Re: [PATCHv6 perf/core 09/22] uprobes/x86: Add uprobe syscall to speed up uprobe

[Jiri Olsa <olsajiri@gmail.com>]

On Thu, Sep 04, 2025 at 11:32:06AM -0700, Andrii Nakryiko wrote:
> On Thu, Sep 4, 2025 at 7:03 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> >
> > On Thu, Sep 04, 2025 at 11:39:33AM +0200, Jann Horn wrote:
> > > On Thu, Sep 4, 2025 at 9:56 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> > > > On Wed, Sep 03, 2025 at 04:12:37PM -0700, Andrii Nakryiko wrote:
> > > > > On Wed, Sep 3, 2025 at 2:01 PM Peter Zijlstra <peterz@infradead.org> wrote:
> > > > > >
> > > > > > On Wed, Sep 03, 2025 at 10:56:10PM +0200, Jiri Olsa wrote:
> > > > > >
> > > > > > > > > +SYSCALL_DEFINE0(uprobe)
> > > > > > > > > +{
> > > > > > > > > +       struct pt_regs *regs = task_pt_regs(current);
> > > > > > > > > +       struct uprobe_syscall_args args;
> > > > > > > > > +       unsigned long ip, sp;
> > > > > > > > > +       int err;
> > > > > > > > > +
> > > > > > > > > +       /* Allow execution only from uprobe trampolines. */
> > > > > > > > > +       if (!in_uprobe_trampoline(regs->ip))
> > > > > > > > > +               goto sigill;
> > > > > > > >
> > > > > > > > Hey Jiri,
> > > > > > > >
> > > > > > > > So I've been thinking what's the simplest and most reliable way to
> > > > > > > > feature-detect support for this sys_uprobe (e.g., for libbpf to know
> > > > > > > > whether we should attach at nop5 vs nop1), and clearly that would be
> > > > > > > > to try to call uprobe() syscall not from trampoline, and expect some
> > > > > > > > error code.
> > > > > > > >
> > > > > > > > How bad would it be to change this part to return some unique-enough
> > > > > > > > error code (-ENXIO, -EDOM, whatever).
> > > > > > > >
> > > > > > > > Is there any reason not to do this? Security-wise it will be just fine, right?
> > > > > > >
> > > > > > > good question.. maybe :) the sys_uprobe sigill error path followed the
> > > > > > > uprobe logic when things go bad, seem like good idea to be strict
> > > > > > >
> > > > > > > I understand it'd make the detection code simpler, but it could just
> > > > > > > just fork and check for sigill, right?
> > > > > >
> > > > > > Can't you simply uprobe your own nop5 and read back the text to see what
> > > > > > it turns into?
> > > > >
> > > > > Sure, but none of that is neither fast, nor cheap, nor that simple...
> > > > > (and requires elevated permissions just to detect)
> > > > >
> > > > > Forking is also resource-intensive. (think from libbpf's perspective,
> > > > > it's not cool for library to fork some application just to check such
> > > > > a seemingly simple thing as whether to
> > > > >
> > > > > The question is why all that? That SIGILL when !in_uprobe_trampoline()
> > > > > is just paranoid. I understand killing an application if it tries to
> > > > > screw up "protocol" in all the subsequent checks. But here it's
> > > > > equally secure to just fail that syscall with normal error, instead of
> > > > > punishing by death.
> > > >
> > > > adding Jann to the loop, any thoughts on this ^^^ ?
> > >
> > > If I understand correctly, the main reason for the SIGILL is that if
> > > you hit an error in here when coming from an actual uprobe, and if the
> > > syscall were to just return an error, then you'd end up not restoring
> > > registers as expected which would probably end up crashing the process
> > > in a pretty ugly way?
> >
> > for some cases yes, for the initial checks I think we could just skip
> > the uprobe and process would continue just fine
> >
> 
> For non-buggy kernel implementation in_uprobe_trampoline(regs->ip)
> will (should) always be true when triggered for kernel-installed
> uprobe. So this check can fail only for cases when someone
> intentionally called sys_uprobe not from kernel-generated and
> kernel-controlled trampoline.
> 
> At which point it's totally fine to just return an error and do nothing.
> 
> > we use sigill because the trap code paths use it for errors and to be
> > paranoid about the !in_uprobe_trampoline check
> 
> Yeah, and it should be totally fine to keep doing that.
> 
> It's just about that entry in_uprobe_trampoline() check. And that's
> sufficient to make all this nicely integrated with USDT use cases.
> 
> (I'd say it would be nice to also amend this into original patch to
> avoid someone cherry picking original commit and forgetting/missing
> the follow up change. But that's up to Peter.)
> 
> Jiri, can you please send a quick patch and see how that goes? Thanks!

