On 2017/8/11 2:13, Arnaldo Carvalho de Melo wrote:
Thomas, please try to find who wrote that test and CC him, I'm doing it
this time, Wang, can you please take a look at this?

I also added lkml to the CC list, here we have more users of perf, lkml
is the more developer centric perf list, as perf touches way too many
places in the kernel.

Best regards,

- Arnaldo

Em Wed, Aug 09, 2017 at 11:24:19AM +0200, Thomas Richter escreveu:
I investigate perf test BPF for s390x and have a question regarding
the 38.3 subtest (bpf-prologue test) which fails on s390x.

When I turn on trace_printk in tests/bpf-script-test-prologue.c
I see this output in /sys/kernel/debug/tracing/trace:

[root@s8360047 perf]# cat /sys/kernel/debug/tracing/trace
perf-30229 [000] d..2 170161.535791: : f_mode 2001d00000000 offset:0 orig:0
perf-30229 [000] d..2 170161.535809: : f_mode 6001f00000000 offset:0 orig:0
perf-30229 [000] d..2 170161.535815: : f_mode 6001f00000000 offset:1 orig:0
perf-30229 [000] d..2 170161.535819: : f_mode 2001d00000000 offset:1 orig:0
perf-30229 [000] d..2 170161.535822: : f_mode 2001d00000000 offset:2 orig:1
perf-30229 [000] d..2 170161.535825: : f_mode 6001f00000000 offset:2 orig:1
perf-30229 [000] d..2 170161.535828: : f_mode 6001f00000000 offset:3 orig:1
perf-30229 [000] d..2 170161.535832: : f_mode 2001d00000000 offset:3 orig:1
perf-30229 [000] d..2 170161.535835: : f_mode 2001d00000000 offset:4 orig:0
perf-30229 [000] d..2 170161.535841: : f_mode 6001f00000000 offset:4 orig:0


There are 3 parameters the eBPF program tests/bpf-script-test-prologue.c
accesses: f_mode (member of struct file at offset 140) offset and orig.
They are parameters of the lseek() system call triggered in this
test case in function llseek_loop().

What is really strange is the value of f_mode. It is an 8 byte
value, whereas in the probe event it is defined as a 4 byte value.
The lower 4 bytes are all zero and do not belong to member f_mode.
The correct value should be 2001d for read-only and 6001f for
read-write open mode.

Here is the output of the 'perf test -vv bpf' trace:
    Try to find probe point from debuginfo.
    Matched function: null_lseek [2d9310d]
    Probe point found: null_lseek+0
    Searching 'file' variable in context.
    Converting variable file into trace event.
    converting f_mode in file
    f_mode type is unsigned int.
    Opening /sys/kernel/debug/tracing//README write=0
    Searching 'offset' variable in context.
    Converting variable offset into trace event.
    offset type is long long int.
    Searching 'orig' variable in context.
    Converting variable orig into trace event.
    orig type is int.
    Found 1 probe_trace_events.
    Opening /sys/kernel/debug/tracing//kprobe_events write=1
    Writing event: p:perf_bpf_probe/func _text+8794224 f_mode=+140(%r2):x32
        offset=%r3:s64 orig=%r4:s32

Thank you for your information. This is an endianess problem.

Here f_mode is x32, so perf probe and debuginfo in vmlinux is correct.
However, BPF prologue generator doesn't obey type, but unconditionally
fetch 8 bytes (on 64-bit machine) and pass it to parameter of
bpf_func__null_lseek. This is reasonable, because all args should be
unsigned long. However, to recover its original value, we need a casting.

Please help me verify if the following fix works on your platform:

diff --git a/tools/perf/tests/bpf-script-test-prologue.c b/tools/perf/tests/bpf-script-test-prologue.c
index b4ebc75..43f1e16 100644
--- a/tools/perf/tests/bpf-script-test-prologue.c
+++ b/tools/perf/tests/bpf-script-test-prologue.c
@@ -26,9 +26,11 @@ static void (*bpf_trace_printk)(const char *fmt, int fmt_size, ...) =
        (void *) 6;

 SEC("func=null_lseek file->f_mode offset orig")
-int bpf_func__null_lseek(void *ctx, int err, unsigned long f_mode,
+int bpf_func__null_lseek(void *ctx, int err, unsigned long _f_mode,
                         unsigned long offset, unsigned long orig)
+       fmode_t f_mode = (fmode_t)_f_mode;
        if (err)
                return 0;
        if (f_mode & FMODE_WRITE)

Thank you.

Reply via email to