[PATCH 0/5] Fix segfaults related to missing BTF support

[PATCH 0/5] Fix segfaults related to missing BTF support

[Arnaldo Carvalho de Melo]

Hi,

	While looking for reports to fix before release 1.28 I got to
Matthias reports about segfaults in systems where BTF isn't present, so
I introduced a regression test and infrastructure to allow testing
handling such a system, please take a look.

Thanks,

- Arnaldo

Arnaldo Carvalho de Melo (5):
  core: Add method to get the vmlinux BTF filename, allow overriding it via env var
  tests default_vmlinux_btf: Introduce test for using BTF by default
  pahole: Honour exclusive BTF loading
  tests default_vmlinux_btf: Cover the no args segfault too
  core, libctf: Check if constructor arguments are NULL before using them

 dwarves.c                    | 33 ++++++++++++++++++++++++++++++---
 dwarves.h                    |  2 ++
 libctf.c                     |  3 +++
 man-pages/pahole.1           |  4 ++++
 pahole.c                     | 11 +++++++++--
 tests/default_vmlinux_btf.sh | 32 ++++++++++++++++++++++++++++++++
 6 files changed, 80 insertions(+), 5 deletions(-)
 create mode 100755 tests/default_vmlinux_btf.sh

-- 
2.47.0

[PATCH 0/5] Fix segfaults related to missing BTF support

[Arnaldo Carvalho de Melo]

Sorry for the top post: Only the cover letter ended up reaching the
mailing list, so resending to have it there :-\

Hi,

	While looking for reports to fix before release 1.28 I got to
Matthias reports about segfaults in systems where BTF isn't present, so
I introduced a regression test and infrastructure to allow testing
handling such a system, please take a look.

Thanks,

- Arnaldo

Arnaldo Carvalho de Melo (5):
  core: Add method to get the vmlinux BTF filename, allow overriding it via env var
  tests default_vmlinux_btf: Introduce test for using BTF by default
  pahole: Honour exclusive BTF loading
  tests default_vmlinux_btf: Cover the no args segfault too
  core, libctf: Check if constructor arguments are NULL before using them

 dwarves.c                    | 33 ++++++++++++++++++++++++++++++---
 dwarves.h                    |  2 ++
 libctf.c                     |  3 +++
 man-pages/pahole.1           |  4 ++++
 pahole.c                     | 11 +++++++++--
 tests/default_vmlinux_btf.sh | 32 ++++++++++++++++++++++++++++++++
 6 files changed, 80 insertions(+), 5 deletions(-)
 create mode 100755 tests/default_vmlinux_btf.sh

-- 
2.47.0

[PATCH 1/5] core: Add method to get the vmlinux BTF filename, allow overriding it via env var

[Arnaldo Carvalho de Melo]

From: Arnaldo Carvalho de Melo <acme@redhat.com>

In systems where BTF isn't available there were reports that the
simplest pahole call, without any args, segfaults.

To have a proper test before fixing this problem, allow overriding the
/sys/kernel/btf/vmlinux filename, so that even in systems with BTF we an
point it to a invalid location, making pahole think that there is no BTF
available and thus fallback to something that currently segfaults.

Using it:

  $ pahole list_head
  struct list_head {
  	struct list_head *         next;                 /*     0     8 */
  	struct list_head *         prev;                 /*     8     8 */

  	/* size: 16, cachelines: 1, members: 2 */
  	/* last cacheline: 16 bytes */
  };
  $
  $ PAHOLE_VMLINUX_BTF_FILENAME=foobar pahole list_head
  Segmentation fault (core dumped)
  $

Reported-by: Matthias Schwarzott <zzam@gentoo.org>
Cc: Alan Maguire <alan.maguire@oracle.com>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Song Liu <song@kernel.org>
Cc: Yonghong Song <yonghong.song@linux.dev>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 dwarves.c          | 17 +++++++++++++++--
 dwarves.h          |  2 ++
 man-pages/pahole.1 |  4 ++++
 pahole.c           |  2 +-
 4 files changed, 22 insertions(+), 3 deletions(-)

diff --git a/dwarves.c b/dwarves.c
index 14ba4f0c6dd7f90f..5b51b55191177e21 100644
--- a/dwarves.c
+++ b/dwarves.c
@@ -2807,6 +2807,19 @@ static int filename__sprintf_build_id(const char *pathname, char *sbuild_id)
 static int vmlinux_path__nr_entries;
 static char **vmlinux_path;
 
