[PATCH bpf-next] libbpf: Introduce 'custom_btf_path' to 'bpf_obj_open_opts'.

In order to enable the older kernel to use the CO-RE feature, load the
vmlinux btf of the specified path.

Learn from Andrii's comments in [0], add the custom_btf_path parameter
to bpf_obj_open_opts, you can directly use the skeleton's
<objname>_bpf__open_opts function to pass in the custom_btf_path
parameter.

Prior to this, there was also a developer who provided a patch with
similar functions. It is a pity that the follow-up did not continue to
advance. See [1].

	[0]https://lore.kernel.org/bpf/CAEf4BzbJZLjNoiK8_VfeVg_Vrg=9iYFv+po-38SMe=UzwDKJ=Q@mail.gmail.com/#t
	[1]https://yhbt.net/lore/all/CAEf4Bzbgw49w2PtowsrzKQNcxD4fZRE6AKByX-5-dMo-+oWHHA@mail.gmail.com/

Signed-off-by: Shuyi Cheng <redacted>
---
 tools/lib/bpf/libbpf.c | 23 ++++++++++++++++++++---
 tools/lib/bpf/libbpf.h |  6 +++++-
 2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
index 1e04ce7..518b19f 100644
--- a/tools/lib/bpf/libbpf.c
+++ b/tools/lib/bpf/libbpf.c

@@ -509,6 +509,8 @@ struct bpf_object {
 	void *priv;
 	bpf_object_clear_priv_t clear_priv;
 
+	char *custom_btf_path;
+
 	char path[];
 };
 #define obj_elf_valid(o)	((o)->efile.elf)

@@ -2679,8 +2681,15 @@ static int bpf_object__load_vmlinux_btf(struct bpf_object *obj, bool force)
 	if (!force && !obj_needs_vmlinux_btf(obj))
 		return 0;
 
-	obj->btf_vmlinux = libbpf_find_kernel_btf();
-	err = libbpf_get_error(obj->btf_vmlinux);
+	if (obj->custom_btf_path) {
+		obj->btf_vmlinux = btf__parse(obj->custom_btf_path, NULL);
+		err = libbpf_get_error(obj->btf_vmlinux);
+		pr_debug("loading custom vmlinux BTF '%s': %d\n", obj->custom_btf_path, err);
+	} else {
+		obj->btf_vmlinux = libbpf_find_kernel_btf();
+		err = libbpf_get_error(obj->btf_vmlinux);
+	}
+
 	if (err) {
 		pr_warn("Error loading vmlinux BTF: %d\n", err);
 		obj->btf_vmlinux = NULL;

@@ -7554,7 +7563,7 @@ int bpf_program__load(struct bpf_program *prog, char *license, __u32 kern_ver)
 __bpf_object__open(const char *path, const void *obj_buf, size_t obj_buf_sz,
 		   const struct bpf_object_open_opts *opts)
 {
-	const char *obj_name, *kconfig;
+	const char *obj_name, *kconfig, *tmp_btf_path;
 	struct bpf_program *prog;
 	struct bpf_object *obj;
 	char tmp_name[64];

@@ -7584,6 +7593,13 @@ int bpf_program__load(struct bpf_program *prog, char *license, __u32 kern_ver)
 	obj = bpf_object__new(path, obj_buf, obj_buf_sz, obj_name);
 	if (IS_ERR(obj))
 		return obj;
+
+	tmp_btf_path = OPTS_GET(opts, custom_btf_path, NULL);
+	if (tmp_btf_path && strlen(tmp_btf_path) < PATH_MAX) {
+		obj->custom_btf_path = strdup(tmp_btf_path);
+		if (!obj->custom_btf_path)
+			return ERR_PTR(-ENOMEM);
+	}
 
 	kconfig = OPTS_GET(opts, kconfig, NULL);
 	if (kconfig) {

@@ -8702,6 +8718,7 @@ void bpf_object__close(struct bpf_object *obj)
 	for (i = 0; i < obj->nr_maps; i++)
 		bpf_map__destroy(&obj->maps[i]);
 
+	zfree(&obj->custom_btf_path);
 	zfree(&obj->kconfig);
 	zfree(&obj->externs);
 	obj->nr_extern = 0;

diff --git a/tools/lib/bpf/libbpf.h b/tools/lib/bpf/libbpf.h
index 6e61342..16e0f01 100644
--- a/tools/lib/bpf/libbpf.h
+++ b/tools/lib/bpf/libbpf.h

