[RFC,5/5] selftests/bpf: Add iter_task_vma_buildid test

  Testing iterator access to build id in vma->vm_file object
by storing each binary with buildid into map and checking
it against buildid retrieved in user space.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 .../selftests/bpf/prog_tests/bpf_iter.c       | 88 +++++++++++++++++++
 .../bpf/progs/bpf_iter_task_vma_buildid.c     | 49 +++++++++++
 2 files changed, 137 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
  

On Wed, Feb 1, 2023 at 5:58 AM Jiri Olsa <jolsa@kernel.org> wrote:
>
> Testing iterator access to build id in vma->vm_file object
> by storing each binary with buildid into map and checking
> it against buildid retrieved in user space.
>
> Signed-off-by: Jiri Olsa <jolsa@kernel.org>
> ---
>  .../selftests/bpf/prog_tests/bpf_iter.c       | 88 +++++++++++++++++++
>  .../bpf/progs/bpf_iter_task_vma_buildid.c     | 49 +++++++++++
>  2 files changed, 137 insertions(+)
>  create mode 100644 tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
>
> diff --git a/tools/testing/selftests/bpf/prog_tests/bpf_iter.c b/tools/testing/selftests/bpf/prog_tests/bpf_iter.c
> index 3af6450763e9..fd3217b68c2e 100644
> --- a/tools/testing/selftests/bpf/prog_tests/bpf_iter.c
> +++ b/tools/testing/selftests/bpf/prog_tests/bpf_iter.c
> @@ -33,6 +33,7 @@
>  #include "bpf_iter_bpf_link.skel.h"
>  #include "bpf_iter_ksym.skel.h"
>  #include "bpf_iter_sockmap.skel.h"
> +#include "bpf_iter_task_vma_buildid.skel.h"
>
>  static int duration;
>
> @@ -1536,6 +1537,91 @@ static void test_task_vma_dead_task(void)
>         bpf_iter_task_vma__destroy(skel);
>  }
>
> +#define D_PATH_BUF_SIZE                1024
> +#define BUILD_ID_SIZE_MAX      20
> +
> +struct build_id {
> +       u32 sz;
> +       char data[BUILD_ID_SIZE_MAX];
> +};
> +
> +#define BUILDID_STR_SIZE (BPF_BUILD_ID_SIZE*2 + 1)
> +
> +static void test_task_vma_buildid(void)
> +{
> +       int err, iter_fd = -1, proc_maps_fd = -1;
> +       struct bpf_iter_task_vma_buildid *skel;
> +       char key[D_PATH_BUF_SIZE], *prev_key;
> +       char bpf_build_id[BUILDID_STR_SIZE];
> +       int len, files_fd, i, cnt = 0;
> +       struct build_id val;
> +       char *build_id;
> +       char c;
> +
> +       skel = bpf_iter_task_vma_buildid__open();
> +       if (!ASSERT_OK_PTR(skel, "bpf_iter_task_vma_buildid__open"))
> +               return;
> +
> +       err = bpf_iter_task_vma_buildid__load(skel);
> +       if (!ASSERT_OK(err, "bpf_iter_task_vma_buildid__load"))
> +               goto out;

minor: you can do __open_and_load() in one step

> +
> +       skel->links.proc_maps = bpf_program__attach_iter(
> +               skel->progs.proc_maps, NULL);
> +
> +       if (!ASSERT_OK_PTR(skel->links.proc_maps, "bpf_program__attach_iter")) {
> +               skel->links.proc_maps = NULL;
> +               goto out;
> +       }
> +
> +       iter_fd = bpf_iter_create(bpf_link__fd(skel->links.proc_maps));
> +       if (!ASSERT_GE(iter_fd, 0, "create_iter"))
> +               goto out;
> +
> +       /* trigger the iterator, there's no output, just map */
> +       len = read(iter_fd, &c, 1);
> +       ASSERT_EQ(len, 0, "len_check");
> +
> +       files_fd = bpf_map__fd(skel->maps.files);
> +
> +       prev_key = NULL;
> +
> +       while (true) {
> +               err = bpf_map_get_next_key(files_fd, prev_key, &key);
> +               if (err) {
> +                       if (errno == ENOENT)
> +                               err = 0;
> +                       break;
> +               }
> +               if (bpf_map_lookup_elem(files_fd, key, &val))
> +                       break;
> +               if (!ASSERT_LE(val.sz, BUILD_ID_SIZE_MAX, "buildid_size"))
> +                       break;
> +
> +               memset(bpf_build_id, 0x0, sizeof(bpf_build_id));
> +               for (i = 0; i < val.sz; i++) {
> +                       sprintf(bpf_build_id + i*2, "%02x",
> +                               (unsigned char) val.data[i]);
> +               }
> +
> +               if (!ASSERT_OK(read_buildid(key, &build_id), "read_buildid"))
> +                       break;
> +
> +               printf("BUILDID %s %s %s\n", bpf_build_id, build_id, key);

debugging leftover or intentional?

> +               ASSERT_OK(strncmp(bpf_build_id, build_id, strlen(bpf_build_id)), "buildid_cmp");
> +
> +               free(build_id);
> +               prev_key = key;
> +               cnt++;
> +       }
> +
> +       printf("checked %d files\n", cnt);

ditto

> +out:
> +       close(proc_maps_fd);
> +       close(iter_fd);
> +       bpf_iter_task_vma_buildid__destroy(skel);
> +}
> +
>  void test_bpf_sockmap_map_iter_fd(void)
>  {
>         struct bpf_iter_sockmap *skel;
> @@ -1659,6 +1745,8 @@ void test_bpf_iter(void)
>                 test_task_vma();
>         if (test__start_subtest("task_vma_dead_task"))
>                 test_task_vma_dead_task();
> +       if (test__start_subtest("task_vma_buildid"))
> +               test_task_vma_buildid();
>         if (test__start_subtest("task_btf"))
>                 test_task_btf();
>         if (test__start_subtest("tcp4"))
> diff --git a/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c b/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
> new file mode 100644
> index 000000000000..25e2179ae5f4
> --- /dev/null
> +++ b/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
> @@ -0,0 +1,49 @@
> +// SPDX-License-Identifier: GPL-2.0
> +
> +#include "bpf_iter.h"
> +#include <bpf/bpf_helpers.h>
> +#include <string.h>
> +
> +char _license[] SEC("license") = "GPL";
> +
> +#define VM_EXEC                0x00000004
> +#define D_PATH_BUF_SIZE        1024
> +
> +struct {
> +       __uint(type, BPF_MAP_TYPE_HASH);
> +       __uint(max_entries, 10000);
> +       __type(key, char[D_PATH_BUF_SIZE]);
> +       __type(value, struct build_id);
> +} files SEC(".maps");
> +
> +static char tmp_key[D_PATH_BUF_SIZE];
> +static struct build_id tmp_data;
> +
> +SEC("iter/task_vma") int proc_maps(struct bpf_iter__task_vma *ctx)

nit: let's keep SEC() on separate line from function itself

> +{
> +       struct vm_area_struct *vma = ctx->vma;
> +       struct seq_file *seq = ctx->meta->seq;
> +       struct task_struct *task = ctx->task;
> +       unsigned long file_key;
> +       struct file *file;
> +
> +       if (task == (void *)0 || vma == (void *)0)
> +               return 0;
> +
> +       if (!(vma->vm_flags & VM_EXEC))
> +               return 0;
> +
> +       file = vma->vm_file;
> +       if (!file)
> +               return 0;
> +
> +       memset(tmp_key, 0x0, D_PATH_BUF_SIZE);

__builtin_memset() to not rely on compiler optimization?

> +       bpf_d_path(&file->f_path, (char *) &tmp_key, D_PATH_BUF_SIZE);
> +
> +       if (bpf_map_lookup_elem(&files, &tmp_key))
> +               return 0;
> +
> +       memcpy(&tmp_data, file->f_bid, sizeof(*file->f_bid));

same about __builtin_memcpy()

> +       bpf_map_update_elem(&files, &tmp_key, &tmp_data, 0);
> +       return 0;
> +}
> --
> 2.39.1
>
  

On Wed, Feb 08, 2023 at 04:01:42PM -0800, Andrii Nakryiko wrote:

SNIP

> > +static void test_task_vma_buildid(void)
> > +{
> > +       int err, iter_fd = -1, proc_maps_fd = -1;
> > +       struct bpf_iter_task_vma_buildid *skel;
> > +       char key[D_PATH_BUF_SIZE], *prev_key;
> > +       char bpf_build_id[BUILDID_STR_SIZE];
> > +       int len, files_fd, i, cnt = 0;
> > +       struct build_id val;
> > +       char *build_id;
> > +       char c;
> > +
> > +       skel = bpf_iter_task_vma_buildid__open();
> > +       if (!ASSERT_OK_PTR(skel, "bpf_iter_task_vma_buildid__open"))
> > +               return;
> > +
> > +       err = bpf_iter_task_vma_buildid__load(skel);
> > +       if (!ASSERT_OK(err, "bpf_iter_task_vma_buildid__load"))
> > +               goto out;
> 
> minor: you can do __open_and_load() in one step

right, I copied that from another test, but removed all the
setup in between, so we can actually call just __open_and_load

