Re: [PATCH bpf-next 1/3] bpf: support nocsr patterns for calls to kfuncs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 8/15/24 3:16 PM, Yonghong Song wrote:

On 8/15/24 2:24 PM, Andrii Nakryiko wrote:
On Mon, Aug 12, 2024 at 4:44 PM Eduard Zingerman <eddyz87@xxxxxxxxx> wrote:
Recognize nocsr patterns around kfunc calls.
For example, suppose bpf_cast_to_kern_ctx() follows nocsr contract
(which it does, it is rewritten by verifier as "r0 = r1" insn),
in such a case, rewrite BPF program below:

   r2 = 1;
   *(u64 *)(r10 - 32) = r2;
   call %[bpf_cast_to_kern_ctx];
   r2 = *(u64 *)(r10 - 32);
   r0 = r2;

Removing the spill/fill pair:

   r2 = 1;
   call %[bpf_cast_to_kern_ctx];
   r0 = r2;

Add a KF_NOCSR flag to mark kfuncs that follow nocsr contract.

Signed-off-by: Eduard Zingerman <eddyz87@xxxxxxxxx>
---
  include/linux/btf.h   |  1 +
  kernel/bpf/verifier.c | 36 ++++++++++++++++++++++++++++++++++++
  2 files changed, 37 insertions(+)

diff --git a/include/linux/btf.h b/include/linux/btf.h
index cffb43133c68..59ca37300423 100644
--- a/include/linux/btf.h
+++ b/include/linux/btf.h
@@ -75,6 +75,7 @@
  #define KF_ITER_NEXT    (1 << 9) /* kfunc implements BPF iter next method */   #define KF_ITER_DESTROY (1 << 10) /* kfunc implements BPF iter destructor */   #define KF_RCU_PROTECTED (1 << 11) /* kfunc should be protected by rcu cs when they are invoked */ +#define KF_NOCSR        (1 << 12) /* kfunc follows nocsr calling contract */

  /*
   * Tag marking a kernel function as a kfunc. This is meant to minimize the
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index df3be12096cf..c579f74be3f9 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -16140,6 +16140,28 @@ static bool verifier_inlines_helper_call(struct bpf_verifier_env *env, s32 imm)
         }
  }

+/* Same as helper_nocsr_clobber_mask() but for kfuncs, see comment above */ +static u32 kfunc_nocsr_clobber_mask(struct bpf_kfunc_call_arg_meta *meta)
+{
+       const struct btf_param *params;
+       u32 vlen, i, mask;
+
+       params = btf_params(meta->func_proto);
+       vlen = btf_type_vlen(meta->func_proto);
+       mask = 0;
+       if (!btf_type_is_void(btf_type_by_id(meta->btf, meta->func_proto->type)))
+               mask |= BIT(BPF_REG_0);
+       for (i = 0; i < vlen; ++i)
+               mask |= BIT(BPF_REG_1 + i);
Somewhere deep in btf_dump implementation of libbpf, there is a
special handling of `<whatever> func(void)` (no args) function as
having vlen == 1 and type being VOID (i.e., zero). I don't know if
that still can happen, but I believe at some point we could get this
vlen==1 and type=VOID for no-args functions. So I wonder if we should
handle that here as well, or is it some compiler atavism we can forget
about?

The case to have vlen=1 and type=VOID only happens for
bpf programs with llvm19 and later.
For example,

$ cat t.c
int foo(); // a kfunc or a helper
int bar() {
  return foo(1, 2);
}

$ clang --target=bpf -O2 -g -c t.c && llvm-dwarfdump t.o
t.c:3:13: warning: passing arguments to 'foo' without a prototype is deprecated in all versions of C and is not supported in C23 [-Wdeprecated-non-prototype]
    3 |   return foo(1, 2);
      |             ^
1 warning generated.
t.o:    file format elf64-bpf
...
0x00000039:   DW_TAG_subprogram
                DW_AT_name      ("foo")
                DW_AT_decl_file ("/home/yhs/t.c")
                DW_AT_decl_line (1)
                DW_AT_type      (0x00000043 "int")
                DW_AT_declaration       (true)
                DW_AT_external  (true)

0x00000041:     DW_TAG_unspecified_parameters

0x00000042:     NULL
...

If we do see a BPF kfunc/helper with vlen=1 and type is VOID,
that means the number of arguments is actual UNKNOWN
based on dwarf DW_TAG_subprogram tag. Although it is unlikely
people to write code like above, it might be still useful
to add check with vlen=1 and type=VOID and reject such a case.


For vmlinux BTF, this is not possible since eventually all function
has a definition which will define the function precisely w.r.t.
the number of arguments and their types.





+       return mask;
+}
+
+/* Same as verifier_inlines_helper_call() but for kfuncs, see comment above */ +static bool verifier_inlines_kfunc_call(struct bpf_kfunc_call_arg_meta *meta)
+{
+       return false;
+}
+
  /* GCC and LLVM define a no_caller_saved_registers function attribute.
   * This attribute means that function scratches only some of
   * the caller saved registers defined by ABI.
@@ -16238,6 +16260,20 @@ static void mark_nocsr_pattern_for_call(struct bpf_verifier_env *env,
bpf_jit_inlines_helper_call(call->imm));
         }

+       if (bpf_pseudo_kfunc_call(call)) {
+               struct bpf_kfunc_call_arg_meta meta;
+               int err;
+
+               err = fetch_kfunc_meta(env, call, &meta, NULL);
+               if (err < 0)
+                       /* error would be reported later */
+                       return;
+
+               clobbered_regs_mask = kfunc_nocsr_clobber_mask(&meta);
+               can_be_inlined = (meta.kfunc_flags & KF_NOCSR) &&
+ verifier_inlines_kfunc_call(&meta);
+       }
+
         if (clobbered_regs_mask == ALL_CALLER_SAVED_REGS)
                 return;

--
2.45.2





[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux