From: Xu Kuohai xukuohai@huawei.com
dst_reg is marked as unknown when BPF_MOD instruction is verified, causing the following bpf prog to be incorrectly rejected.
0: r0 = 0 1: r0 %= 1 // r0 is marked as unknown 2: r1 = 0 3: r1 += 1 4: if r1 < r0 goto pc-2 // verifier concludes the loop is unbounded 5: exit
To teach verifier to accept the above prog, this series adds bound tracking for BPF_MOD.
v2: - fix build warning reported by kernel test robot lkp@intel.com [0] - add two more cases and update commit message
[0] https://lore.kernel.org/oe-kbuild-all/202303060036.zK05OC5M-lkp@intel.com
v1: https://lore.kernel.org/bpf/20230306033119.2634976-1-xukuohai@huaweicloud.co...
Xu Kuohai (2): bpf: add bound tracking for BPF_MOD selftests/bpf: check if verifier tracks dst_reg bound for BPF_MOD
kernel/bpf/verifier.c | 98 ++++++- tools/testing/selftests/bpf/verifier/mod.c | 320 +++++++++++++++++++++ 2 files changed, 413 insertions(+), 5 deletions(-) create mode 100644 tools/testing/selftests/bpf/verifier/mod.c
From: Xu Kuohai xukuohai@huawei.com
dst_reg is marked as unknown when BPF_MOD instruction is verified, causing the following bpf prog to be incorrectly rejected.
0: r0 = 0 1: r0 %= 1 // r0 is marked as unknown 2: r1 = 0 3: r1 += 1 4: if r1 < r0 goto pc-2 // verifier treats the loop as unbounded 5: exit
To teach verifier to accept the above prog, this patch adds bound tracking for BPF_MOD.
The approach is based on the following rules:
1. BPF_MOD is unsigned;
2. For an unsigned constant divisor x:
a. when x != 0, the resulted dst_reg bits are in the range [0, x - 1], and if no wrapping occurs, the result can be further narrowed down to [umin mod x, umax mod x];
b. when x == 0, dst_reg is truncated to 32 bits by mod32 or remains unchanged by mod64.
Signed-off-by: Xu Kuohai xukuohai@huawei.com --- kernel/bpf/verifier.c | 98 ++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 93 insertions(+), 5 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 64f06f6e16bf..e8e37f587d6c 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -12085,6 +12085,87 @@ static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg, __update_reg_bounds(dst_reg); }
+static void scalar32_min_max_mod(struct bpf_reg_state *dst_reg, + struct bpf_reg_state *src_reg) +{ + u32 val = (u32)src_reg->var_off.value; /* src_reg is constant */ + u64 umax = dst_reg->u32_max_value; /* do_div requires u64 */ + u64 umin = dst_reg->u32_min_value; /* do_div requires u64 */ + u32 umax_rem, umin_rem; + + /* dst_reg is 32-bit truncated when mod32 zero, since + * adjust_scalar_min_max_vals invokes zext_32_to_64 to do truncation + * for all alu32 ops, here we do nothing and just return. + */ + if (!val) + return; + + umax_rem = do_div(umax, val); + umin_rem = do_div(umin, val); + + /* no wrapping */ + if (umax - umin < val && umin_rem <= umax_rem) { + dst_reg->var_off = tnum_range(umin_rem, umax_rem); + dst_reg->u32_min_value = umin_rem; + dst_reg->u32_max_value = umax_rem; + } else { + dst_reg->var_off = tnum_range(0, val - 1); + dst_reg->u32_min_value = 0; + dst_reg->u32_max_value = val - 1; + } + + /* cross the sign boundary */ + if ((s32)dst_reg->u32_min_value > (s32)dst_reg->u32_max_value) { + dst_reg->s32_min_value = S32_MIN; + dst_reg->s32_max_value = S32_MAX; + } else { + dst_reg->s32_min_value = (s32)dst_reg->u32_min_value; + dst_reg->s32_max_value = (s32)dst_reg->u32_max_value; + } + + /* mark reg64 unbounded to deduce 64-bit bounds from var_off */ + __mark_reg64_unbounded(dst_reg); +} + +static void scalar_min_max_mod(struct bpf_reg_state *dst_reg, + struct bpf_reg_state *src_reg) +{ + u64 val = src_reg->var_off.value; /* src_reg is constant */ + u64 umax = dst_reg->umax_value; + u64 umin = dst_reg->umin_value; + u64 umax_rem, umin_rem; + + /* dst_reg is untouched when mod64 zero */ + if (!val) + return; + + div64_u64_rem(umin, val, &umin_rem); + div64_u64_rem(umax, val, &umax_rem); + + /* no wrapping */ + if (umax - umin < val && umin_rem <= umax_rem) { + dst_reg->var_off = tnum_range(umin_rem, umax_rem); + dst_reg->umin_value = umin_rem; + dst_reg->umax_value = umax_rem; + } else { + dst_reg->var_off = tnum_range(0, val - 1); + dst_reg->umin_value = 0; + dst_reg->umax_value = val - 1; + } + + /* cross the sign boundary */ + if ((s64)dst_reg->umin_value > (s64)dst_reg->umax_value) { + dst_reg->smin_value = S64_MIN; + dst_reg->smax_value = S64_MAX; + } else { + dst_reg->smin_value = (s64)dst_reg->umin_value; + dst_reg->smax_value = (s64)dst_reg->umax_value; + } + + /* mark reg32 unbounded to deduce 32-bit bounds from var_off */ + __mark_reg32_unbounded(dst_reg); +} + /* WARNING: This function does calculations on 64-bit values, but the actual * execution may occur on 32-bit values. Therefore, things like bitshifts * need extra checks in the 32-bit case. @@ -12159,11 +12240,12 @@ static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env, * and BPF_OR. This is possible because these ops have fairly easy to * understand and calculate behavior in both 32-bit and 64-bit alu ops. * See alu32 verifier tests for examples. The second class of - * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy - * with regards to tracking sign/unsigned bounds because the bits may - * cross subreg boundaries in the alu64 case. When this happens we mark - * the reg unbounded in the subreg bound space and use the resulting - * tnum to calculate an approximation of the sign/unsigned bounds. + * operations, BPF_LSH, BPF_RSH, BPF_ARSH and BPF_MOD, however are not + * so easy with regards to tracking sign/unsigned bounds because the + * bits may cross subreg boundaries in the alu64 case. When this happens + * we mark the reg unbounded in the subreg bound space and use the + * resulting tnum to calculate an approximation of the sign/unsigned + * bounds. */ switch (opcode) { case BPF_ADD: @@ -12235,6 +12317,12 @@ static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env, else scalar_min_max_arsh(dst_reg, &src_reg); break; + case BPF_MOD: + if (alu32) + scalar32_min_max_mod(dst_reg, &src_reg); + else + scalar_min_max_mod(dst_reg, &src_reg); + break; default: mark_reg_unknown(env, regs, insn->dst_reg); break;
On Thu, Mar 23, 2023 at 8:59 AM Xu Kuohai xukuohai@huaweicloud.com wrote:
From: Xu Kuohai xukuohai@huawei.com
dst_reg is marked as unknown when BPF_MOD instruction is verified, causing the following bpf prog to be incorrectly rejected.
0: r0 = 0 1: r0 %= 1 // r0 is marked as unknown 2: r1 = 0 3: r1 += 1 4: if r1 < r0 goto pc-2 // verifier treats the loop as unbounded 5: exit
To teach verifier to accept the above prog, this patch adds bound tracking for BPF_MOD.
The approach is based on the following rules:
BPF_MOD is unsigned;
For an unsigned constant divisor x:
a. when x != 0, the resulted dst_reg bits are in the range [0, x - 1], and if no wrapping occurs, the result can be further narrowed down to [umin mod x, umax mod x];
b. when x == 0, dst_reg is truncated to 32 bits by mod32 or remains unchanged by mod64.
Signed-off-by: Xu Kuohai xukuohai@huawei.com
Same Nack as before. You haven't answered _why_ anyone needs it.
On 3/24/2023 1:16 AM, Alexei Starovoitov wrote:
On Thu, Mar 23, 2023 at 8:59 AM Xu Kuohai xukuohai@huaweicloud.com wrote:
From: Xu Kuohai xukuohai@huawei.com
dst_reg is marked as unknown when BPF_MOD instruction is verified, causing the following bpf prog to be incorrectly rejected.
0: r0 = 0 1: r0 %= 1 // r0 is marked as unknown 2: r1 = 0 3: r1 += 1 4: if r1 < r0 goto pc-2 // verifier treats the loop as unbounded 5: exit
To teach verifier to accept the above prog, this patch adds bound tracking for BPF_MOD.
The approach is based on the following rules:
BPF_MOD is unsigned;
For an unsigned constant divisor x:
a. when x != 0, the resulted dst_reg bits are in the range [0, x - 1], and if no wrapping occurs, the result can be further narrowed down to [umin mod x, umax mod x];
b. when x == 0, dst_reg is truncated to 32 bits by mod32 or remains unchanged by mod64.
Signed-off-by: Xu Kuohai xukuohai@huawei.com
Same Nack as before.
Sorry, I did not receive Nack for this patch before.
You haven't answered _why_ anyone needs it.
No idea, I simply believe it's good to have a closer estimate of the result.
From: Xu Kuohai xukuohai@huawei.com
Test cases to check if verifier tracks dst_reg bound for BPF_MOD.
