These macros are a temporary stop-gap until bpf exceptions support
unwinding acquired entities. Basically these macros act as if they take
a callback which only get executed if the assertion fails.

Signed-off-by: Daniel Xu <dxu@dxuuu.xyz>
---
 .../testing/selftests/bpf/bpf_experimental.h  | 22 +++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/tools/testing/selftests/bpf/bpf_experimental.h b/tools/testing/selftests/bpf/bpf_experimental.h
index 1386baf9ae4a..d63f415bef26 100644
--- a/tools/testing/selftests/bpf/bpf_experimental.h
+++ b/tools/testing/selftests/bpf/bpf_experimental.h
@@ -263,6 +263,17 @@ extern void bpf_throw(u64 cookie) __ksym;
  */
 #define bpf_assert(cond) if (!(cond)) bpf_throw(0);
 
+/* Description
+ *	Assert that a conditional expression is true. If false, runs code in the
+ *	body before throwing.
+ * Returns
+ *	Void.
+ * Throws
+ *	An exception with the value zero when the assertion fails.
+ */
+#define bpf_assert_if(cond) \
+	for (int ___i = 0, ___j = !!(cond); !(___j) && !___i; bpf_throw(0), ___i++)
+
 /* Description
  *	Assert that a conditional expression is true.
  * Returns
@@ -272,6 +283,17 @@ extern void bpf_throw(u64 cookie) __ksym;
  */
 #define bpf_assert_with(cond, value) if (!(cond)) bpf_throw(value);
 
+/* Description
+ *	Assert that a conditional expression is true. If false, runs code in the
+ *	body before throwing.
+ * Returns
+ *	Void.
+ * Throws
+ *	An exception with the given value when the assertion fails.
+ */
+#define bpf_assert_with_if(cond, value) \
+	for (int ___i = 0, ___j = !!(cond); !(___j) && !___i; bpf_throw(value), ___i++)
+
 /* Description
  *	Assert that LHS is equal to RHS. This statement updates the known value
  *	of LHS during verification. Note that RHS must be a constant value, and
-- 
2.42.1

On Thu, Dec 14, 2023 at 2:56 PM Daniel Xu <dxu@dxuuu.xyz> wrote:
>
> These macros are a temporary stop-gap until bpf exceptions support
> unwinding acquired entities. Basically these macros act as if they take
> a callback which only get executed if the assertion fails.
>
> Signed-off-by: Daniel Xu <dxu@dxuuu.xyz>
> ---
>  .../testing/selftests/bpf/bpf_experimental.h  | 22 +++++++++++++++++++
>  1 file changed, 22 insertions(+)
>
> diff --git a/tools/testing/selftests/bpf/bpf_experimental.h b/tools/testing/selftests/bpf/bpf_experimental.h
> index 1386baf9ae4a..d63f415bef26 100644
> --- a/tools/testing/selftests/bpf/bpf_experimental.h
> +++ b/tools/testing/selftests/bpf/bpf_experimental.h
> @@ -263,6 +263,17 @@ extern void bpf_throw(u64 cookie) __ksym;
>   */
>  #define bpf_assert(cond) if (!(cond)) bpf_throw(0);
>
> +/* Description
> + *     Assert that a conditional expression is true. If false, runs code in the
> + *     body before throwing.
> + * Returns
> + *     Void.
> + * Throws
> + *     An exception with the value zero when the assertion fails.
> + */
> +#define bpf_assert_if(cond) \
> +       for (int ___i = 0, ___j = !!(cond); !(___j) && !___i; bpf_throw(0), ___i++)

Kumar,

Is this approach reliable?
I suspect the compiler can still optimize it.
I feel it will be annoying to clean up if folks start using it now,
since there won't be a drop in replacement.
Every such bpf_assert_if() would need to be manually patched.

If 2nd part of exception is far, how about we add an equivalent
of __bpf_assert() macroses with conditional ops in asm,
but with extra 'asm volatile goto' that can be used to construct
release of resources.

bpf_do_assert_eq(var1, 0) { bpf_spin_unlock(...); }
bpf_do_assert_lt(var2, 0) { bpf_spin_unlock(...); }