SNIP

> > +               memset(bpf_build_id, 0x0, sizeof(bpf_build_id));
> > +               for (i = 0; i < val.sz; i++) {
> > +                       sprintf(bpf_build_id + i*2, "%02x",
> > +                               (unsigned char) val.data[i]);
> > +               }
> > +
> > +               if (!ASSERT_OK(read_buildid(key, &build_id), "read_buildid"))
> > +                       break;
> > +
> > +               printf("BUILDID %s %s %s\n", bpf_build_id, build_id, key);
> 
> debugging leftover or intentional?
> 
> > +               ASSERT_OK(strncmp(bpf_build_id, build_id, strlen(bpf_build_id)), "buildid_cmp");
> > +
> > +               free(build_id);
> > +               prev_key = key;
> > +               cnt++;
> > +       }
> > +
> > +       printf("checked %d files\n", cnt);
> 
> ditto

both intentional, first one can go out I guess, but the
number of checked files seemed interesting to me ;-)

SNIP

> > diff --git a/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c b/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
> > new file mode 100644
> > index 000000000000..25e2179ae5f4
> > --- /dev/null
> > +++ b/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
> > @@ -0,0 +1,49 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +
> > +#include "bpf_iter.h"
> > +#include <bpf/bpf_helpers.h>
> > +#include <string.h>
> > +
> > +char _license[] SEC("license") = "GPL";
> > +
> > +#define VM_EXEC                0x00000004
> > +#define D_PATH_BUF_SIZE        1024
> > +
> > +struct {
> > +       __uint(type, BPF_MAP_TYPE_HASH);
> > +       __uint(max_entries, 10000);
> > +       __type(key, char[D_PATH_BUF_SIZE]);
> > +       __type(value, struct build_id);
> > +} files SEC(".maps");
> > +
> > +static char tmp_key[D_PATH_BUF_SIZE];
> > +static struct build_id tmp_data;
> > +
> > +SEC("iter/task_vma") int proc_maps(struct bpf_iter__task_vma *ctx)
> 
> nit: let's keep SEC() on separate line from function itself

ok

> 
> > +{
> > +       struct vm_area_struct *vma = ctx->vma;
> > +       struct seq_file *seq = ctx->meta->seq;
> > +       struct task_struct *task = ctx->task;
> > +       unsigned long file_key;
> > +       struct file *file;
> > +
> > +       if (task == (void *)0 || vma == (void *)0)
> > +               return 0;
> > +
> > +       if (!(vma->vm_flags & VM_EXEC))
> > +               return 0;
> > +
> > +       file = vma->vm_file;
> > +       if (!file)
> > +               return 0;
> > +
> > +       memset(tmp_key, 0x0, D_PATH_BUF_SIZE);
> 
> __builtin_memset() to not rely on compiler optimization?
> 
> > +       bpf_d_path(&file->f_path, (char *) &tmp_key, D_PATH_BUF_SIZE);
> > +
> > +       if (bpf_map_lookup_elem(&files, &tmp_key))
> > +               return 0;
> > +
> > +       memcpy(&tmp_data, file->f_bid, sizeof(*file->f_bid));
> 
> same about __builtin_memcpy()

ah ok, did not know that, will check.. curious what could
go wrong by using not '__builtin_...' version?

