> + *file = a2l->filename ? strdup(a2l->filename) : NULL;
> + ret = *file ? 1 : 0;
> + }
> +
> + if (line)
> + *line = (unsigned int)a2l->line;
> +
> + return ret;
> +}
> +
> +void dso__free_a2l(struct dso *dso)
> +{
> + struct a2l_data *a2l = dso->a2l;
> +
> + if (!a2l)
> + return;
> +
> + addr2line_cleanup(a2l);
> +
> + dso->a2l = NULL;
> +}
> +
> +#endif /* HAVE_LIBBFD_SUPPORT */
> +
> +#else /* HAVE_SYMBOLIZER_SUPPORT */
> +
> static int filename_split(char *filename, unsigned int *line_nr)
> {
> char *sep;
> @@ -788,7 +1063,7 @@ void dso__free_a2l(struct dso *dso)
> dso->a2l = NULL;
> }
>
> -#endif /* HAVE_LIBBFD_SUPPORT */
> +#endif /* HAVE_SYMBOLIZER_SUPPORT */
>
> static struct inline_node *addr2inlines(const char *dso_name, u64 addr,
> struct dso *dso, struct symbol *sym)
> --
> 2.44.0
>

2024-04-01 23:34:04

by Namhyung Kim

[permalink] [raw]
Subject: Re: [PATCH] perf srcline: Implement addr2line using libdw

Hello,

On Mon, Apr 1, 2024 at 9:56 AM Ian Rogers <[email protected]> wrote:
>
> On Mon, Apr 1, 2024 at 9:08 AM Martin Rodriguez Reboredo
> <[email protected]> wrote:
> >
> > `perf script` can be quite slow when inlines are enabled, by default it
> > spawns an `addr2line` process and communicates with it via pipes, that
> > comes with a certain overhead. The other option is to build perf with
> > libbfd enabled so that its methods are called directly instead, but due
> > to its licensing the resulting binary cannot be redistribuited.
> >
> > This commit adds the ability for perf to use libdw to query the source
> > lines of binaries from the passed addresses. This avoids the process
> > spawn overhead and because libdw is licensed dually under
> > GPL-2.0-or-later and LGPL-3.0-or-later it can be distributed by
> > downstream packagers. Another thing to consider is that if libdebuginfod
> > was enabled for libdw then it's used to download debug info, it can be
> > switched off by unsetting the `DEBUGINFOD_URLS` envvar.

Thanks for doing this! Yep, we could unset the env temporarily.

As a general comment, perf has some helper functions for libdw
in util/dwarf-aux.[ch]. Please take a look and use/update them.