@@ -94,8 +94,12 @@ struct bpf_object_open_opts {
 	 * system Kconfig for CONFIG_xxx externs.
 	 */
 	const char *kconfig;
+	/* Specify the path of vmlinux btf to facilitate the use of CO-RE features
+	 * in the old kernel.
+	 */
+	char *custom_btf_path;
 };
-#define bpf_object_open_opts__last_field kconfig
+#define bpf_object_open_opts__last_field custom_btf_path
 
 LIBBPF_API struct bpf_object *bpf_object__open(const char *path);
 LIBBPF_API struct bpf_object *

-- 
1.8.3.1

On 6/24/21 6:03 AM, Shuyi Cheng wrote:

nit: This should rather go to the 'Parse and load BTF vmlinux if any of [...]'
section of struct bpf_object, and for consistency, I'd keep the btf_ prefix,
like: char *btf_custom_path

Couldn't we do something like (only compile-tested):

diff --git a/tools/lib/bpf/btf.c b/tools/lib/bpf/btf.c
index b46760b93bb4..5b88ce3e483c 100644
--- a/tools/lib/bpf/btf.c
+++ b/tools/lib/bpf/btf.c

@@ -4394,7 +4394,7 @@ static int btf_dedup_remap_types(struct btf_dedup *d)
   * Probe few well-known locations for vmlinux kernel image and try to load BTF
   * data out of it to use for target BTF.
   */
-struct btf *libbpf_find_kernel_btf(void)
+static struct btf *__libbpf_find_kernel_btf(char *btf_custom_path)
  {
  	struct {
  		const char *path_fmt;

@@ -4402,6 +4402,8 @@ struct btf *libbpf_find_kernel_btf(void)
  	} locations[] = {
  		/* try canonical vmlinux BTF through sysfs first */
  		{ "/sys/kernel/btf/vmlinux", true /* raw BTF */ },
+		/* try user defined vmlinux ELF if a path was specified */
+		{ btf_custom_path },
  		/* fall back to trying to find vmlinux ELF on disk otherwise */
  		{ "/boot/vmlinux-%1$s" },
  		{ "/lib/modules/%1$s/vmlinux-%1$s" },

@@ -4419,11 +4421,11 @@ struct btf *libbpf_find_kernel_btf(void)
  	uname(&buf);

  	for (i = 0; i < ARRAY_SIZE(locations); i++) {
+		if (!locations[i].path_fmt)
+			continue;
  		snprintf(path, PATH_MAX, locations[i].path_fmt, buf.release);
-
  		if (access(path, R_OK))
  			continue;
-
  		if (locations[i].raw_btf)
  			btf = btf__parse_raw(path);
  		else

@@ -4440,6 +4442,11 @@ struct btf *libbpf_find_kernel_btf(void)
  	return libbpf_err_ptr(-ESRCH);
  }

+struct btf *libbpf_find_kernel_btf(void)
+{
+	return __libbpf_find_kernel_btf(NULL);
+}
+
  int btf_type_visit_type_ids(struct btf_type *t, type_id_visit_fn visit, void *ctx)
  {
  	int i, n, err;

And then you just call it as:

	obj->btf_vmlinux = __libbpf_find_kernel_btf(obj->btf_custom_path);
	err = libbpf_get_error(obj->btf_vmlinux);

On 6/24/21 11:06 PM, Daniel Borkmann wrote:

On 6/24/21 6:03 AM, Shuyi Cheng wrote:

quoted

In order to enable the older kernel to use the CO-RE feature, load the
vmlinux btf of the specified path.

Learn from Andrii's comments in [0], add the custom_btf_path parameter
to bpf_obj_open_opts, you can directly use the skeleton's
<objname>_bpf__open_opts function to pass in the custom_btf_path
parameter.

Prior to this, there was also a developer who provided a patch with
similar functions. It is a pity that the follow-up did not continue to
advance. See [1].

