bpf_trace_printk() uses conditional operators to attempt to pass
different types to __trace_printk() depending on the format operators.
This doesn't work as intended on 32-bit architectures where u32 & long
are passed differently to u64, since the result of C conditional
operators follows the "usual arithmetic conversions" rules, such that
the values passed to __trace_printk() will always be u64.
For example the samples/bpf/tracex5 test printed lines like below on
MIPS, where the fd and buf have come from the u64 fd argument, and the
size from the buf argument:
dd-1176 [000] .... 1180.941542: 0x00000001: write(fd=1, buf= (null),
size=6258688)
Instead of this:
dd-1217 [000] .... 1625.616026: 0x00000001: write(fd=1, buf=009e4000,
size=512)
Work around this with an ugly hack which expands each combination of
argument types for the 3 arguments. On 64-bit kernels it is assumed that
u32, long & u64 are all passed the same way so no casting takes place
(it has apparently worked implicitly until now). On 32-bit kernels it is
assumed that long and u32 pass the same way so there are 8 combinations.
On 32-bit kernels bpf_trace_printk() increases in size but should now
work correctly. On 64-bit kernels it actually reduces in size slightly,
I presume due to removal of some of the casts (which as far as I can
tell are unnecessary for printk anyway due to the controlled nature of
the interpretation):
arch function old new delta
x86_64 bpf_trace_printk 532 412 -120
x86 bpf_trace_printk 676 1120 +444
MIPS64 bpf_trace_printk 760 612 -148
MIPS32 bpf_trace_printk 768 996 +228
Fixes: 9c959c863f82 ("tracing: Allow BPF programs to call bpf_trace_printk()")
Signed-off-by: James Hogan <james.ho...@imgtec.com>
Cc: Alexei Starovoitov <a...@kernel.org>
Cc: Daniel Borkmann <dan...@iogearbox.net>
Cc: Steven Rostedt <rost...@goodmis.org>
Cc: Ingo Molnar <mi...@redhat.com>
Cc: netdev@vger.kernel.org
---
I'm open to nicer ways of fixing this.
This is tested with samples/bpf/tracex5 on MIPS32 and MIPS64. Only build
tested on x86.
---
kernel/trace/bpf_trace.c | 26 ++++++++++++++++++++++----
1 file changed, 22 insertions(+), 4 deletions(-)
diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index 37385193a608..32dcbe1b48f2 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -204,10 +204,28 @@ BPF_CALL_5(bpf_trace_printk, char *, fmt, u32, fmt_size,
u64, arg1,
fmt_cnt++;
}
- return __trace_printk(1/* fake ip will not be printed */, fmt,
- mod[0] == 2 ? arg1 : mod[0] == 1 ? (long) arg1 :
(u32) arg1,
- mod[1] == 2 ? arg2 : mod[1] == 1 ? (long) arg2 :
(u32) arg2,
- mod[2] == 2 ? arg3 : mod[2] == 1 ? (long) arg3 :
(u32) arg3);
+ /*
+ * This is a horribly ugly hack to allow different combinations of
+ * argument types to be used, particularly on 32-bit architectures where
+ * u32 & long pass the same as one another, but differently to u64.
+ *
+ * On 64-bit architectures it is assumed u32, long & u64 pass in the
+ * same way.
+ */
+
+#define __BPFTP_P(...) __trace_printk(1/* fake ip will not be printed */, \
+ fmt, ##__VA_ARGS__)
+#define __BPFTP_1(...) ((mod[0] == 2 || __BITS_PER_LONG == 64) \
+ ? __BPFTP_P(arg1, ##__VA_ARGS__) \
+ : __BPFTP_P((long)arg1, ##__VA_ARGS__))
+#define __BPFTP_2(...) ((mod[1] == 2 || __BITS_PER_LONG == 64) \
+ ? __BPFTP_1(arg2, ##__VA_ARGS__) \
+ : __BPFTP_1((long)arg2, ##__VA_ARGS__))
+#define __BPFTP_3(...) ((mod[2] == 2 || __BITS_PER_LONG == 64) \
+ ? __BPFTP_2(arg3, ##__VA_ARGS__) \
+ : __BPFTP_2((long)arg3, ##__VA_ARGS__))
+
+ return __BPFTP_3();
}
static const struct bpf_func_proto bpf_trace_printk_proto = {
--
2.13.2
