On Fri, Apr 25, 2025 at 03:40:55PM +0200, Thomas Weißschuh wrote:
<snip>
Some more information:
The crash comes from the address arithmetic in "vc = &vc[CS_RAW]" going wrong.
That appears to be because it's not doing any arithmetic, but using value
from some linker-generated symbol (I'll refer to it as "7a8").
The below is presumably the check that compares clk != CLOCK_MONOTONIC_RAW,
CLOCK_MONOTONIC_RAW is 4. And it should choose between first and second
vdso_clock
in 2nd vvar page:
# readelf -a /root/kernel-ark/arch/arm64/kernel/vdso/vdso.so.dbg | grep 7a8
25: 00000000000007a8 0 NOTYPE LOCAL DEFAULT 7 $d
objdump -d -S:
658: 17ffffef b 614
<__cvdso_clock_gettime_data.constprop.0+0x104>
vc = &vc[CS_HRES_COARSE];
65c: 58000a62 ldr x2, 7a8
<__cvdso_clock_gettime_data.constprop.0+0x298>
660: 7100101f cmp w0, #0x4
664: 10e7cce3 adr x3, fffffffffffd0000 <vdso_u_data+0x10000>
668: 9a821063 csel x3, x3, x2, ne // ne = any
while (unlikely((seq = READ_ONCE(vc->seq)) & 1))
66c: b9400065 ldr w5, [x3]
...
7a8: fffd00e0 .word 0xfffd00e0
7ac: ffffffff .word 0xffffffff
(gdb) r
Thread 2.1 "a.out" received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0xfffff7ff2dc0 (LWP 44638)]
0x0000fffff7fa066c in ?? ()
(gdb) disassemble 0x0000fffff7fa0658,0x0000fffff7fa067c
--Type <RET> for more, q to quit, c to continue without paging--
0x0000fffff7fa0658: b 0xfffff7fa0614
0x0000fffff7fa065c: ldr x2, 0xfffff7fa07a8
0x0000fffff7fa0660: cmp w0, #0x4
0x0000fffff7fa0664: adr x3, 0xfffff7f70000
0x0000fffff7fa0668: csel x3, x3, x2, ne // ne = any
=> 0x0000fffff7fa066c: ldr w5, [x3]
$x3 here is using correct vvar runtime address, but $x2 uses
a bogus one from "7a8". And when it tries to load from it, it crashes:
(gdb) info proc map
process 44638
Mapped address spaces:
Start Addr End Addr Size Offset
Perms File
0x0000000000400000 0x0000000000410000 0x10000 0x0
r-xp /root/ltp/testcases/kernel/syscalls/clock_gettime/a.out
0x0000000000410000 0x0000000000420000 0x10000 0x0
r--p /root/ltp/testcases/kernel/syscalls/clock_gettime/a.out
0x0000000000420000 0x0000000000430000 0x10000 0x10000
rw-p /root/ltp/testcases/kernel/syscalls/clock_gettime/a.out
0x0000fffff7d80000 0x0000fffff7f30000 0x1b0000 0x0
r-xp /usr/lib64/libc.so.6
0x0000fffff7f30000 0x0000fffff7f40000 0x10000 0x1a0000
r--p /usr/lib64/libc.so.6
0x0000fffff7f40000 0x0000fffff7f50000 0x10000 0x1b0000
rw-p /usr/lib64/libc.so.6
0x0000fffff7f60000 0x0000fffff7fa0000 0x40000 0x0
r--p [vvar]
0x0000fffff7fa0000 0x0000fffff7fb0000 0x10000 0x0
r-xp [vdso]
0x0000fffff7fb0000 0x0000fffff7fe0000 0x30000 0x0
r-xp /usr/lib/ld-linux-aarch64.so.1
0x0000fffff7fe0000 0x0000fffff7ff0000 0x10000 0x20000
r--p /usr/lib/ld-linux-aarch64.so.1
0x0000fffff7ff0000 0x0000fffff8000000 0x10000 0x30000
rw-p /usr/lib/ld-linux-aarch64.so.1
0x0000fffffffd0000 0x0001000000000000 0x30000 0x0
rw-p [stack]
$x3 was the beginning of 2nd vvar page (1st vdso_clock), and $x2 looks
like it should have been 2nd vdso_clock, but the value that is used is not
the address of vvar during runtime.
vdso_clock has size 224 == 0xe0 (according to pahole)
(gdb) x/2x 0xfffff7fa07a8
0xfffff7fa07a8: 0xfffd00e0 0xffffffff
(gdb) p/x $x2
$1 = 0xfffffffffffd00e0
(gdb) x/1x $x2
0xfffffffffffd00e0: Cannot access memory at address 0xfffffffffffd00e0
But it does match the symbol value from vdso.so:
# readelf -a /root/kernel-ark/arch/arm64/kernel/vdso/vdso.so.dbg | grep vdso_u
37: fffffffffffc0000 0 NOTYPE LOCAL DEFAULT 1 vdso_u_data
39: fffffffffffc0000 0 NOTYPE LOCAL DEFAULT ABS vdso_u_time_data
40: fffffffffffe0000 0 NOTYPE LOCAL DEFAULT ABS vdso_u_rng_data
print(hex(0xfffffffffffc0000+65536+224))
0xfffffffffffd00e0 -> and this is address where it crashes
Regards,
Jan
This should just do "vc = vc + 1", advancing the pointer by sizeof(*vc).
But I get these example values of "vc" before and after the arithmetic:
0x00ffffbc060000 -> 0xfffffffffffd00e8
Which is obviously wrong.
The arithmetic can be fixed with any one of the following changes:
* OPTIMIZER_HIDE_VAR() as above
* Replacement of -mcmodel=tiny with -mcmode=small in
arch/arm64/kernel/vdso/Makefile
-mcmodel=tiny is supposed to cover 1MiB programs, the vDSO only needs 300KiB
here
* Removal of the (clk != CLOCK_MONOTONIC_RAW) case
Shuffling code around, even without any impact on semantics, sometimes fixes
the issue. Compiling the vDSO with UBSAN didn't show anything.
Thomas
> > > > [1]
https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/syscalls/clock_gettime/clock_gettime03.c
> > > > [2] https://koji.fedoraproject.org/koji/buildinfo?buildID=2704401
> > >
>