Signed-off-by: Xu Kuohai xukuohai@huawei.com --- tools/testing/selftests/bpf/verifier/mod.c | 320 +++++++++++++++++++++ 1 file changed, 320 insertions(+) create mode 100644 tools/testing/selftests/bpf/verifier/mod.c
diff --git a/tools/testing/selftests/bpf/verifier/mod.c b/tools/testing/selftests/bpf/verifier/mod.c new file mode 100644 index 000000000000..3aec856d5c9f --- /dev/null +++ b/tools/testing/selftests/bpf/verifier/mod.c @@ -0,0 +1,320 @@ +{ + "mod64 positive imm", + .insns = { + BPF_MOV64_IMM(BPF_REG_0, 0), + BPF_ALU64_IMM(BPF_MOD, BPF_REG_0, 1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 0 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod64 positive reg", + .insns = { + BPF_MOV64_IMM(BPF_REG_0, 0), + BPF_MOV64_IMM(BPF_REG_1, 1), + BPF_ALU64_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 0 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod64 zero", + .insns = { + BPF_MOV64_IMM(BPF_REG_0, 0), + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 0 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod64 negative 1", + .insns = { + BPF_MOV64_IMM(BPF_REG_0, 1), + BPF_MOV64_IMM(BPF_REG_1, -1), + BPF_ALU64_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 1 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, + .retval = 1, +}, +{ + "mod64 negative 2", + .insns = { + BPF_MOV64_IMM(BPF_REG_0, -4), + BPF_MOV32_IMM(BPF_REG_1, 5), + BPF_ALU64_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 2 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, + .retval = 2, +}, +{ + "mod64 negative 3", + .insns = { + BPF_MOV32_IMM(BPF_REG_0, -4), + BPF_MOV32_IMM(BPF_REG_1, -5), + BPF_ALU64_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 1 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, + .retval = 1, +}, +{ + "mod64 variable dividend cross signed boundary, with JLT", + .insns = { + BPF_LDX_MEM(BPF_W, BPF_REG_2, BPF_REG_1, + offsetof(struct xdp_md, data)), + BPF_LDX_MEM(BPF_W, BPF_REG_3, BPF_REG_1, + offsetof(struct xdp_md, data_end)), + BPF_MOV64_REG(BPF_REG_1, BPF_REG_2), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + BPF_JMP_REG(BPF_JGT, BPF_REG_1, BPF_REG_3, 11), + + BPF_LDX_MEM(BPF_B, BPF_REG_1, BPF_REG_2, 0), + BPF_LD_IMM64(BPF_REG_0, 0x7fffffffffffff10), + BPF_ALU64_REG(BPF_ADD, BPF_REG_1, BPF_REG_0), + BPF_LD_IMM64(BPF_REG_0, 0x80000000000000ff), + BPF_ALU64_REG(BPF_MOD, BPF_REG_1, BPF_REG_0), + + BPF_LD_IMM64(BPF_REG_0, 0x8000000000000000), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_0, 1), + /* r1 unsigned range is [0x7fffffffffffff10, 0x800000000000000f] */ + BPF_JMP_REG(BPF_JLT, BPF_REG_0, BPF_REG_1, -2), + + BPF_MOV64_IMM(BPF_REG_0, 0), + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod64 variable dividend cross signed boundary, with JSLT", + .insns = { + BPF_LDX_MEM(BPF_W, BPF_REG_2, BPF_REG_1, + offsetof(struct xdp_md, data)), + BPF_LDX_MEM(BPF_W, BPF_REG_3, BPF_REG_1, + offsetof(struct xdp_md, data_end)), + BPF_MOV64_REG(BPF_REG_1, BPF_REG_2), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + BPF_JMP_REG(BPF_JGT, BPF_REG_1, BPF_REG_3, 11), + + BPF_LDX_MEM(BPF_B, BPF_REG_1, BPF_REG_2, 0), + BPF_LD_IMM64(BPF_REG_0, 0x7fffffffffffff10), + BPF_ALU64_REG(BPF_ADD, BPF_REG_1, BPF_REG_0), + BPF_LD_IMM64(BPF_REG_0, 0x80000000000000ff), + BPF_ALU64_REG(BPF_MOD, BPF_REG_1, BPF_REG_0), + + BPF_LD_IMM64(BPF_REG_0, 0x8000000000000000), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_0, 1), + /* r1 signed range is [S64_MIN, S64_MAX] */ + BPF_JMP_REG(BPF_JSLT, BPF_REG_0, BPF_REG_1, -2), + + BPF_MOV64_IMM(BPF_REG_0, 