On 11/14/17 12:25 PM, Daniel Borkmann wrote:
On 11/14/2017 07:15 PM, Yonghong Song wrote:
On 11/14/17 6:19 AM, Daniel Borkmann wrote:
On 11/14/2017 02:42 PM, Arnaldo Carvalho de Melo wrote:
Em Tue, Nov 14, 2017 at 02:09:34PM +0100, Daniel Borkmann escreveu:
On 11/14/2017 01:58 PM, Arnaldo Carvalho de Melo wrote:
Em Tue, Nov 14, 2017 at 01:09:39AM +0100, Daniel Borkmann escreveu:
On 11/13/2017 04:08 PM, Arnaldo Carvalho de Melo wrote:
libbpf: -- BEGIN DUMP LOG ---
libbpf:
0: (79) r3 = *(u64 *)(r1 +104)
1: (b7) r2 = 0
2: (bf) r6 = r1
3: (bf) r1 = r10
4: (07) r1 += -128
5: (b7) r2 = 128
6: (85) call bpf_probe_read_str#45
7: (bf) r1 = r0
8: (07) r1 += -1
9: (67) r1 <<= 32
10: (77) r1 >>= 32
11: (25) if r1 > 0x7f goto pc+11
Right, so the compiler is optimizing the two tests into a single one above,
which means lower bound cannot properly be derived again by the verifier due
to this and thus you'll get the error. Similar issue was seen recently [1].
Does the below hack work for you?
int prog([...])
{
char filename[128];
int ret = bpf_probe_read_str(filename, sizeof(filename),
filename_ptr);
if (ret > 0)
bpf_perf_event_output(ctx, &__bpf_stdout__,
BPF_F_CURRENT_CPU, filename,
ret & (sizeof(filename) - 1));
return 1;
}
r0 should keep on tracking bounds here at least:
prog:
0: bf 16 00 00 00 00 00 00 r6 = r1
1: bf a1 00 00 00 00 00 00 r1 = r10
2: 07 01 00 00 80 ff ff ff r1 += -128
3: b7 02 00 00 80 00 00 00 r2 = 128
4: 85 00 00 00 2d 00 00 00 call 45
5: 67 00 00 00 20 00 00 00 r0 <<= 32
6: c7 00 00 00 20 00 00 00 r0 s>>= 32
7: b7 01 00 00 01 00 00 00 r1 = 1
8: 6d 01 0a 00 00 00 00 00 if r1 s> r0 goto 10
9: 57 00 00 00 7f 00 00 00 r0 &= 127
10: bf a4 00 00 00 00 00 00 r4 = r10
11: 07 04 00 00 80 ff ff ff r4 += -128
12: bf 61 00 00 00 00 00 00 r1 = r6
13: 18 02 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r2 = 0ll
15: 18 03 00 00 ff ff ff ff 00 00 00 00 00 00 00 00 r3 =
4294967295ll
17: bf 05 00 00 00 00 00 00 r5 = r0
18: 85 00 00 00 19 00 00 00 call 25
[1]
https://urldefense.proofpoint.com/v2/url?u=http-3A__patchwork.ozlabs.org_project_netdev_list_-3Fseries-3D13211&d=DwIDaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=Qp3xFfXEz-CT8rzYtrHeXbow2M6FlsUzwcY32i3_2Q0&s=z0d6b_hxStA845Kh7epJ-JiFwkiWqUH_z3fEadwqAQY&e=
Not yet:
6: (85) call bpf_probe_read_str#45
7: (bf) r1 = r0
8: (67) r1 <<= 32
9: (77) r1 >>= 32
10: (15) if r1 == 0x0 goto pc+10
R0=inv(id=0) R1=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff))
R6=ctx(id=0,off=0,imm=0) R10=fp0
11: (57) r0 &= 127
12: (bf) r4 = r10
13: (07) r4 += -128
14: (bf) r1 = r6
15: (18) r2 = 0xffff92bfc2aba840u
17: (18) r3 = 0xffffffff
19: (bf) r5 = r0
20: (85) call bpf_perf_event_output#25
invalid stack type R4 off=-128 access_size=0
I'll try updating clang/llvm...
Full details:
[root@jouet bpf]# cat open.c
#include "bpf.h"
SEC("prog=do_sys_open filename")
int prog(void *ctx, int err, const char __user *filename_ptr)
{
char filename[128];
const unsigned len = bpf_probe_read_str(filename, sizeof(filename),
filename_ptr);
Btw, I was using 'int' here above instead of 'unsigned' as strncpy_from_unsafe()
could potentially return errors like -EFAULT.
I changed to int, didn't help
Currently having a version compiled from the git tree:
# llc --version
LLVM
(https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_&d=DwIDaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=Qp3xFfXEz-CT8rzYtrHeXbow2M6FlsUzwcY32i3_2Q0&s=BKC_Gu9s1hw0v13OCgCpfsGtAY2hE7dujFqg8LNaK2I&e=):
LLVM version 6.0.0git-2d810c2
Optimized build.
