[bpf-next,1/2] selftests/bpf: Fix uprobe consumer test

Commit Message

Jiri Olsa Sept. 24, 2024, 11:07 a.m. UTC

With newly merged code the uprobe behaviour is slightly different
and affects uprobe consumer test.

We no longer need to check if the uprobe object is still preserved
after removing last uretprobe, because it stays as long as there's
pending/installed uretprobe instance.

This allows to run uretprobe consumers registered 'after' uprobe was
hit even if previous uretprobe got unregistered before being hit.

The uprobe object will be now removed after the last uprobe ref is
released and in such case it's held by ri->uprobe (return instance)
which is released after the uretprobe is hit.

Reported-by: Ihor Solodrai <[e-mail]>
Closes: https://lore.kernel.org/bpf/w6U8Z9fdhjnkSp2UaFaV1fGqJXvfLEtDKEUyGDkwmoruDJ_AgF_c0FFhrkeKW18OqiP-05s9yDKiT6X-Ns-avN_ABf0dcUkXqbSJN1TQSXo=@pm.me/
Signed-off-by: Jiri Olsa <[e-mail]>
---
 .../testing/selftests/bpf/prog_tests/uprobe_multi_test.c | 9 +--------
 1 file changed, 1 insertion(+), 8 deletions(-)

Comments

Ihor Solodrai Sept. 24, 2024, 5:11 p.m. UTC | #1

On Tuesday, September 24th, 2024 at 4:07 AM, Jiri Olsa <[e-mail]> wrote:

> 
> 
> With newly merged code the uprobe behaviour is slightly different
> and affects uprobe consumer test.
> 
> We no longer need to check if the uprobe object is still preserved
> after removing last uretprobe, because it stays as long as there's
> pending/installed uretprobe instance.
> 
> This allows to run uretprobe consumers registered 'after' uprobe was
> hit even if previous uretprobe got unregistered before being hit.
> 
> The uprobe object will be now removed after the last uprobe ref is
> released and in such case it's held by ri->uprobe (return instance)
> 
> which is released after the uretprobe is hit.
> 
> Reported-by: Ihor Solodrai [e-mail]
> 
> Closes: https://lore.kernel.org/bpf/w6U8Z9fdhjnkSp2UaFaV1fGqJXvfLEtDKEUyGDkwmoruDJ_AgF_c0FFhrkeKW18OqiP-05s9yDKiT6X-Ns-avN_ABf0dcUkXqbSJN1TQSXo=@pm.me/
> Signed-off-by: Jiri Olsa [e-mail]
> 
> ---
> .../testing/selftests/bpf/prog_tests/uprobe_multi_test.c | 9 +--------
> 1 file changed, 1 insertion(+), 8 deletions(-)
> 
> diff --git a/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c b/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c
> index 844f6fc8487b..c1ac813ff9ba 100644
> --- a/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c
> +++ b/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c
> @@ -869,21 +869,14 @@ static void consumer_test(struct uprobe_multi_consumers skel,
> fmt = "prog 0/1: uprobe";
> } else {
> /
> - * uprobe return is tricky ;-)
> - *
> * to trigger uretprobe consumer, the uretprobe needs to be installed,
> * which means one of the 'return' uprobes was alive when probe was hit:
> *
> * idxs: 2/3 uprobe return in 'installed' mask
> - *
> - * in addition if 'after' state removes everything that was installed in
> - * 'before' state, then uprobe kernel object goes away and return uprobe
> - * is not installed and we won't hit it even if it's in 'after' state.
> /
> unsigned long had_uretprobes = before & 0b1100; / is uretprobe installed /
> - unsigned long probe_preserved = before & after; / did uprobe go away */
> 
> - if (had_uretprobes && probe_preserved && test_bit(idx, after))
> + if (had_uretprobes && test_bit(idx, after))
> val++;
> fmt = "idx 2/3: uretprobe";
> }
> --
> 2.46.0

Hi Jiri,

I tested this change on top of bpf-next/master, and
selftests/bpf/vmtest.sh passes.

Thank you for the fix!

Tested-by: Ihor Solodrai <[e-mail]>

[e-mail] Sept. 25, 2024, 10 a.m. UTC | #2

Hello:

This series was applied to bpf/bpf-next.git (master)
by Daniel Borkmann <[e-mail]>:

On Tue, 24 Sep 2024 13:07:30 +0200 you wrote:
> With newly merged code the uprobe behaviour is slightly different
> and affects uprobe consumer test.
> 
> We no longer need to check if the uprobe object is still preserved
> after removing last uretprobe, because it stays as long as there's
> pending/installed uretprobe instance.
> 
> [...]

Here is the summary with links:
  - [bpf-next,1/2] selftests/bpf: Fix uprobe consumer test
    https://git.kernel.org/bpf/bpf-next/c/510c654dfa3b
  - [bpf-next,2/2] selftests/bpf: Bail out quickly from failing consumer test
    https://git.kernel.org/bpf/bpf-next/c/9eac650cd120

You are awesome, thank you!

Patch

diff --git a/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c b/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c
index 844f6fc8487b..c1ac813ff9ba 100644
--- a/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c
+++ b/tools/testing/selftests/bpf/prog_tests/uprobe_multi_test.c
@@ -869,21 +869,14 @@  static void consumer_test(struct uprobe_multi_consumers *skel,
 			fmt = "prog 0/1: uprobe";
 		} else {
 			/*
-			 * uprobe return is tricky ;-)
-			 *
 			 * to trigger uretprobe consumer, the uretprobe needs to be installed,
 			 * which means one of the 'return' uprobes was alive when probe was hit:
 			 *
 			 *   idxs: 2/3 uprobe return in 'installed' mask
-			 *
-			 * in addition if 'after' state removes everything that was installed in
-			 * 'before' state, then uprobe kernel object goes away and return uprobe
-			 * is not installed and we won't hit it even if it's in 'after' state.
 			 */
 			unsigned long had_uretprobes  = before & 0b1100; /* is uretprobe installed */
-			unsigned long probe_preserved = before & after;  /* did uprobe go away */
 
-			if (had_uretprobes && probe_preserved && test_bit(idx, after))
+			if (had_uretprobes && test_bit(idx, after))
 				val++;
 			fmt = "idx 2/3: uretprobe";
 		}