    [0]https://lore.kernel.org/bpf/CAEf4BzbJZLjNoiK8_VfeVg_Vrg=9iYFv+po-38SMe=UzwDKJ=Q@mail.gmail.com/#t 

    [1]https://yhbt.net/lore/all/CAEf4Bzbgw49w2PtowsrzKQNcxD4fZRE6AKByX-5-dMo-+oWHHA@mail.gmail.com/ 


Signed-off-by: Shuyi Cheng <redacted>
---
  tools/lib/bpf/libbpf.c | 23 ++++++++++++++++++++---
  tools/lib/bpf/libbpf.h |  6 +++++-
  2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
index 1e04ce7..518b19f 100644
--- a/tools/lib/bpf/libbpf.c
+++ b/tools/lib/bpf/libbpf.c

@@ -509,6 +509,8 @@ struct bpf_object {

      void *priv;
      bpf_object_clear_priv_t clear_priv;
+    char *custom_btf_path;
+

nit: This should rather go to the 'Parse and load BTF vmlinux if any of 
[...]'
section of struct bpf_object, and for consistency, I'd keep the btf_ 
prefix,
like: char *btf_custom_path

Thank you very much for your reply.

Agree.

quoted

      char path[];
  };
  #define obj_elf_valid(o)    ((o)->efile.elf)

@@ -2679,8 +2681,15 @@ static int bpf_object__load_vmlinux_btf(struct 

bpf_object *obj, bool force)
      if (!force && !obj_needs_vmlinux_btf(obj))
          return 0;
-    obj->btf_vmlinux = libbpf_find_kernel_btf();
-    err = libbpf_get_error(obj->btf_vmlinux);
+    if (obj->custom_btf_path) {
+        obj->btf_vmlinux = btf__parse(obj->custom_btf_path, NULL);
+        err = libbpf_get_error(obj->btf_vmlinux);
+        pr_debug("loading custom vmlinux BTF '%s': %d\n", 
obj->custom_btf_path, err);
+    } else {
+        obj->btf_vmlinux = libbpf_find_kernel_btf();
+        err = libbpf_get_error(obj->btf_vmlinux);
+    }

Couldn't we do something like (only compile-tested):

Your approach is very inspiring to me. But I did it for two reasons.

1. When the developer specifies btf_custom_path, btf should only be 
loaded from btf_custom_path;
2. Now pahole supports saving vmlinux's btf in raw format, so the old 
kernel can provide btf in elf format or raw format. see [0].

	[0] 
https://git.kernel.org/pub/scm/devel/pahole/pahole.git/tree/pahole.c#n1157

What do you think?
	
Regards,
Shuyi

On Wed, Jun 23, 2021 at 9:04 PM Shuyi Cheng
[off-list ref] wrote:

I think it will be more flexible to treat custom_btf as an vmlinux BTF
override, just like [1] did. I can see how in some situations users
might want to treat this custom BTF as either a replacement of vmlinux
BTF or as an augmentation of vmlinux BTF for the purpose of extra
custom CO-RE relocations (e.g., something along the XDP hints that
were discussed recently). For now it's probably enough to implement
"custom BTF is a replacement for vmlinux BTF" policy and, if
necessary, add "custom BTF is in addition to vmlinux BTF" later with
extra opts flag/field.

Keep in mind that this custom BTF is only useful for BPF CO-RE
relocation. Any other kernel feature relying on vmlinux BTF (e.g.,
fentry) won't work with custom BTF because it expects correct BTF type
IDs.

if strlen() is >= PATH_MAX you'll just silently ignore it? We should
either truncate silently (because PATH_MAX is totally reasonable
assumption) or error out.

On Thu, Jun 24, 2021 at 8:06 AM Daniel Borkmann [off-list ref] wrote:

On 6/24/21 6:03 AM, Shuyi Cheng wrote:

quoted

In order to enable the older kernel to use the CO-RE feature, load the
vmlinux btf of the specified path.

Learn from Andrii's comments in [0], add the custom_btf_path parameter
to bpf_obj_open_opts, you can directly use the skeleton's
<objname>_bpf__open_opts function to pass in the custom_btf_path
parameter.

Prior to this, there was also a developer who provided a patch with
similar functions. It is a pity that the follow-up did not continue to
advance. See [1].