> >
> > Signed-off-by: Martin Rodriguez Reboredo <[email protected]>
>
> Awesome sauce! Namhyung was just mentioning the idea to do this to me.
> I wonder when this lands we can just work to remove all of the
> BUILD_NONDISTRO options, namely libbfd, libiberty, etc. I suspect we
> have dead/broken code hiding there.
>
> > ---
> > tools/perf/Makefile.config | 7 +-
> > tools/perf/util/srcline.c | 277 ++++++++++++++++++++++++++++++++++++-
> > 2 files changed, 281 insertions(+), 3 deletions(-)
> >
> > diff --git a/tools/perf/Makefile.config b/tools/perf/Makefile.config
> > index 1fe8df97fe88..ab6d41e7a6b6 100644
> > --- a/tools/perf/Makefile.config
> > +++ b/tools/perf/Makefile.config
> > @@ -947,13 +947,16 @@ ifdef BUILD_NONDISTRO
> > $(call feature_check,disassembler-init-styled)
> > endif
> >
> > - CFLAGS += -DHAVE_LIBBFD_SUPPORT
> > - CXXFLAGS += -DHAVE_LIBBFD_SUPPORT
> > + CFLAGS += -DHAVE_LIBBFD_SUPPORT -DHAVE_SYMBOLIZER_SUPPORT
> > + CXXFLAGS += -DHAVE_LIBBFD_SUPPORT -DHAVE_SYMBOLIZER_SUPPORT
>
> What does SYMBOLIZER mean in this context? Shouldn't the code be gated
> on say a HAVE_LIBDW?
>
> > ifeq ($(feature-libbfd-buildid), 1)
> > CFLAGS += -DHAVE_LIBBFD_BUILDID_SUPPORT
> > else
> > $(warning Old version of libbfd/binutils things like PE executable profiling will not be available)
> > endif
> > +else ifndef NO_LIBDW_DWARF_UNWIND
> > + CFLAGS += -DHAVE_SYMBOLIZER_SUPPORT
> > + CXXFLAGS += -DHAVE_SYMBOLIZER_SUPPORT
> > endif
> >
> > ifndef NO_DEMANGLE
> > diff --git a/tools/perf/util/srcline.c b/tools/perf/util/srcline.c
> > index 7addc34afcf5..8117424137cc 100644
> > --- a/tools/perf/util/srcline.c
> > +++ b/tools/perf/util/srcline.c
> > @@ -130,6 +130,8 @@ static struct symbol *new_inline_sym(struct dso *dso,
> >
> > #define MAX_INLINE_NEST 1024
> >
> > +#ifdef HAVE_SYMBOLIZER_SUPPORT
> > +
> > #ifdef HAVE_LIBBFD_SUPPORT
> >
> > /*
> > @@ -372,6 +374,279 @@ void dso__free_a2l(struct dso *dso)
> >
> > #else /* HAVE_LIBBFD_SUPPORT */
> >
> > +#include <elfutils/libdwfl.h>
> > +#include <dwarf.h>
> > +
> > +static char *debuginfo_path = NULL;
> > +
> > +static const Dwfl_Callbacks offline_callbacks = {
> > + /* We use this table for core files too. */
> > + .find_elf = dwfl_build_id_find_elf,
> > + .find_debuginfo = dwfl_standard_find_debuginfo,
> > + .section_address = dwfl_offline_section_address,
> > + .debuginfo_path = &debuginfo_path,
> > +
> > +};
> > +
> > +struct a2l_data {
>
> Perhaps libdw_a2l_data to avoid confusion with data used for forked
> addr2line. Could you comment the variables? Names like "input" are
> fairly generic so you could provide an example of what their values
> are. It is also useful to comment when something like a string is
> owned by the struct, so that cleaning it up can be checked.
>
> > + const char *input;
> > + Dwarf_Addr addr;
> > + Dwarf_Addr bias;
> > +
> > + bool found;
> > + const char *filename;
> > + const char *funcname;
> > + int line;
>
> Moving "found" and "line" later will avoid padding. As this data is
> attached to a DSO, does there need to be some kind of locking protocol
> for >1 symbolizing the same DSO? Perhaps these should be filled in as
> out arguments to avoid this kind of complexity.

Right, we cannot attach this to a DSO. Maybe only dwfl can be
attached, but we might want to use debuginfo abstraction instead.

Also I don't think we can keep all debuginfo/dwfl descriptors at
once due to the file descriptor limit. But I'm not sure if we have
something for the external addr2line.


>
> Also, there is some DSO clean up happening in:
> https://lore.kernel.org/lkml/CAM9d7chqnsDBCVFoK2hSs=22QrXBS=13Px5hGA4hM=ho7CZd2g@mail.gmail.com/
> where accessor functions are used for the sake of reference count checking:
> https://perf.wiki.kernel.org/index.php/Reference_Count_Checking
> which may cause some minor conflicts with this patch.
>
> > +
> > + Dwfl *dwfl;
> > + Dwfl_Module *mod;
> > + GElf_Sym **syms;
> > +};
> > +
> > +static const char *get_diename(Dwarf_Die *die)
> > +{
> > + Dwarf_Attribute attr;
> > + const char *name;
> > +
> > + name = dwarf_formstring(
> > + dwarf_attr_integrate(die, DW_AT_MIPS_linkage_name, &attr) ?:
> > + dwarf_attr_integrate(die, DW_AT_linkage_name, &attr));

We have die_get_linkage_name() but it doesn't seem to
handle MIPS_linkage_name though.


> > +
> > + if (name == NULL)
> > + name = dwarf_diename(die) ?: "??";
> > +
> > + return name;
> > +}
> > +
> > +static void find_address_in_section(struct a2l_data *a2l)

Probably we can use cu_find_lineinfo().


> > +{
> > + Dwarf_Addr addr;
> > + Dwfl_Line *line;
> > +
> > + if (a2l->found)
> > + return;
> > +
> > + a2l->mod = dwfl_addrmodule(a2l->dwfl, a2l->addr);
> > + if (!a2l->mod)
> > + return;
> > +
> > + dwfl_module_getdwarf(a2l->mod, &a2l->bias);
> > + addr = a2l->addr + a2l->bias;
> > +
> > + line = dwfl_module_getsrc(a2l->mod, addr);
> > + if (!line)
> > + line = dwfl_getsrc(a2l->dwfl, addr);
> > + if (!line)
> > + return;
> > +
> > + a2l->filename = dwfl_lineinfo(line, NULL, &a2l->line, NULL, NULL, NULL);
> > + a2l->funcname = dwfl_module_addrname(a2l->mod, addr);
> > +
> > + if (a2l->filename && !strlen(a2l->filename))
> > + a2l->filename = NULL;
> > + else
> > + a2l->found = true;
> > +}
> > +
> > +static struct a2l_data *addr2line_init(const char *path)

debuginfo__new()?


> > +{
> > + Dwfl *dwfl;
> > + struct a2l_data *a2l = NULL;
> > +
> > + dwfl = dwfl_begin(&offline_callbacks);
> > + if (!dwfl)
> > + goto out;
> > +
> > + if (dwfl_report_offline(dwfl, "", path, -1) == NULL)
> > + return NULL;
> > +
> > + a2l = zalloc(sizeof(*a2l));
> > + if (a2l == NULL)
> > + goto out;
> > +
> > + a2l->dwfl = dwfl;
> > + a2l->input = strdup(path);
> > + if (a2l->input == NULL)
> > + goto out;
> > +
> > + if (dwfl_report_end(dwfl, NULL, NULL))
> > + goto out;
> > +
> > + return a2l;
> > +
> > +out:
> > + if (a2l) {
> > + zfree((char **)&a2l->input);
> > + free(a2l);
> > + }
> > + dwfl_end(dwfl);
> > + return NULL;
> > +}
> > +
> > +static void addr2line_cleanup(struct a2l_data *a2l)
> > +{
> > + if (a2l->dwfl)
> > + dwfl_end(a2l->dwfl);
> > + zfree((char **)&a2l->input);
> > + zfree(&a2l->syms);
> > + free(a2l);
> > +}
> > +
> > +static int inline_list__append_dso_a2l(struct dso *dso,
> > + struct inline_node *node,
> > + struct symbol *sym)
> > +{
> > + struct a2l_data *a2l = dso->a2l;
> > + struct symbol *inline_sym = new_inline_sym(dso, sym, a2l->funcname);
> > + char *srcline = NULL;
> > +
> > + if (a2l->filename)
> > + srcline = srcline_from_fileline(a2l->filename, a2l->line);
> > +
> > + return inline_list__append(inline_sym, srcline, node);
> > +}
> > +
> > +static int get_inline_function(struct dso *dso, struct inline_node *node,
> > + struct symbol *sym)
> > +{
> > + struct a2l_data *a2l = dso->a2l;
> > + Dwarf_Addr addr = a2l->addr + a2l->bias;
> > + Dwarf_Addr bias = 0;
> > + Dwarf_Die *cudie = dwfl_module_addrdie(a2l->mod, addr, &bias);
> > +
> > + Dwarf_Die *scopes = NULL;
> > + int nscopes = dwarf_getscopes(cudie, addr - bias, &scopes);

It's not clear to me how this dwarf_getscopes() and later
dwarf_getscopes_die() work together. Can you please add some
comment? Also we have die_get_scopes() and I think it's simpler.


> > + if (nscopes < 0)
> > + return 0;
> > +
> > + if (nscopes > 0) {
>
> Try to prefer early return to avoid indenting the code. So I think the
> above can just be:
>
> if (nscopes <= 0)
> return 0;
>
> and then you don't need this "if (nscopes > 0) {"
>
> > + Dwarf_Die subroutine;
> > + Dwarf_Off dieoff = dwarf_dieoffset(&scopes[0]);
> > + dwarf_offdie(dwfl_module_getdwarf(a2l->mod, &bias), dieoff,
> > + &subroutine);
> > + free(scopes);
> > + scopes = NULL;
>
> Is this dead code?

It seems you can use zfree().


>
> > +
> > + nscopes = dwarf_getscopes_die(&subroutine, &scopes);
> > + if (nscopes > 1) {
>
> Similar early return comment to above to avoid indentation.
>
> Thanks,
> Ian
>
> > + Dwarf_Die cu;
> > + Dwarf_Files *files;
> > + if (dwarf_diecu(&scopes[0], &cu, NULL, NULL) != NULL &&
> > + dwarf_getsrcfiles(cudie, &files, NULL) == 0) {
> > + for (int i = 0; i < nscopes - 1; i++) {
> > + Dwarf_Word val;
> > + Dwarf_Attribute attr;
> > + Dwarf_Die *die = &scopes[i];
> > + if (dwarf_tag(die) !=
> > + DW_TAG_inlined_subroutine)
> > + continue;
> > +
> > + for (int j = i + 1; j < nscopes; j++) {
> > + Dwarf_Die *parent = &scopes[j];
> > + int tag = dwarf_tag(parent);
> > + if (tag == DW_TAG_inlined_subroutine ||
> > + tag == DW_TAG_entry_point ||
> > + tag == DW_TAG_subprogram) {
> > + a2l->funcname =
> > + get_diename(
> > + parent);

Why not get the function name from the abstract origin?


> > + break;
> > + }
> > + }
> > +
> > + a2l->filename = NULL;
> > + a2l->line = 0;
> > + if (dwarf_formudata(
> > + dwarf_attr(die,
> > + DW_AT_call_file,
> > + &attr),
> > + &val) == 0)
> > + a2l->filename = dwarf_filesrc(
> > + files, val, NULL, NULL);
> > +
> > + if (dwarf_formudata(
> > + dwarf_attr(die,
> > + DW_AT_call_line,
> > + &attr),
> > + &val) == 0)
> > + a2l->line = val;

We have die_get_call_file() and die_get_call_lineno().

Thanks,
Namhyung


> > +
> > + if (a2l->filename != NULL)
> > + if (inline_list__append_dso_a2l(
> > + dso, node, sym))
> > + return 0;
> > + }
> > + }
> > + }
> > + }
> > + free(scopes);
> > +
> > + return 1;
> > +}
> > +
> > +static int addr2line(const char *dso_name, u64 addr, char **file,
> > + unsigned int *line, struct dso *dso, bool unwind_inlines,
> > + struct inline_node *node, struct symbol *sym)
> > +{
> > + int ret = 0;
> > + struct a2l_data *a2l = dso->a2l;
> > +
> > + if (!a2l) {
> > + dso->a2l = addr2line_init(dso_name);
> > + a2l = dso->a2l;
> > + }
> > +
> > + if (a2l == NULL) {
> > + if (!symbol_conf.disable_add2line_warn)
> > + pr_warning("addr2line_init failed for %s\n", dso_name);
> > + return 0;
> > + }
> > +
> > + a2l->addr = addr;
> > + a2l->found = false;
> > +
> > + find_address_in_section(a2l);
> > +
> > + if (!a2l->found)
> > + return 0;
> > +
> > + if (unwind_inlines) {
> > + if (node && inline_list__append_dso_a2l(dso, node, sym))
> > + return 0;
> > +
> > + if (node && !get_inline_function(dso, node, sym))
> > + return 0;
> > +
> > + ret = 1;
> > + }
> > +
> > + if (file) {
> > + *file = a2l->filename ? strdup(a2l->filename) : NULL;
> > + ret = *file ? 1 : 0;
> > + }
> > +
> > + if (line)
> > + *line = (unsigned int)a2l->line;
> > +
> > + return ret;
> > +}
> > +
> > +void dso__free_a2l(struct dso *dso)
> > +{
> > + struct a2l_data *a2l = dso->a2l;
> > +
> > + if (!a2l)
> > + return;
> > +
> > + addr2line_cleanup(a2l);
> > +
> > + dso->a2l = NULL;
> > +}
> > +
> > +#endif /* HAVE_LIBBFD_SUPPORT */
> > +
> > +#else /* HAVE_SYMBOLIZER_SUPPORT */
> > +
> > static int filename_split(char *filename, unsigned int *line_nr)
> > {
> > char *sep;
> > @@ -788,7 +1063,7 @@ void dso__free_a2l(struct dso *dso)
> > dso->a2l = NULL;
> > }
> >
> > -#endif /* HAVE_LIBBFD_SUPPORT */
> > +#endif /* HAVE_SYMBOLIZER_SUPPORT */
> >
> > static struct inline_node *addr2inlines(const char *dso_name, u64 addr,
> > struct dso *dso, struct symbol *sym)
> > --
> > 2.44.0
> >