seems like it's as easy as the change below, I'll send formal patches
later if I don't hear otherwise.. we will also need man page change

jirka


---
diff --git a/arch/x86/kernel/uprobes.c b/arch/x86/kernel/uprobes.c
index 0a8c0a4a5423..845aeaf36b8d 100644
--- a/arch/x86/kernel/uprobes.c
+++ b/arch/x86/kernel/uprobes.c
@@ -810,7 +810,7 @@ SYSCALL_DEFINE0(uprobe)
 
 	/* Allow execution only from uprobe trampolines. */
 	if (!in_uprobe_trampoline(regs->ip))
-		goto sigill;
+		return -ENXIO;
 
 	err = copy_from_user(&args, (void __user *)regs->sp, sizeof(args));
 	if (err)
diff --git a/tools/testing/selftests/bpf/prog_tests/uprobe_syscall.c b/tools/testing/selftests/bpf/prog_tests/uprobe_syscall.c
index 5da0b49eeaca..6d75ede16e7c 100644
--- a/tools/testing/selftests/bpf/prog_tests/uprobe_syscall.c
+++ b/tools/testing/selftests/bpf/prog_tests/uprobe_syscall.c
@@ -757,34 +757,12 @@ static void test_uprobe_race(void)
 #define __NR_uprobe 336
 #endif
 
-static void test_uprobe_sigill(void)
+static void test_uprobe_error(void)
 {
-	int status, err, pid;
+	long err = syscall(__NR_uprobe);
 
-	pid = fork();
-	if (!ASSERT_GE(pid, 0, "fork"))
-		return;
-	/* child */
-	if (pid == 0) {
-		asm volatile (
-			"pushq %rax\n"
-			"pushq %rcx\n"
-			"pushq %r11\n"
-			"movq $" __stringify(__NR_uprobe) ", %rax\n"
-			"syscall\n"
-			"popq %r11\n"
-			"popq %rcx\n"
-			"retq\n"
-		);
-		exit(0);
-	}
-
-	err = waitpid(pid, &status, 0);
-	ASSERT_EQ(err, pid, "waitpid");
-
-	/* verify the child got killed with SIGILL */
-	ASSERT_EQ(WIFSIGNALED(status), 1, "WIFSIGNALED");
-	ASSERT_EQ(WTERMSIG(status), SIGILL, "WTERMSIG");
+	ASSERT_EQ(err, -1, "error");
+	ASSERT_EQ(errno, ENXIO, "errno");
 }
 
 static void __test_uprobe_syscall(void)
@@ -805,8 +783,8 @@ static void __test_uprobe_syscall(void)
 		test_uprobe_usdt();
 	if (test__start_subtest("uprobe_race"))
 		test_uprobe_race();
-	if (test__start_subtest("uprobe_sigill"))
-		test_uprobe_sigill();
+	if (test__start_subtest("uprobe_error"))
+		test_uprobe_error();
 	if (test__start_subtest("uprobe_regs_equal"))
 		test_uprobe_regs_equal(false);
 	if (test__start_subtest("regs_change"))