+const char *vmlinux_path__btf_filename(void)
+{
+	static const char *vmlinux_btf;
+
+	if (vmlinux_btf == NULL) {
+		vmlinux_btf = getenv("PAHOLE_VMLINUX_BTF_FILENAME");
+		if (vmlinux_btf == NULL)
+			vmlinux_btf = "/sys/kernel/btf/vmlinux";
+	}
+
+	return vmlinux_btf;
+}
+
 static void vmlinux_path__exit(void)
 {
 	while (--vmlinux_path__nr_entries >= 0)
@@ -2933,7 +2946,7 @@ static int cus__load_running_kernel(struct cus *cus, struct conf_load *conf)
 	int err = 0;
 
 	if ((!conf || conf->format_path == NULL || strncmp(conf->format_path, "btf", 3) == 0) &&
-	    access("/sys/kernel/btf/vmlinux", R_OK) == 0) {
+	    access(vmlinux_path__btf_filename(), R_OK) == 0) {
 		int loader = debugging_formats__loader("btf");
 		if (loader == -1)
 			goto try_elf;
@@ -2941,7 +2954,7 @@ static int cus__load_running_kernel(struct cus *cus, struct conf_load *conf)
 		if (conf && conf->conf_fprintf)
 			conf->conf_fprintf->has_alignment_info = debug_fmt_table[loader]->has_alignment_info;
 
-		if (debug_fmt_table[loader]->load_file(cus, conf, "/sys/kernel/btf/vmlinux") == 0)
+		if (debug_fmt_table[loader]->load_file(cus, conf, vmlinux_path__btf_filename()) == 0)
 			return 0;
 	}
 try_elf:
diff --git a/dwarves.h b/dwarves.h
index 49988f1cf9d45be8..1cb0d629f5265523 100644
--- a/dwarves.h
+++ b/dwarves.h
@@ -1628,6 +1628,8 @@ void dwarves__resolve_cacheline_size(const struct conf_load *conf, uint16_t user
 
 const char *dwarf_tag_name(const uint32_t tag);
 
+const char *vmlinux_path__btf_filename(void);
+
 const char *vmlinux_path__find_running_kernel(void);
 
 struct argp_state;
diff --git a/man-pages/pahole.1 b/man-pages/pahole.1
index bd28259b65429acc..39e7b465c64e7e2c 100644
--- a/man-pages/pahole.1
+++ b/man-pages/pahole.1
@@ -75,6 +75,10 @@ struct list_head {
 };
 $
 .fi
+It is possible to override the /sys/kernel/btf/vmlinux file location by setting
+the PAHOLE_VMLINUX_BTF_FILENAME environment variable. This may be useful for
+testing, scripting when using a different BTF for vmlinux. Used in the pahole
+regression tests.
 
 If BTF is not present and no file is passed, then a vmlinux that matches the
 build-id for the running kernel will be looked up in the usual places,
diff --git a/pahole.c b/pahole.c
index 55d04cf82a3da683..b94cb1a979a6923d 100644
--- a/pahole.c
+++ b/pahole.c
@@ -3754,7 +3754,7 @@ try_sole_arg_as_class_names:
 		if (filename &&
 		    strstarts(filename, "/sys/kernel/btf/") &&
 		    strstr(filename, "/vmlinux") == NULL) {
-			base_btf_file = "/sys/kernel/btf/vmlinux";
+			base_btf_file = vmlinux_path__btf_filename();
 			conf_load.base_btf = btf__parse(base_btf_file, NULL);
 			if (libbpf_get_error(conf_load.base_btf)) {
 				fprintf(stderr, "Failed to parse base BTF '%s': %ld\n",
-- 
2.47.0

[PATCH 2/5] tests default_vmlinux_btf: Introduce test for using BTF by default

[Arnaldo Carvalho de Melo]

From: Arnaldo Carvalho de Melo <acme@redhat.com>

On a system without any debugging info or when one specifies BTF as the
only BTF info desired but no BTF is available (or invalidated using
PAHOLE_VMLINUX_BTF_FILENAME to an invalid/non-existent file), we're
getting a segfault:

  root@x1:/home/acme/git/pahole# export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
  root@x1:/home/acme/git/pahole# pahole --running_kernel_vmlinux
  pahole: couldn't find a vmlinux that matches the running kernel
  HINT: Maybe you're inside a container or missing a debuginfo package?
  root@x1:/home/acme/git/pahole# pahole
  Segmentation fault (core dumped)
  root@x1:/home/acme/git/pahole#

So add a test that checks for that before we fix it:

  root@x1:/home/acme/git/pahole# tests/default_vmlinux_btf.sh
  Default BTF on a system without BTF: FAILED
  root@x1:/home/acme/git/pahole#