Subject: Re: [PATCH] perf srcline: Implement addr2line using libdw

On 4/1/24 1:56 PM, Ian Rogers wrote:
> On Mon, Apr 1, 2024 at 9:08 AM Martin Rodriguez Reboredo
> <[email protected]> wrote:
> [...]
>
> Awesome sauce! Namhyung was just mentioning the idea to do this to me.
> I wonder when this lands we can just work to remove all of the
> BUILD_NONDISTRO options, namely libbfd, libiberty, etc. I suspect we
> have dead/broken code hiding there.

I thought about the same, though I think there's some disassembler
things to tackle, otherwise it'd be easy to do.

> [...]
>
> What does SYMBOLIZER mean in this context? Shouldn't the code be gated
> on say a HAVE_LIBDW?

Accourding to LLVM `addr2line` is a "symbolizer" program, that
`SYMBOLIZER` means that we have a library for translating an address or
a symbol with an offset into a source file and line.

> [...]
>
> Perhaps libdw_a2l_data to avoid confusion with data used for forked
> addr2line. Could you comment the variables? Names like "input" are
> fairly generic so you could provide an example of what their values
> are. It is also useful to comment when something like a string is
> owned by the struct, so that cleaning it up can be checked.

I've left out some unused and suboptimal fields, a mistake from my part.
Though `filename` and `funcname` come as read-only from `dwfl` so they
don't have to be copied.

>> + const char *input;
>> + Dwarf_Addr addr;
>> + Dwarf_Addr bias;
>> +
>> + bool found;
>> + const char *filename;
>> + const char *funcname;
>> + int line;
>
> Moving "found" and "line" later will avoid padding. As this data is
> attached to a DSO, does there need to be some kind of locking protocol
> for >1 symbolizing the same DSO? Perhaps these should be filled in as
> out arguments to avoid this kind of complexity.

Maybe not, I'm not sure about it.

> Also, there is some DSO clean up happening in:
> https://lore.kernel.org/lkml/CAM9d7chqnsDBCVFoK2hSs=22QrXBS=13Px5hGA4hM=ho7CZd2g@mail.gmail.com/
> where accessor functions are used for the sake of reference count checking:
> https://perf.wiki.kernel.org/index.php/Reference_Count_Checking
> which may cause some minor conflicts with this patch.

Will rebase it, then.

> [...]
>
>> + Dwarf_Die subroutine;
>> + Dwarf_Off dieoff = dwarf_dieoffset(&scopes[0]);
>> + dwarf_offdie(dwfl_module_getdwarf(a2l->mod, &bias), dieoff,
>> + &subroutine);
>> + free(scopes);
>> + scopes = NULL;
>
> Is this dead code?

I don't think so, as the scopes could probably differ in each call, I
will have to investigate.

>> +
>> + nscopes = dwarf_getscopes_die(&subroutine, &scopes);
>> + if (nscopes > 1) {
>
> Similar early return comment to above to avoid indentation.
>
> Thanks,
> Ian
> [...]

Subject: Re: [PATCH] perf srcline: Implement addr2line using libdw

On 4/1/24 8:31 PM, Namhyung Kim wrote:
> [...]
>
> Thanks for doing this! Yep, we could unset the env temporarily.
>
> As a general comment, perf has some helper functions for libdw
> in util/dwarf-aux.[ch]. Please take a look and use/update them.

