On 2022/11/28 9:57, Alexei Starovoitov wrote:
On Sat, Nov 26, 2022 at 05:45:27PM +0800, Yang Jihong wrote:
For ARM32 architecture, if data width of kfunc return value is 32 bits,
need to do explicit zero extension for high 32-bit, insn_def_regno should
return dst_reg for BPF_JMP type of BPF_PSEUDO_KFUNC_CALL. Otherwise,
opt_subreg_zext_lo32_rnd_hi32 returns -EFAULT, resulting in BPF failure.
Signed-off-by: Yang Jihong <yangjihong1@xxxxxxxxxx>
---
kernel/bpf/verifier.c | 44 ++++++++++++++++++++++++++++++++++++++++---
1 file changed, 41 insertions(+), 3 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 264b3dc714cc..193ea927aa69 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -1927,6 +1927,21 @@ find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
}
+static int kfunc_desc_cmp_by_imm(const void *a, const void *b);
+
+static const struct bpf_kfunc_desc *
+find_kfunc_desc_by_imm(const struct bpf_prog *prog, s32 imm)
+{
+ struct bpf_kfunc_desc desc = {
+ .imm = imm,
+ };
+ struct bpf_kfunc_desc_tab *tab;
+
+ tab = prog->aux->kfunc_tab;
+ return bsearch(&desc, tab->descs, tab->nr_descs,
+ sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
+}
+
static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
s16 offset)
{
@@ -2342,6 +2357,13 @@ static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
*/
if (insn->src_reg == BPF_PSEUDO_CALL)
return false;
+
+ /* Kfunc call will reach here because of insn_has_def32,
+ * conservatively return TRUE.
+ */
+ if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
+ return true;
+
/* Helper call will reach here because of arg type
* check, conservatively return TRUE.
*/
@@ -2405,10 +2427,26 @@ static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
}
/* Return the regno defined by the insn, or -1. */
-static int insn_def_regno(const struct bpf_insn *insn)
+static int insn_def_regno(struct bpf_verifier_env *env, const struct bpf_insn *insn)
{
switch (BPF_CLASS(insn->code)) {
case BPF_JMP:
+ if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
+ const struct bpf_kfunc_desc *desc;
+
+ /* The value of desc cannot be NULL */
+ desc = find_kfunc_desc_by_imm(env->prog, insn->imm);
+
+ /* A kfunc can return void.
+ * The btf type of the kfunc's return value needs
+ * to be checked against "void" first
+ */
+ if (desc->func_model.ret_size == 0)
+ return -1;
+ else
+ return insn->dst_reg;
+ }
+ fallthrough;
I cannot make any sense of this patch.
insn->dst_reg above is 0.
The kfunc call doesn't define a register from insn_def_regno() pov.
Are you hacking insn_def_regno() to return 0 so that
if (WARN_ON(load_reg == -1)) {
verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
return -EFAULT;
}
in opt_subreg_zext_lo32_rnd_hi32() doesn't trigger ?
But this verifier message should have been a hint that you need
to analyze why zext_dst is set on this kfunc call.
Maybe it shouldn't ?
Did you analyze the logic of mark_btf_func_reg_size() ?
make r0 zext is not caused by mark_btf_func_reg_size.
This problem occurs when running the kfunc_call_test_ref_btf_id test
case in the 32-bit ARM environment.
The bpf prog is as follows:
int kfunc_call_test_ref_btf_id(struct __sk_buff *skb)
{
struct prog_test_ref_kfunc *pt;
unsigned long s = 0;
int ret = 0;
pt = bpf_kfunc_call_test_acquire(&s);
if (pt) {
// here, do_check clears the upper 32bits of r0 through:
// check_alu_op
// ->check_reg_arg
// ->mark_insn_zext
if (pt->a != 42 || pt->b != 108)
ret = -1;
bpf_kfunc_call_test_release(pt);
}
return ret;
}
Before producing any patches please understand the logic fully.
Your commit log
"insn_def_regno should
return dst_reg for BPF_JMP type of BPF_PSEUDO_KFUNC_CALL."
Makes no sense to me, since dst_reg is unused in JMP insn.
There is no concept of a src or dst register in a JMP insn.
32-bit x86 supports calling kfuncs. See emit_kfunc_call().
And we don't have this "verifier bug. zext_dst is set" issue there, right?
But what you're saying in the commit log:
"if data width of kfunc return value is 32 bits"
should have been applicable to x86-32 as well.
So please start with a test that demonstrates the issue on x86-32 and
then we can discuss the way to fix it.
The patch 2 sort-of makes sense.
For patch 3 pls add new test funcs to bpf_testmod.
We will move all of them from net/bpf/test_run.c to bpf_testmod eventually.
.