Reported-by: Matthias Schwarzott <zzam@gentoo.org>
Cc: Alan Maguire <alan.maguire@oracle.com>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Song Liu <song@kernel.org>
Cc: Yonghong Song <yonghong.song@linux.dev>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 tests/default_vmlinux_btf.sh | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)
 create mode 100755 tests/default_vmlinux_btf.sh

diff --git a/tests/default_vmlinux_btf.sh b/tests/default_vmlinux_btf.sh
new file mode 100755
index 0000000000000000..a9effa2d6d37e0ee
--- /dev/null
+++ b/tests/default_vmlinux_btf.sh
@@ -0,0 +1,21 @@
+#!/bin/bash
+
+echo -n "Default BTF on a system without BTF: "
+
+ulimit -c 0
+
+# To suppress the "Segmentation fault core dumped" message in bash we
+# pipe it to some other command, if it segfaults it will not produce any
+# lines and thus we can infer from the number of lines that the segfault
+# took place, tricky, but couldn't find any other way to check this
+# while suppressing the core dumped message. -acme
+
+nr_lines=$(PAHOLE_VMLINUX_BTF_FILENAME=foobar pahole -F btf list_head 2>&1 | wc -l)
+
+if [ $nr_lines -eq 0 ] ; then
+	echo "FAILED"
+	exit 1
+fi
+
+echo "Ok"
+exit 0
-- 
2.47.0

[PATCH 3/5] pahole: Honour exclusive BTF loading

[Arnaldo Carvalho de Melo]

From: Arnaldo Carvalho de Melo <acme@redhat.com>

When we specify 'btf' and BTF info is available in the system, in the
/sys/kernel/btf/vmlinux file, pahole will find it and use it to get the
type information asked for:

  root@x1:~# perf trace -o /tmp/output -e openat pahole -F btf list_head ; tail -1 /tmp/output
  struct list_head {
  	struct list_head *         next;                 /*     0     8 */
  	struct list_head *         prev;                 /*     8     8 */

  	/* size: 16, cachelines: 1, members: 2 */
  	/* last cacheline: 16 bytes */
  };

       1.579 ( 0.010 ms): pahole/764777 openat(dfd: CWD, filename: "/sys/kernel/btf/vmlinux", flags: RDONLY|CLOEXEC) = 3
  root@x1:~#

But, if the system doesn't have BTF info, which we now can simulate by using an
environment variable it is tryng DWARF after failing for BTF, which isn't what
the user asked for having specified just btf in the -F pahole command line:

  root@x1:~# export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
  root@x1:~# perf trace -o /tmp/output -e openat,access pahole -F btf list_head ; tail -12 /tmp/output
  struct list_head {
  	struct list_head *         next;                 /*     0     8 */
  	struct list_head *         prev;                 /*     8     8 */

  	/* size: 16, cachelines: 1, members: 2 */
  	/* last cacheline: 16 bytes */
  };

  1.643 ( 0.003 ms): pahole/765220 access(filename: "list_head", mode: R)      = -1 ENOENT (No such file or directory)
  1.658 ( 0.002 ms): pahole/765220 access(filename: "non-existent", mode: R)   = -1 ENOENT (No such file or directory)
  1.697 ( 0.018 ms): pahole/765220 openat(dfd: CWD, filename: "/sys/kernel/notes") = 3
  1.801 ( 0.004 ms): pahole/765220 openat(dfd: CWD, filename: "vmlinux")       = -1 ENOENT (No such file or directory)
  1.807 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/boot/vmlinux") = -1 ENOENT (No such file or directory)
  1.927 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/usr/lib/debug/lib/modules/6.11.6-200.fc40.x86_64/vmlinux", flags: RDONLY|CLOEXEC) = 3
  root@x1:~#

So honour the request and fails when just BTF is asked and no BTF file is
found:

  # export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
  # pahole -F btf list_head
  pahole: couldn't find any btf debug information on this system.
  #

Reported-by: Matthias Schwarzott <zzam@gentoo.org>
Cc: Alan Maguire <alan.maguire@oracle.com>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Song Liu <song@kernel.org>
Cc: Yonghong Song <yonghong.song@linux.dev>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 dwarves.c | 15 +++++++++++++--
 pahole.c  |  9 ++++++++-
 2 files changed, 21 insertions(+), 3 deletions(-)

diff --git a/dwarves.c b/dwarves.c
index 5b51b55191177e21..459311e64c8b45b2 100644
--- a/dwarves.c
+++ b/dwarves.c
@@ -2944,9 +2944,17 @@ const char *vmlinux_path__find_running_kernel(void)
 static int cus__load_running_kernel(struct cus *cus, struct conf_load *conf)
 {
 	int err = 0;
+	bool btf_only = false;
 