      [0]https://lore.kernel.org/bpf/CAEf4BzbJZLjNoiK8_VfeVg_Vrg=9iYFv+po-38SMe=UzwDKJ=Q@mail.gmail.com/#t
      [1]https://yhbt.net/lore/all/CAEf4Bzbgw49w2PtowsrzKQNcxD4fZRE6AKByX-5-dMo-+oWHHA@mail.gmail.com/

Signed-off-by: Shuyi Cheng <redacted>
---
  tools/lib/bpf/libbpf.c | 23 ++++++++++++++++++++---
  tools/lib/bpf/libbpf.h |  6 +++++-
  2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
index 1e04ce7..518b19f 100644
--- a/tools/lib/bpf/libbpf.c
+++ b/tools/lib/bpf/libbpf.c

@@ -509,6 +509,8 @@ struct bpf_object {
      void *priv;
      bpf_object_clear_priv_t clear_priv;

+     char *custom_btf_path;
+

nit: This should rather go to the 'Parse and load BTF vmlinux if any of [...]'
section of struct bpf_object, and for consistency, I'd keep the btf_ prefix,
like: char *btf_custom_path

quoted

      char path[];
  };
  #define obj_elf_valid(o)    ((o)->efile.elf)

@@ -2679,8 +2681,15 @@ static int bpf_object__load_vmlinux_btf(struct bpf_object *obj, bool force)
      if (!force && !obj_needs_vmlinux_btf(obj))
              return 0;

-     obj->btf_vmlinux = libbpf_find_kernel_btf();
-     err = libbpf_get_error(obj->btf_vmlinux);
+     if (obj->custom_btf_path) {
+             obj->btf_vmlinux = btf__parse(obj->custom_btf_path, NULL);
+             err = libbpf_get_error(obj->btf_vmlinux);
+             pr_debug("loading custom vmlinux BTF '%s': %d\n", obj->custom_btf_path, err);
+     } else {
+             obj->btf_vmlinux = libbpf_find_kernel_btf();
+             err = libbpf_get_error(obj->btf_vmlinux);
+     }

Couldn't we do something like (only compile-tested):

I wonder what are the benefits of this approach, though. My
expectation is that if the user specifies a custom BTF path and BTF is
missing then the whole bpf_object load process should fail, but in
this case it will be silently ignored. Also, if custom BTF is
specified, that custom BTF has to be used even if
/sys/kernel/btf/vmlinux is present, but the patch below will still
prefer /sys/kernel/btf/vmlinux.

So the semantics is different. I'm not saying it's wrong, but I think
it means we need to discuss what behavior we are after first.

[...]

On 7/8/21 4:52 AM, Andrii Nakryiko wrote:

On Wed, Jun 23, 2021 at 9:04 PM Shuyi Cheng
[off-list ref] wrote:

quoted

In order to enable the older kernel to use the CO-RE feature, load the
vmlinux btf of the specified path.

Learn from Andrii's comments in [0], add the custom_btf_path parameter
to bpf_obj_open_opts, you can directly use the skeleton's
<objname>_bpf__open_opts function to pass in the custom_btf_path
parameter.

Prior to this, there was also a developer who provided a patch with
similar functions. It is a pity that the follow-up did not continue to
advance. See [1].

         [0]https://lore.kernel.org/bpf/CAEf4BzbJZLjNoiK8_VfeVg_Vrg=9iYFv+po-38SMe=UzwDKJ=Q@mail.gmail.com/#t
         [1]https://yhbt.net/lore/all/CAEf4Bzbgw49w2PtowsrzKQNcxD4fZRE6AKByX-5-dMo-+oWHHA@mail.gmail.com/

Signed-off-by: Shuyi Cheng <redacted>
---
  tools/lib/bpf/libbpf.c | 23 ++++++++++++++++++++---
  tools/lib/bpf/libbpf.h |  6 +++++-
  2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
index 1e04ce7..518b19f 100644
--- a/tools/lib/bpf/libbpf.c
+++ b/tools/lib/bpf/libbpf.c

@@ -509,6 +509,8 @@ struct bpf_object {
         void *priv;
         bpf_object_clear_priv_t clear_priv;

+       char *custom_btf_path;
+
         char path[];
  };
  #define obj_elf_valid(o)       ((o)->efile.elf)

@@ -2679,8 +2681,15 @@ static int bpf_object__load_vmlinux_btf(struct bpf_object *obj, bool force)
         if (!force && !obj_needs_vmlinux_btf(obj))
                 return 0;