thanks,
jirka
  

On Thu, Feb 9, 2023 at 6:04 AM Jiri Olsa <olsajiri@gmail.com> wrote:
>
> On Wed, Feb 08, 2023 at 04:01:42PM -0800, Andrii Nakryiko wrote:
>
> SNIP
>
> > > +static void test_task_vma_buildid(void)
> > > +{
> > > +       int err, iter_fd = -1, proc_maps_fd = -1;
> > > +       struct bpf_iter_task_vma_buildid *skel;
> > > +       char key[D_PATH_BUF_SIZE], *prev_key;
> > > +       char bpf_build_id[BUILDID_STR_SIZE];
> > > +       int len, files_fd, i, cnt = 0;
> > > +       struct build_id val;
> > > +       char *build_id;
> > > +       char c;
> > > +
> > > +       skel = bpf_iter_task_vma_buildid__open();
> > > +       if (!ASSERT_OK_PTR(skel, "bpf_iter_task_vma_buildid__open"))
> > > +               return;
> > > +
> > > +       err = bpf_iter_task_vma_buildid__load(skel);
> > > +       if (!ASSERT_OK(err, "bpf_iter_task_vma_buildid__load"))
> > > +               goto out;
> >
> > minor: you can do __open_and_load() in one step
>
> right, I copied that from another test, but removed all the
> setup in between, so we can actually call just __open_and_load
>
> SNIP
>
> > > +               memset(bpf_build_id, 0x0, sizeof(bpf_build_id));
> > > +               for (i = 0; i < val.sz; i++) {
> > > +                       sprintf(bpf_build_id + i*2, "%02x",
> > > +                               (unsigned char) val.data[i]);
> > > +               }
> > > +
> > > +               if (!ASSERT_OK(read_buildid(key, &build_id), "read_buildid"))
> > > +                       break;
> > > +
> > > +               printf("BUILDID %s %s %s\n", bpf_build_id, build_id, key);
> >
> > debugging leftover or intentional?
> >
> > > +               ASSERT_OK(strncmp(bpf_build_id, build_id, strlen(bpf_build_id)), "buildid_cmp");
> > > +
> > > +               free(build_id);
> > > +               prev_key = key;
> > > +               cnt++;
> > > +       }
> > > +
> > > +       printf("checked %d files\n", cnt);
> >
> > ditto
>
> both intentional, first one can go out I guess, but the
> number of checked files seemed interesting to me ;-)
>
> SNIP
>
> > > diff --git a/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c b/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
> > > new file mode 100644
> > > index 000000000000..25e2179ae5f4
> > > --- /dev/null
> > > +++ b/tools/testing/selftests/bpf/progs/bpf_iter_task_vma_buildid.c
> > > @@ -0,0 +1,49 @@
> > > +// SPDX-License-Identifier: GPL-2.0
> > > +
> > > +#include "bpf_iter.h"
> > > +#include <bpf/bpf_helpers.h>
> > > +#include <string.h>
> > > +
> > > +char _license[] SEC("license") = "GPL";
> > > +
> > > +#define VM_EXEC                0x00000004
> > > +#define D_PATH_BUF_SIZE        1024
> > > +
> > > +struct {
> > > +       __uint(type, BPF_MAP_TYPE_HASH);
> > > +       __uint(max_entries, 10000);
> > > +       __type(key, char[D_PATH_BUF_SIZE]);
> > > +       __type(value, struct build_id);
> > > +} files SEC(".maps");
> > > +
> > > +static char tmp_key[D_PATH_BUF_SIZE];
> > > +static struct build_id tmp_data;
> > > +
> > > +SEC("iter/task_vma") int proc_maps(struct bpf_iter__task_vma *ctx)
> >
> > nit: let's keep SEC() on separate line from function itself
>
> ok
>
> >
> > > +{
> > > +       struct vm_area_struct *vma = ctx->vma;
> > > +       struct seq_file *seq = ctx->meta->seq;
> > > +       struct task_struct *task = ctx->task;
> > > +       unsigned long file_key;
> > > +       struct file *file;
> > > +
> > > +       if (task == (void *)0 || vma == (void *)0)
> > > +               return 0;
> > > +
> > > +       if (!(vma->vm_flags & VM_EXEC))
> > > +               return 0;
> > > +
> > > +       file = vma->vm_file;
> > > +       if (!file)
> > > +               return 0;
> > > +
> > > +       memset(tmp_key, 0x0, D_PATH_BUF_SIZE);
> >
> > __builtin_memset() to not rely on compiler optimization?
> >
> > > +       bpf_d_path(&file->f_path, (char *) &tmp_key, D_PATH_BUF_SIZE);
> > > +
> > > +       if (bpf_map_lookup_elem(&files, &tmp_key))
> > > +               return 0;
> > > +
> > > +       memcpy(&tmp_data, file->f_bid, sizeof(*file->f_bid));
> >
> > same about __builtin_memcpy()
>
> ah ok, did not know that, will check.. curious what could
> go wrong by using not '__builtin_...' version?

if compiler doesn't optimize it into __builtin_memcpy() (which results
in just explicit assembly code to copy/set data word-by-word), then
BPF program will do actual call to memset(), which with C rules would
be inferred as extern symbol, which would fail BPF object loading with
error along the lines of "couldn't resolve memset extern".

>
> thanks,
> jirka