-	if ((!conf || conf->format_path == NULL || strncmp(conf->format_path, "btf", 3) == 0) &&
-	    access(vmlinux_path__btf_filename(), R_OK) == 0) {
+	if (!conf || conf->format_path == NULL)
+		goto try_btf;
+
+	if (!strstr(conf->format_path, "btf"))
+		goto try_elf;
+
+	btf_only = strcmp(conf->format_path, "btf") == 0;
+try_btf:
+	if (access(vmlinux_path__btf_filename(), R_OK) == 0) {
 		int loader = debugging_formats__loader("btf");
 		if (loader == -1)
 			goto try_elf;
@@ -2958,6 +2966,9 @@ static int cus__load_running_kernel(struct cus *cus, struct conf_load *conf)
 			return 0;
 	}
 try_elf:
+	if (btf_only)
+		return -1;
+
 	elf_version(EV_CURRENT);
 	vmlinux_path__init();
 
diff --git a/pahole.c b/pahole.c
index b94cb1a979a6923d..fa5d8c707bd4717d 100644
--- a/pahole.c
+++ b/pahole.c
@@ -3782,7 +3782,14 @@ try_sole_arg_as_class_names:
 			remaining = argc;
 			goto try_sole_arg_as_class_names;
 		}
-		cus__fprintf_load_files_err(cus, "pahole", argv + remaining, err, stderr);
+		if (argv[remaining] != NULL) {
+			cus__fprintf_load_files_err(cus, "pahole", argv + remaining, err, stderr);
+		} else {
+			fprintf(stderr, "pahole: couldn't find any%s%s debug information on this system.\n",
+					conf_load.format_path ? " " : "",
+					conf_load.format_path ?: "");
+		}
+
 		goto out_cus_delete;
 	}
 
-- 
2.47.0

[PATCH 4/5] tests default_vmlinux_btf: Cover the no args segfault too

[Arnaldo Carvalho de Melo]

From: Arnaldo Carvalho de Melo <acme@redhat.com>

  root@x1:/home/acme/git/pahole# pahole --running_kernel_vmlinux
  pahole: couldn't find a vmlinux that matches the running kernel
  HINT: Maybe you're inside a container or missing a debuginfo package?
  root@x1:/home/acme/git/pahole# tests/default_vmlinux_btf.sh
  Default BTF on a system without BTF: FAILED
  root@x1:/home/acme/git/pahole# pahole
  Segmentation fault (core dumped)
  root@x1:/home/acme/git/pahole#

Now to fix this one as well.

Reported-by: Matthias Schwarzott <zzam@gentoo.org>
Cc: Alan Maguire <alan.maguire@oracle.com>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Song Liu <song@kernel.org>
Cc: Yonghong Song <yonghong.song@linux.dev>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 tests/default_vmlinux_btf.sh | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/tests/default_vmlinux_btf.sh b/tests/default_vmlinux_btf.sh
index a9effa2d6d37e0ee..a855ca7587380934 100755
--- a/tests/default_vmlinux_btf.sh
+++ b/tests/default_vmlinux_btf.sh
@@ -17,5 +17,16 @@ if [ $nr_lines -eq 0 ] ; then
 	exit 1
 fi
 
+# There is also the case where no debugging info is available, be it DWARF of
+# BTF and it segfaults when calling just 'pahole', with no args, so check for
+# that as well
+#
+nr_lines=$(PAHOLE_VMLINUX_BTF_FILENAME=foobar pahole 2>&1 | wc -l)
+
+if [ $nr_lines -eq 0 ] ; then
+	echo "FAILED"
+	exit 1
+fi
+
 echo "Ok"
 exit 0
-- 
2.47.0

[PATCH 5/5] core, libctf: Check if constructor arguments are NULL before using them

[Arnaldo Carvalho de Melo]

From: Arnaldo Carvalho de Melo <acme@redhat.com>

  # pahole
  Segmentation fault (core dumped)
  #

Now we have:

  # ls -la bla
  ls: cannot access 'bla': No such file or directory
  # export PAHOLE_VMLINUX_BTF_FILENAME=bla
  # pahole
  pahole: couldn't find any debug information on this system.
  # pahole list_head
  pahole: couldn't find any debug information on this system.
  # pahole -F btf list_head
  pahole: couldn't find any btf debug information on this system.
  # pahole -F dwarf list_head
  pahole: couldn't find any dwarf debug information on this system.
  # tests/default_vmlinux_btf.sh
  Default BTF on a system without BTF: Ok
  #

Reported-by: Matthias Schwarzott <zzam@gentoo.org>
Cc: Alan Maguire <alan.maguire@oracle.com>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Song Liu <song@kernel.org>
Cc: Yonghong Song <yonghong.song@linux.dev>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 dwarves.c | 3 +++
 libctf.c  | 3 +++
 2 files changed, 6 insertions(+)

diff --git a/dwarves.c b/dwarves.c
index 459311e64c8b45b2..ae512b9f46ddcc0a 100644
--- a/dwarves.c
+++ b/dwarves.c
@@ -777,6 +777,9 @@ struct cu *cu__new(const char *name, uint8_t addr_size,
 		if (cu->use_obstack)
 			obstack_init(&cu->obstack);
 
+		if (name == NULL || filename == NULL)
+			goto out_free;
+
 		cu->name = strdup(name);
 		if (cu->name == NULL)
 			goto out_free;
diff --git a/libctf.c b/libctf.c
index edb34dbf10de09e7..0641b10586527a6e 100644
--- a/libctf.c
+++ b/libctf.c
@@ -179,6 +179,9 @@ struct ctf *ctf__new(const char *filename, Elf *elf)
 	struct ctf *ctf = zalloc(sizeof(*ctf));
 
 	if (ctf != NULL) {
+		if (filename == NULL)
+			goto out_delete;
+
 		ctf->filename = strdup(filename);
 		if (ctf->filename == NULL)
 			goto out_delete;
-- 
2.47.0

Re: [PATCH 3/5] pahole: Honour exclusive BTF loading

[Alan Maguire]

On 19/11/2024 13:40, Arnaldo Carvalho de Melo wrote:
> From: Arnaldo Carvalho de Melo <acme@redhat.com>
> 
> When we specify 'btf' and BTF info is available in the system, in the
> /sys/kernel/btf/vmlinux file, pahole will find it and use it to get the
> type information asked for:
> 
>   root@x1:~# perf trace -o /tmp/output -e openat pahole -F btf list_head ; tail -1 /tmp/output
>   struct list_head {
>   	struct list_head *         next;                 /*     0     8 */
>   	struct list_head *         prev;                 /*     8     8 */
> 
>   	/* size: 16, cachelines: 1, members: 2 */
>   	/* last cacheline: 16 bytes */
>   };
> 
>        1.579 ( 0.010 ms): pahole/764777 openat(dfd: CWD, filename: "/sys/kernel/btf/vmlinux", flags: RDONLY|CLOEXEC) = 3
>   root@x1:~#
> 
> But, if the system doesn't have BTF info, which we now can simulate by using an
> environment variable it is tryng DWARF after failing for BTF, which isn't what
> the user asked for having specified just btf in the -F pahole command line:
> 
>   root@x1:~# export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
>   root@x1:~# perf trace -o /tmp/output -e openat,access pahole -F btf list_head ; tail -12 /tmp/output
>   struct list_head {
>   	struct list_head *         next;                 /*     0     8 */
>   	struct list_head *         prev;                 /*     8     8 */
> 
>   	/* size: 16, cachelines: 1, members: 2 */
>   	/* last cacheline: 16 bytes */
>   };
> 
>   1.643 ( 0.003 ms): pahole/765220 access(filename: "list_head", mode: R)      = -1 ENOENT (No such file or directory)
>   1.658 ( 0.002 ms): pahole/765220 access(filename: "non-existent", mode: R)   = -1 ENOENT (No such file or directory)
>   1.697 ( 0.018 ms): pahole/765220 openat(dfd: CWD, filename: "/sys/kernel/notes") = 3
>   1.801 ( 0.004 ms): pahole/765220 openat(dfd: CWD, filename: "vmlinux")       = -1 ENOENT (No such file or directory)
>   1.807 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/boot/vmlinux") = -1 ENOENT (No such file or directory)
>   1.927 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/usr/lib/debug/lib/modules/6.11.6-200.fc40.x86_64/vmlinux", flags: RDONLY|CLOEXEC) = 3
>   root@x1:~#
> 
> So honour the request and fails when just BTF is asked and no BTF file is
> found:
> 
>   # export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
>   # pahole -F btf list_head
>   pahole: couldn't find any btf debug information on this system.
>   #
> 
> Reported-by: Matthias Schwarzott <zzam@gentoo.org>
> Cc: Alan Maguire <alan.maguire@oracle.com>
> Cc: Andrii Nakryiko <andrii@kernel.org>
> Cc: Eduard Zingerman <eddyz87@gmail.com>
> Cc: Jiri Olsa <jolsa@kernel.org>
> Cc: Song Liu <song@kernel.org>
> Cc: Yonghong Song <yonghong.song@linux.dev>
> Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

For the series (since some of my replies didn't reach the list, apologies)

Reviewed-by: Alan Maguire <alan.maguire@oracle.com>
Tested-by: Alan Maguire <alan.maguire@oracle.com>