Default target: x86_64-unknown-linux-gnu
Host CPU: skylake
[root@jouet bpf]# llc --version
LLVM
(https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_&d=DwIDaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=Qp3xFfXEz-CT8rzYtrHeXbow2M6FlsUzwcY32i3_2Q0&s=BKC_Gu9s1hw0v13OCgCpfsGtAY2hE7dujFqg8LNaK2I&e=):
LLVM version 4.0.0svn
Old stuff! ;-) Will change, but improving these messages should be on
the radar, I think :-)
Yep, agree, I think we need a generic, better solution for this type of
issue instead of converting individual helpers to handle 0 min bound and
then only bailing out in such case; need to brainstorm a bit on that.
I think for the above in your case ...
[...]
6: (85) call bpf_probe_read_str#45
7: (bf) r1 = r0
8: (67) r1 <<= 32
9: (77) r1 >>= 32
10: (15) if r1 == 0x0 goto pc+10
R0=inv(id=0) R1=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff))
R6=ctx(id=0,off=0,imm=0) R10=fp0
11: (57) r0 &= 127
[...]
... the shifts on r1 might be due to using 32 bit type, so if you find
a way to avoid these and have the test on r0 directly, we might get there.
Perhaps keep using a 64 bit type to avoid them. It would be useful to
propagate the deduced bound information back to r0 when we know that
neither r0 nor r1 has changed in the meantime.
It is tricky to do in the bpf_program. Compiler tries hard to optimize :-).
The issue is at "r0 &= 127".
9: (6d) if r1 s> r0 goto pc+10
R0=inv(id=0,umin_value=1,umax_value=9223372036854775807,var_off=(0x0;
0x7fffffffffffffff)) R1=inv1 R6=ctx(id=0,off=0,imm=0) R10=fp0
10: R0=inv(id=0,umin_value=1,umax_value=9223372036854775807,var_off=(0x0;
0x7fffffffffffffff)) R1=inv1 R6=ctx(id=0,off=0,imm=0) R10=fp0
10: (57) r0 &= 127
11: R0=inv(id=0,umax_value=127,var_off=(0x0; 0x7f)) R1=inv1
R6=ctx(id=0,off=0,imm=0) R10=fp0
One possible solution for this problem is to relax the arg4 type
from ARG_CONST_SIZE to ARG_CONST_SIZE_OR_ZERO.
Yeah, I know, that's what I mentioned earlier in this thread to resolve it,
but do we really want to add this hack everywhere? :( Potentially any function
having ARG_CONST_SIZE would need to handle size 0 and bail out again in their
helper implementation and it ends up that progs start relying on this runtime
check where we won't be able to get rid of it later on anymore.
The compiler actually does the right thing for the below code:
int ret = bpf_probe_read_str(filename, sizeof(filename),
filename_ptr);
if (ret > 0)
bpf_perf_event_output(ctx, &__bpf_stdout__,BPF_F_CURRENT_CPU,
filename, ret & (sizeof(filename) - 1));
Just from the above code without consulting bpf_probe_read_str
internals, it is totally possible that ret = 128, then
ret & (sizeof(filename) - 1) = 0.
The issue is that the verifier did not set the "ret" initial range as
(-inf, sizeof(filename) - 1). We could have this information associated
with helper and feed back to verifier.
If we have this range, later for ret & (sizeof(filename) - 1) with ret
>= 1, the verifier should be able to conclude
ret & (sizeof(filename) - 1) >= 1.
To workaround the immediate problem, I tested the following hack
with bcc and it works fine.
BPF_PERF_OUTPUT(events);
int trace(struct pt_regs *ctx) {
char filename[128];
int ret = bpf_probe_read_str(filename, sizeof(filename), 0);
if (ret > 0) {
if (ret == 1)
events.perf_submit(ctx, filename, ret);
else if (ret < 128)
events.perf_submit(ctx, filename, ret);
}
return 1;
}
The idea is to make control flow more complex to prevent llvm
do certain optimizations.
diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index a5580c6..a68d8bd 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -393,6 +393,9 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs,
struct bpf_map *, map,
},
};
+ if (unlikely(size == 0))
+ return 0;
+
if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
return -EINVAL;
@@ -407,7 +410,7 @@ static const struct bpf_func_proto
bpf_perf_event_output_proto = {
.arg2_type = ARG_CONST_MAP_PTR,
.arg3_type = ARG_ANYTHING,
.arg4_type = ARG_PTR_TO_MEM,
- .arg5_type = ARG_CONST_SIZE,
+ .arg5_type = ARG_CONST_SIZE_OR_ZERO,
};