0), + BPF_EXIT_INSN(), + }, + .errstr = "BPF program is too large.", + .result = REJECT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod32 positive imm", + .insns = { + BPF_MOV32_IMM(BPF_REG_0, 0), + BPF_ALU32_IMM(BPF_MOD, BPF_REG_0, 1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 0 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod32 positive reg", + .insns = { + BPF_MOV32_IMM(BPF_REG_0, 0), + BPF_MOV32_IMM(BPF_REG_1, 1), + BPF_ALU32_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 0 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod32 zero", + .insns = { + BPF_MOV32_IMM(BPF_REG_0, 0), + BPF_MOV32_IMM(BPF_REG_1, 0), + BPF_ALU32_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 0 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod32 negative 1", + .insns = { + BPF_MOV32_IMM(BPF_REG_0, 1), + BPF_MOV32_IMM(BPF_REG_1, -1), + BPF_ALU32_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 1 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, + .retval = 1, +}, +{ + "mod32 negative 2", + .insns = { + BPF_MOV32_IMM(BPF_REG_0, -4), + BPF_MOV32_IMM(BPF_REG_1, 5), + BPF_ALU32_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 2 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, + .retval = 2, +}, +{ + "mod32 negative 3", + .insns = { + BPF_MOV32_IMM(BPF_REG_0, -4), + BPF_MOV32_IMM(BPF_REG_1, -5), + BPF_ALU32_REG(BPF_MOD, BPF_REG_0, BPF_REG_1), + + BPF_MOV64_IMM(BPF_REG_1, 0), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + /* r0 = 1 */ + BPF_JMP_REG(BPF_JLT, BPF_REG_1, BPF_REG_0, -2), + + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, + .retval = 1, +}, +{ + "mod32 variable dividend cross signed boundary, with JLT", + .insns = { + BPF_LDX_MEM(BPF_W, BPF_REG_2, BPF_REG_1, + offsetof(struct xdp_md, data)), + BPF_LDX_MEM(BPF_W, BPF_REG_3, BPF_REG_1, + offsetof(struct xdp_md, data_end)), + BPF_MOV64_REG(BPF_REG_1, BPF_REG_2), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + BPF_JMP_REG(BPF_JGT, BPF_REG_1, BPF_REG_3, 7), + + BPF_LDX_MEM(BPF_B, BPF_REG_1, BPF_REG_2, 0), + BPF_MOV64_IMM(BPF_REG_0, 0x7fffff10), + BPF_ALU64_REG(BPF_ADD, BPF_REG_1, BPF_REG_0), + BPF_ALU32_IMM(BPF_MOD, BPF_REG_1, 0x800000ff), + + BPF_MOV32_IMM(BPF_REG_0, 0x80000000), + BPF_ALU32_IMM(BPF_ADD, BPF_REG_0, 1), + /* r1 unsigned 32-bit range is [0x7fffff10, 0x8000000f] */ + BPF_JMP32_REG(BPF_JLT, BPF_REG_0, BPF_REG_1, -2), + + BPF_MOV64_IMM(BPF_REG_0, 0), + BPF_EXIT_INSN(), + }, + .result = ACCEPT, + .prog_type = BPF_PROG_TYPE_XDP, +}, +{ + "mod32 variable dividend cross signed boundary, with JSLT", + .insns = { + BPF_LDX_MEM(BPF_W, BPF_REG_2, BPF_REG_1, + offsetof(struct xdp_md, data)), + BPF_LDX_MEM(BPF_W, BPF_REG_3, BPF_REG_1, + offsetof(struct xdp_md, data_end)), + BPF_MOV64_REG(BPF_REG_1, BPF_REG_2), + BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, 1), + BPF_JMP_REG(BPF_JGT, BPF_REG_1, BPF_REG_3, 7), + + BPF_LDX_MEM(BPF_B, BPF_REG_1, BPF_REG_2, 0), + BPF_MOV64_IMM(BPF_REG_0, 0x7fffff10), + BPF_ALU64_REG(BPF_ADD, BPF_REG_1, BPF_REG_0), + BPF_ALU32_IMM(BPF_MOD, BPF_REG_1, 0x800000ff), + + BPF_MOV32_IMM(BPF_REG_0, 0x80000000), + BPF_ALU32_IMM(BPF_ADD, BPF_REG_0, 1), + /* r1 signed 32-bit range is [S32_MIN, S32_MAX] */ + BPF_JMP32_REG(BPF_JSLT, BPF_REG_0, BPF_REG_1, -2), + + BPF_MOV64_IMM(BPF_REG_0, 0), + BPF_EXIT_INSN(), + }, + .errstr = "BPF program is too large.", + .result = REJECT, + .prog_type = BPF_PROG_TYPE_XDP, +},
linux-kselftest-mirror@lists.linaro.org
-
Alexei Starovoitov -
Xu Kuohai