Re: [PATCH 3/5] pahole: Honour exclusive BTF loading

[Arnaldo Carvalho de Melo]

On Tue, Nov 19, 2024 at 05:47:21PM +0000, Alan Maguire wrote:
> On 19/11/2024 13:40, Arnaldo Carvalho de Melo wrote:
> > From: Arnaldo Carvalho de Melo <acme@redhat.com>
> > 
> > When we specify 'btf' and BTF info is available in the system, in the
> > /sys/kernel/btf/vmlinux file, pahole will find it and use it to get the
> > type information asked for:
> > 
> >   root@x1:~# perf trace -o /tmp/output -e openat pahole -F btf list_head ; tail -1 /tmp/output
> >   struct list_head {
> >   	struct list_head *         next;                 /*     0     8 */
> >   	struct list_head *         prev;                 /*     8     8 */
> > 
> >   	/* size: 16, cachelines: 1, members: 2 */
> >   	/* last cacheline: 16 bytes */
> >   };
> > 
> >        1.579 ( 0.010 ms): pahole/764777 openat(dfd: CWD, filename: "/sys/kernel/btf/vmlinux", flags: RDONLY|CLOEXEC) = 3
> >   root@x1:~#
> > 
> > But, if the system doesn't have BTF info, which we now can simulate by using an
> > environment variable it is tryng DWARF after failing for BTF, which isn't what
> > the user asked for having specified just btf in the -F pahole command line:
> > 
> >   root@x1:~# export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
> >   root@x1:~# perf trace -o /tmp/output -e openat,access pahole -F btf list_head ; tail -12 /tmp/output
> >   struct list_head {
> >   	struct list_head *         next;                 /*     0     8 */
> >   	struct list_head *         prev;                 /*     8     8 */
> > 
> >   	/* size: 16, cachelines: 1, members: 2 */
> >   	/* last cacheline: 16 bytes */
> >   };
> > 
> >   1.643 ( 0.003 ms): pahole/765220 access(filename: "list_head", mode: R)      = -1 ENOENT (No such file or directory)
> >   1.658 ( 0.002 ms): pahole/765220 access(filename: "non-existent", mode: R)   = -1 ENOENT (No such file or directory)
> >   1.697 ( 0.018 ms): pahole/765220 openat(dfd: CWD, filename: "/sys/kernel/notes") = 3
> >   1.801 ( 0.004 ms): pahole/765220 openat(dfd: CWD, filename: "vmlinux")       = -1 ENOENT (No such file or directory)
> >   1.807 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/boot/vmlinux") = -1 ENOENT (No such file or directory)
> >   1.927 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/usr/lib/debug/lib/modules/6.11.6-200.fc40.x86_64/vmlinux", flags: RDONLY|CLOEXEC) = 3
> >   root@x1:~#
> > 
> > So honour the request and fails when just BTF is asked and no BTF file is
> > found:
> > 
> >   # export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
> >   # pahole -F btf list_head
> >   pahole: couldn't find any btf debug information on this system.
> >   #
> > 
> > Reported-by: Matthias Schwarzott <zzam@gentoo.org>
> > Cc: Alan Maguire <alan.maguire@oracle.com>
> > Cc: Andrii Nakryiko <andrii@kernel.org>
> > Cc: Eduard Zingerman <eddyz87@gmail.com>
> > Cc: Jiri Olsa <jolsa@kernel.org>
> > Cc: Song Liu <song@kernel.org>
> > Cc: Yonghong Song <yonghong.song@linux.dev>
> > Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
> 
> For the series (since some of my replies didn't reach the list, apologies)
> 
> Reviewed-by: Alan Maguire <alan.maguire@oracle.com>
> Tested-by: Alan Maguire <alan.maguire@oracle.com>

Re: [PATCH 3/5] pahole: Honour exclusive BTF loading

[Arnaldo Carvalho de Melo]

On Tue, Nov 19, 2024 at 05:47:21PM +0000, Alan Maguire wrote:
> On 19/11/2024 13:40, Arnaldo Carvalho de Melo wrote:
> > From: Arnaldo Carvalho de Melo <acme@redhat.com>
> > 
> > When we specify 'btf' and BTF info is available in the system, in the
> > /sys/kernel/btf/vmlinux file, pahole will find it and use it to get the
> > type information asked for:
> > 
> >   root@x1:~# perf trace -o /tmp/output -e openat pahole -F btf list_head ; tail -1 /tmp/output
> >   struct list_head {
> >   	struct list_head *         next;                 /*     0     8 */
> >   	struct list_head *         prev;                 /*     8     8 */
> > 
> >   	/* size: 16, cachelines: 1, members: 2 */
> >   	/* last cacheline: 16 bytes */
> >   };
> > 
> >        1.579 ( 0.010 ms): pahole/764777 openat(dfd: CWD, filename: "/sys/kernel/btf/vmlinux", flags: RDONLY|CLOEXEC) = 3
> >   root@x1:~#
> > 
> > But, if the system doesn't have BTF info, which we now can simulate by using an
> > environment variable it is tryng DWARF after failing for BTF, which isn't what
> > the user asked for having specified just btf in the -F pahole command line:
> > 
> >   root@x1:~# export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
> >   root@x1:~# perf trace -o /tmp/output -e openat,access pahole -F btf list_head ; tail -12 /tmp/output
> >   struct list_head {
> >   	struct list_head *         next;                 /*     0     8 */
> >   	struct list_head *         prev;                 /*     8     8 */
> > 
> >   	/* size: 16, cachelines: 1, members: 2 */
> >   	/* last cacheline: 16 bytes */
> >   };
> > 
> >   1.643 ( 0.003 ms): pahole/765220 access(filename: "list_head", mode: R)      = -1 ENOENT (No such file or directory)
> >   1.658 ( 0.002 ms): pahole/765220 access(filename: "non-existent", mode: R)   = -1 ENOENT (No such file or directory)
> >   1.697 ( 0.018 ms): pahole/765220 openat(dfd: CWD, filename: "/sys/kernel/notes") = 3
> >   1.801 ( 0.004 ms): pahole/765220 openat(dfd: CWD, filename: "vmlinux")       = -1 ENOENT (No such file or directory)
> >   1.807 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/boot/vmlinux") = -1 ENOENT (No such file or directory)
> >   1.927 ( 0.005 ms): pahole/765220 openat(dfd: CWD, filename: "/usr/lib/debug/lib/modules/6.11.6-200.fc40.x86_64/vmlinux", flags: RDONLY|CLOEXEC) = 3
> >   root@x1:~#
> > 
> > So honour the request and fails when just BTF is asked and no BTF file is
> > found:
> > 
> >   # export PAHOLE_VMLINUX_BTF_FILENAME=non-existent
> >   # pahole -F btf list_head
> >   pahole: couldn't find any btf debug information on this system.
> >   #
> > 
> > Reported-by: Matthias Schwarzott <zzam@gentoo.org>
> > Cc: Alan Maguire <alan.maguire@oracle.com>
> > Cc: Andrii Nakryiko <andrii@kernel.org>
> > Cc: Eduard Zingerman <eddyz87@gmail.com>
> > Cc: Jiri Olsa <jolsa@kernel.org>
> > Cc: Song Liu <song@kernel.org>
> > Cc: Yonghong Song <yonghong.song@linux.dev>
> > Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
> 
> For the series (since some of my replies didn't reach the list, apologies)
> 
> Reviewed-by: Alan Maguire <alan.maguire@oracle.com>
> Tested-by: Alan Maguire <alan.maguire@oracle.com>

Thanks, added to all csets,

- Arnaldo

Re: [PATCH 3/5] pahole: Honour exclusive BTF loading

[Eduard Zingerman]

On Tue, 2024-11-19 at 10:40 -0300, Arnaldo Carvalho de Melo wrote:

[...]

I tried the following test on top of this patch-set:

----- 8< ----------------------------------------
#!/bin/bash

cat <<EOF | clang -o btf-and-dwarf.o -c -x c -g - || exit 1
struct test_val { int i; } g;
EOF
pahole --btf_encode btf-and-dwarf.o || exit 1
# Note: has to be llvm-strip, otherwise --no-strip-all is not available
llvm-strip -o only-dwarf.o    --no-strip-all -R .BTF -R .BTF.ext	btf-and-dwarf.o
llvm-strip -o only-btf.o      --no-strip-all --strip-debug 		btf-and-dwarf.o

for fmt in none btf dwarf dwarf,btf; do
    for file in no-such-file only-btf.o only-dwarf.o btf-and-dwarf.o; do
        if [ "$fmt" == "none" ]; then
            cmd="pahole"
        else
            cmd="pahole -F ${fmt}"
        fi
        printf -- "------ PAHOLE_VMLINUX_BTF_FILENAME=%-20s %-20s ------\n" "${file}" "${cmd}"
        PAHOLE_VMLINUX_BTF_FILENAME=$file $cmd 2>&1 | grep -P '(test_val|pahole:)'
    done
done
---------------------------------------- >8 -----

And there are a few surprising (to me) results, namely:

------ PAHOLE_VMLINUX_BTF_FILENAME=only-dwarf.o         pahole               ------
pahole: couldn't find any debug information on this system.
...
------ PAHOLE_VMLINUX_BTF_FILENAME=only-dwarf.o         pahole -F dwarf      ------
pahole: couldn't find any debug information on this system.
...
------ PAHOLE_VMLINUX_BTF_FILENAME=btf-and-dwarf.o      pahole -F dwarf      ------
pahole: couldn't find any debug information on this system.
...
------ PAHOLE_VMLINUX_BTF_FILENAME=only-dwarf.o         pahole -F dwarf,btf  ------
pahole: couldn't find any debug information on this system.
...