On Thu, Dec 14, 2023 at 6:46 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Thu, Dec 14, 2023 at 2:56 PM Daniel Xu <dxu@dxuuu.xyz> wrote:
> >
> > These macros are a temporary stop-gap until bpf exceptions support
> > unwinding acquired entities. Basically these macros act as if they take
> > a callback which only get executed if the assertion fails.
> >
> > Signed-off-by: Daniel Xu <dxu@dxuuu.xyz>
> > ---
> >  .../testing/selftests/bpf/bpf_experimental.h  | 22 +++++++++++++++++++
> >  1 file changed, 22 insertions(+)
> >
> > diff --git a/tools/testing/selftests/bpf/bpf_experimental.h b/tools/testing/selftests/bpf/bpf_experimental.h
> > index 1386baf9ae4a..d63f415bef26 100644
> > --- a/tools/testing/selftests/bpf/bpf_experimental.h
> > +++ b/tools/testing/selftests/bpf/bpf_experimental.h
> > @@ -263,6 +263,17 @@ extern void bpf_throw(u64 cookie) __ksym;
> >   */
> >  #define bpf_assert(cond) if (!(cond)) bpf_throw(0);
> >
> > +/* Description
> > + *     Assert that a conditional expression is true. If false, runs code in the
> > + *     body before throwing.
> > + * Returns
> > + *     Void.
> > + * Throws
> > + *     An exception with the value zero when the assertion fails.
> > + */
> > +#define bpf_assert_if(cond) \
> > +       for (int ___i = 0, ___j = !!(cond); !(___j) && !___i; bpf_throw(0), ___i++)
>
> Kumar,
>
> Is this approach reliable?
> I suspect the compiler can still optimize it.
> I feel it will be annoying to clean up if folks start using it now,
> since there won't be a drop in replacement.
> Every such bpf_assert_if() would need to be manually patched.
>
> If 2nd part of exception is far, how about we add an equivalent
> of __bpf_assert() macroses with conditional ops in asm,
> but with extra 'asm volatile goto' that can be used to construct
> release of resources.
>
> bpf_do_assert_eq(var1, 0) { bpf_spin_unlock(...); }
> bpf_do_assert_lt(var2, 0) { bpf_spin_unlock(...); }

Just realized that we can go the other way instead.

We can get rid of bpf_assert_eq/ne/... and replace with:

diff --git a/tools/testing/selftests/bpf/bpf_experimental.h
b/tools/testing/selftests/bpf/bpf_experimental.h
index 1386baf9ae4a..1c500287766d 100644
--- a/tools/testing/selftests/bpf/bpf_experimental.h
+++ b/tools/testing/selftests/bpf/bpf_experimental.h
@@ -254,6 +254,15 @@ extern void bpf_throw(u64 cookie) __ksym;
                }
                 \
         })

+#define _EQ(LHS, RHS) \
+       ({ int var = 1;\
+               asm volatile goto("if %[lhs] == %[rhs] goto %l[l_yes]" \
+               :: [lhs] "r"(LHS), [rhs] "i"(RHS) :: l_yes);\
+       var = 0;\
+l_yes:\
+       var;\
+       })
+
 /* Description
  *     Assert that a conditional expression is true.
  * Returns
diff --git a/tools/testing/selftests/bpf/progs/exceptions.c
b/tools/testing/selftests/bpf/progs/exceptions.c
index 2811ee842b01..1111e852f154 100644
--- a/tools/testing/selftests/bpf/progs/exceptions.c
+++ b/tools/testing/selftests/bpf/progs/exceptions.c
@@ -203,6 +203,7 @@ __noinline int assert_nz_gfunc(u64 c)
        volatile u64 cookie = c;

        bpf_assert(cookie != 0);
+       bpf_assert(_EQ(cookie, 2));
        return 0;
 }

we can probably remove bpf_assert_with() and
all of the bpf_assert_le|ne|qt|eq|_with()

Users can open code everything:
if (!_EQ(foo, bar)) bpf_throw(123);

bpf_assert_if() can work too,
but let's call it bpf_do_assert() and use like:

bpf_do_assert(EQ(time, 0)) {
   // cleanup
}

On Thu, Dec 14, 2023 at 7:10 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> Just realized that we can go the other way instead.
>
> We can get rid of bpf_assert_eq/ne/... and replace with:
>
> diff --git a/tools/testing/selftests/bpf/bpf_experimental.h
> b/tools/testing/selftests/bpf/bpf_experimental.h
> index 1386baf9ae4a..1c500287766d 100644
> --- a/tools/testing/selftests/bpf/bpf_experimental.h
> +++ b/tools/testing/selftests/bpf/bpf_experimental.h
> @@ -254,6 +254,15 @@ extern void bpf_throw(u64 cookie) __ksym;
>                 }
>                  \
>          })
>
> +#define _EQ(LHS, RHS) \
> +       ({ int var = 1;\
> +               asm volatile goto("if %[lhs] == %[rhs] goto %l[l_yes]" \
> +               :: [lhs] "r"(LHS), [rhs] "i"(RHS) :: l_yes);\
> +       var = 0;\
> +l_yes:\
> +       var;\
> +       })

Realized we can do much better.
We can take advantage that bpf assembly syntax resembles C and do:

bpf_assert(CMP(cookie, "!=", 0);

and use it as generic "volatile compare" that compiler cannot optimize out:

Replacing:
if (foo < bar) ...
with
if (CMP(foo, "<", bar)) ...
when the compare operator should be preserved.
I'll try to prototype it soon.

Might go further and use C++ for bpf programs :)
Override operator<, opreator==, ...
then if (foo < bar) will be in asm code as written in C++ bpf prog.