Re: [PATCH perf 3/4] perf tools: Fix unused variables: x86_{32,64}_regoffset_table

From: Wangnan (F)
Date: Fri Jan 22 2016 - 01:23:08 EST




On 2016/1/22 13:56, åæéå / HIRAMATUïMASAMI wrote:
From: Wangnan (F) [mailto:wangnan0@xxxxxxxxxx]

On 2016/1/20 21:59, Arnaldo Carvalho de Melo wrote:
Em Tue, Jan 19, 2016 at 09:33:06PM +0000, Ben Hutchings escreveu:
gcc 5 doesn't seem to care about these, but gcc 6 does and that
results in a build failure.
Ben, please CC the people on the CC list for the patch that introduces
the problem, Wang, He, can I have your Acked-by?

- Arnaldo

This patch lead me find a bug in original code.

If both perf and target ELF binary is x86_64, following command works okay:

# perf probe -v -n --exec /tmp/oxygen_root/lib64/libc.so.6 pselect
data exceptfds readfds writefds nfds sigmask tval timeout
<SNIP>
Opening /sys/kernel/debug/tracing//uprobe_events write=1
Writing event: p:probe_libc/pselect
/home/w00229757/oxygen_root-w00229757/lib64/libc-2.18.so:0xdfef0
data=-216(%sp):u64 exceptfds=%cx:u64 readfds=%si:u64 writefds=%dx:u64
nfds=%di:s32 sigmask=%r9:u64 tval=-232(%sp):u64 timeout=%r8:u64
<SNIP>

But if the library is x86_32, result is incorrect:

# perf probe -v -n --exec /tmp/oxygen_root/lib32/libc.so.6 pselect
data exceptfds readfds writefds nfds sigmask tval
<SNIP>
Writing event: p:probe_libc/pselect
/tmp/oxygen_root-w00229757/lib32/libc-2.18.so:0xd1330 data=-172(%si):u64
exceptfds=+16(%si):u32 readfds=+8(%si):u32 writefds=+12(%si):u32
nfds=+4(%si):s32 sigmask=+24(%si):u32 tval=-180(%si):u64
timeout=+20(%si):u32
<SNIP>

We know that (%si) is used to passing arguments. Here we should see
'%sp' or '$stack'.

Use a x86_32 perf we get currect result:

# ~/perf probe -v -n --exec /tmp/oxygen_root/lib32/libc.so.6 pselect
data exceptfds readfds writefds nfds sigmask tval
<SNIP>
Writing event: p:probe_libc/pselect
/tmp/oxygen_root-w00229757/lib32/libc-2.18.so:0xd1330
data=-172($stack):u64 exceptfds=+16($stack):u32 readfds=+8($stack):u32
writefds=+12($stack):u32 nfds=+4($stack):s32 sigmask=+24($stack):u32
tval=-180($stack):u64
<SNIP>
Ah, I see. Uprobes may not check the target binary is in 32bit mode.
Since the stack of x86-64 and x86-32 on pt_regs are different,
(regs->sp points stack on x86-64, &(regs->pt) points stack on x86-32)
uprobes would better checking and change the behavior.

But anyway, it is also fixed by changing perf's register table.


Use a small test program to check the result:

#include <sys/time.h>
#include <sys/types.h>
#include <unistd.h>
#include <memory.h>

static struct {
fd_set r, w, e;
struct timespec ts;
sigset_t m;
} s;

int main()
{
memset(&s, '\0', sizeof(s));

pselect(0, &s.r, &s.w, &s.e, &s.ts, &s.m);
return 0;
}

# gcc -m32 -g ./test_pselect.c

Use x86_32 perf:

# ./perf probe -v --exec /tmp/oxygen_root/lib32/libc.so.6 pselect data
exceptfds readfds writefds nfds sigmask tval
Writing event: p:probe_libc/pselect
/tmp/oxygen_root-w00229757/lib32/libc-2.18.so:0xd1330
data=-172($stack):u64 exceptfds=+16($stack):u32 readfds=+8($stack):u32
writefds=+12($stack):u32 nfds=+4($stack):s32 sigmask=+24($stack):u32
tval=-180($stack):u64
Added new event:
probe_libc:pselect (on pselect in
/tmp/oxygen_root-w00229757/lib32/libc-2.18.so with data exceptfds
readfds writefds nfds sigmask tval)

You can now use it in all perf tools, such as:

perf record -e probe_libc:pselect -aR sleep 1

# ./perf record -e probe_libc:pselect ./a.out
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.011 MB perf.data (1 samples) ]
# ./perf script
a.out 25336 [006] 64588.457597: probe_libc:pselect:
(f7663330) data=0xf772e00000000000 exceptfds=0x8049880 readfds=0x8049780
writefds=0x8049800 nfds=0 sigmask=0x8049908 tval=0x0

Switch to x86_64 perf:

# ./perf probe -v --exec /tmp/oxygen_root/lib32/libc.so.6 pselect
data exceptfds readfds writefds nfds sigmask tval
<SNIP>
Opening /sys/kernel/debug/tracing//uprobe_events write=1
Writing event: p:probe_libc/pselect
/tmp/oxygen_root-w00229757/lib32/libc-2.18.so:0xd1330 data=-172(%si):u64
exceptfds=+16(%si):u32 readfds=+8(%si):u32 writefds=+12(%si):u32
nfds=+4(%si):s32 sigmask=+24(%si):u32 tval=-180(%si):u64
Added new event:
probe_libc:pselect (on pselect in
/tmp/oxygen_root-w00229757/lib32/libc-2.18.so with data exceptfds
readfds writefds nfds sigmask tval)

You can now use it in all perf tools, such as:

perf record -e probe_libc:pselect -aR sleep 1

# ./perf record -e probe_libc:pselect ./a.out
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.011 MB perf.data (1 samples) ]
# ./perf script
a.out 25599 [002] 64759.743554: probe_libc:pselect:
(f76e7330) data=0x0 exceptfds=0x0 readfds=0x0 writefds=0x0 nfds=0
sigmask=0x0 tval=0x0

Sad...

I think this problem is not introduced by my patch. In fact
there's a fundamental problem in get_arch_regstr() that it is
impossible to switch sub ISA.
Right, but I guess this can fixed by switching %sp (for x86-64)
and +0(%sp) (for x86-32) instead of $stack.

It may not work.

No matter how we change regoffset_table, when get_arch_regstr()
get a register number 4, how can it know whether we are looking
for x86_32 register and return $stack or return %si for x86_64?

The fundamental problem is: we need a way to map dwarf's register
number to register names so uprobe know which register we are
looking for, but the API we designed for this can't distinguish
the sub ISAs dwarf are using.

Thanks.