It looks like DWARF is never loaded.
llvm-dwarfdump shows that DWARF is indeed present in only-dwarf.o and
btf-and-dwarf.o. Is this an expected behaviour?

Re: [PATCH 3/5] pahole: Honour exclusive BTF loading

[Alan Maguire]

On 19/11/2024 22:04, Eduard Zingerman wrote:
> On Tue, 2024-11-19 at 10:40 -0300, Arnaldo Carvalho de Melo wrote:
> 
> [...]
> 
> I tried the following test on top of this patch-set:
> 
> ----- 8< ----------------------------------------
> #!/bin/bash
> 
> cat <<EOF | clang -o btf-and-dwarf.o -c -x c -g - || exit 1
> struct test_val { int i; } g;
> EOF
> pahole --btf_encode btf-and-dwarf.o || exit 1
> # Note: has to be llvm-strip, otherwise --no-strip-all is not available
> llvm-strip -o only-dwarf.o    --no-strip-all -R .BTF -R .BTF.ext	btf-and-dwarf.o
> llvm-strip -o only-btf.o      --no-strip-all --strip-debug 		btf-and-dwarf.o
> 
> for fmt in none btf dwarf dwarf,btf; do
>     for file in no-such-file only-btf.o only-dwarf.o btf-and-dwarf.o; do
>         if [ "$fmt" == "none" ]; then
>             cmd="pahole"
>         else
>             cmd="pahole -F ${fmt}"
>         fi
>         printf -- "------ PAHOLE_VMLINUX_BTF_FILENAME=%-20s %-20s ------\n" "${file}" "${cmd}"
>         PAHOLE_VMLINUX_BTF_FILENAME=$file $cmd 2>&1 | grep -P '(test_val|pahole:)'
>     done
> done

this is a nice test, maybe we could adapt it for inclusion. more below..

> ---------------------------------------- >8 -----
> 
> And there are a few surprising (to me) results, namely:
> 
> ------ PAHOLE_VMLINUX_BTF_FILENAME=only-dwarf.o         pahole               ------
> pahole: couldn't find any debug information on this system.
> ...
> ------ PAHOLE_VMLINUX_BTF_FILENAME=only-dwarf.o         pahole -F dwarf      ------
> pahole: couldn't find any debug information on this system.
> ...
> ------ PAHOLE_VMLINUX_BTF_FILENAME=btf-and-dwarf.o      pahole -F dwarf      ------
> pahole: couldn't find any debug information on this system.
> ...
> ------ PAHOLE_VMLINUX_BTF_FILENAME=only-dwarf.o         pahole -F dwarf,btf  ------
> pahole: couldn't find any debug information on this system.
> ...
> 
> It looks like DWARF is never loaded.
> llvm-dwarfdump shows that DWARF is indeed present in only-dwarf.o and
> btf-and-dwarf.o. Is this an expected behaviour?
> 
> 

So my read of the current intent is for PAHOLE_VMLINUX_BTF_FILENAME to
provide an override path for BTF debug info only. In the above cases we
know where to find BTF but don't go on to assume there is DWARF there
too. I think this behaviour would be more surprising if the variable was
called something more generic like PAHOLE_VMLINUX_DEBUGINFO_FILENAME -
i.e. not specific to debug info type - but that's just my take.

We could certainly make the override mechanism more generic to other
debuginfo types, but I think that could be done in follow-ups. What do
you think? Thanks!

Alan

Re: [PATCH 3/5] pahole: Honour exclusive BTF loading

[Eduard Zingerman]

On Tue, 2024-11-19 at 22:28 +0000, Alan Maguire wrote:

[...]

> So my read of the current intent is for PAHOLE_VMLINUX_BTF_FILENAME to
> provide an override path for BTF debug info only. In the above cases we
> know where to find BTF but don't go on to assume there is DWARF there
> too. I think this behaviour would be more surprising if the variable was
> called something more generic like PAHOLE_VMLINUX_DEBUGINFO_FILENAME -
> i.e. not specific to debug info type - but that's just my take.

Oh, right...
My bad, sorry for the noise.

Acked-by: Eduard Zingerman <eddyz87@gmail.com>