-       obj->btf_vmlinux = libbpf_find_kernel_btf();
-       err = libbpf_get_error(obj->btf_vmlinux);
+       if (obj->custom_btf_path) {
+               obj->btf_vmlinux = btf__parse(obj->custom_btf_path, NULL);
+               err = libbpf_get_error(obj->btf_vmlinux);
+               pr_debug("loading custom vmlinux BTF '%s': %d\n", obj->custom_btf_path, err);
+       } else {
+               obj->btf_vmlinux = libbpf_find_kernel_btf();
+               err = libbpf_get_error(obj->btf_vmlinux);
+       }

I think it will be more flexible to treat custom_btf as an vmlinux BTF
override, just like [1] did. I can see how in some situations users
might want to treat this custom BTF as either a replacement of vmlinux
BTF or as an augmentation of vmlinux BTF for the purpose of extra
custom CO-RE relocations (e.g., something along the XDP hints that
were discussed recently). For now it's probably enough to implement
"custom BTF is a replacement for vmlinux BTF" policy and, if
necessary, add "custom BTF is in addition to vmlinux BTF" later with
extra opts flag/field.

Thank you very much for your reply.

So this part of the code remains the same, right?

Keep in mind that this custom BTF is only useful for BPF CO-RE
relocation. Any other kernel feature relying on vmlinux BTF (e.g.,
fentry) won't work with custom BTF because it expects correct BTF type
IDs.

quoted

+
         if (err) {
                 pr_warn("Error loading vmlinux BTF: %d\n", err);
                 obj->btf_vmlinux = NULL;

@@ -7554,7 +7563,7 @@ int bpf_program__load(struct bpf_program *prog, char *license, __u32 kern_ver)
  __bpf_object__open(const char *path, const void *obj_buf, size_t obj_buf_sz,
                    const struct bpf_object_open_opts *opts)
  {
-       const char *obj_name, *kconfig;
+       const char *obj_name, *kconfig, *tmp_btf_path;
         struct bpf_program *prog;
         struct bpf_object *obj;
         char tmp_name[64];

@@ -7584,6 +7593,13 @@ int bpf_program__load(struct bpf_program *prog, char *license, __u32 kern_ver)
         obj = bpf_object__new(path, obj_buf, obj_buf_sz, obj_name);
         if (IS_ERR(obj))
                 return obj;
+
+       tmp_btf_path = OPTS_GET(opts, custom_btf_path, NULL);
+       if (tmp_btf_path && strlen(tmp_btf_path) < PATH_MAX) {

if strlen() is >= PATH_MAX you'll just silently ignore it? We should
either truncate silently (because PATH_MAX is totally reasonable
assumption) or error out.

Agree. I will send the second version of the patch.

Regards,
Shuyi

quoted

+               obj->custom_btf_path = strdup(tmp_btf_path);
+               if (!obj->custom_btf_path)
+                       return ERR_PTR(-ENOMEM);
+       }

         kconfig = OPTS_GET(opts, kconfig, NULL);
         if (kconfig) {

@@ -8702,6 +8718,7 @@ void bpf_object__close(struct bpf_object *obj)
         for (i = 0; i < obj->nr_maps; i++)
                 bpf_map__destroy(&obj->maps[i]);

+       zfree(&obj->custom_btf_path);
         zfree(&obj->kconfig);
         zfree(&obj->externs);
         obj->nr_extern = 0;

diff --git a/tools/lib/bpf/libbpf.h b/tools/lib/bpf/libbpf.h
index 6e61342..16e0f01 100644
--- a/tools/lib/bpf/libbpf.h
+++ b/tools/lib/bpf/libbpf.h

@@ -94,8 +94,12 @@ struct bpf_object_open_opts {
          * system Kconfig for CONFIG_xxx externs.
          */
         const char *kconfig;
+       /* Specify the path of vmlinux btf to facilitate the use of CO-RE features
+        * in the old kernel.
+        */
+       char *custom_btf_path;
  };
-#define bpf_object_open_opts__last_field kconfig
+#define bpf_object_open_opts__last_field custom_btf_path

  LIBBPF_API struct bpf_object *bpf_object__open(const char *path);
  LIBBPF_API struct bpf_object *
--

1.8.3.1