Whoops! I haven't seen that file, thanks for mentioning it.

> [...]
>
> Probably we can use cu_find_lineinfo().

I could use that, I'll see.

> [...]
>
>>> +static struct a2l_data *addr2line_init(const char *path)
>
> debuginfo__new()?

`libdw_a2l_new` can be another option too.

> [...]
>
>>> +static int get_inline_function(struct dso *dso, struct inline_node *node,
>>> + struct symbol *sym)
>>> +{
>>> + struct a2l_data *a2l = dso->a2l;
>>> + Dwarf_Addr addr = a2l->addr + a2l->bias;
>>> + Dwarf_Addr bias = 0;
>>> + Dwarf_Die *cudie = dwfl_module_addrdie(a2l->mod, addr, &bias);
>>> +
>>> + Dwarf_Die *scopes = NULL;
>>> + int nscopes = dwarf_getscopes(cudie, addr - bias, &scopes);
>
> It's not clear to me how this dwarf_getscopes() and later
> dwarf_getscopes_die() work together. Can you please add some
> comment? Also we have die_get_scopes() and I think it's simpler.

I think the first is to get the scopes at an address and the second is
for the scopes with the addr offset at that DIE. Off the top of my head.

> [...]
>
> Why not get the function name from the abstract origin?

I'm not getting it, where's that abstract origin?

> [...]

2024-04-09 17:03:40

by Namhyung Kim

[permalink] [raw]
Subject: Re: [PATCH] perf srcline: Implement addr2line using libdw

Hello,

On Thu, Apr 4, 2024 at 5:10 PM Martin Rodriguez Reboredo
<[email protected]> wrote:
>
> On 4/1/24 8:31 PM, Namhyung Kim wrote:
> > [...]
> >
> > Thanks for doing this! Yep, we could unset the env temporarily.
> >
> > As a general comment, perf has some helper functions for libdw
> > in util/dwarf-aux.[ch]. Please take a look and use/update them.
>
> Whoops! I haven't seen that file, thanks for mentioning it.
>
> > [...]
> >
> > Probably we can use cu_find_lineinfo().
>
> I could use that, I'll see.
>
> > [...]
> >
> >>> +static struct a2l_data *addr2line_init(const char *path)
> >
> > debuginfo__new()?
>
> `libdw_a2l_new` can be another option too.
>
> > [...]
> >
> >>> +static int get_inline_function(struct dso *dso, struct inline_node *node,
> >>> + struct symbol *sym)
> >>> +{
> >>> + struct a2l_data *a2l = dso->a2l;
> >>> + Dwarf_Addr addr = a2l->addr + a2l->bias;
> >>> + Dwarf_Addr bias = 0;
> >>> + Dwarf_Die *cudie = dwfl_module_addrdie(a2l->mod, addr, &bias);
> >>> +
> >>> + Dwarf_Die *scopes = NULL;
> >>> + int nscopes = dwarf_getscopes(cudie, addr - bias, &scopes);
> >
> > It's not clear to me how this dwarf_getscopes() and later
> > dwarf_getscopes_die() work together. Can you please add some
> > comment? Also we have die_get_scopes() and I think it's simpler.
>
> I think the first is to get the scopes at an address and the second is
> for the scopes with the addr offset at that DIE. Off the top of my head.
>
> > [...]
> >
> > Why not get the function name from the abstract origin?
>
> I'm not getting it, where's that abstract origin?

IIUC DW_TAG_inlined_subroutine has DW_AT_abstract_origin
to point to the original definition of the function. It should have
the name of the function.

<2><5ca8>: Abbrev Number: 40 (DW_TAG_inlined_subroutine)
<5ca9> DW_AT_abstract_origin: <0x5eeb>
<5cad> DW_AT_low_pc : 0x24efc6
<5cb5> DW_AT_high_pc : 0x2d
<5cbd> DW_AT_call_file : 3
<5cbe> DW_AT_call_line : 135
<5cbf> DW_AT_call_column : 9

Thanks,
Namhyung