From: Maxim Mikityanskiy maxim@isovalent.com
The goal of this series is to extend the verifier's capabilities of tracking scalars when they are spilled to stack, especially when the spill or fill is narrowing. It also contains a fix by Eduard for infinite loop detection and a state pruning optimization by Eduard that compensates for a verification complexity regression introduced by tracking unbounded scalars. These improvements reduce the surface of false rejections that I saw while working on Cilium codebase.
Patch 1 (Maxim): Fix for an existing test, it will matter later in the series.
Patches 2-3 (Eduard): Fixes for false rejections in infinite loop detection that happen in the selftests when my patches are applied.
Patches 4-5 (Maxim): Fix the inconsistency of find_equal_scalars that was possible if 32-bit spills were made.
Patches 6-11 (Maxim): Support the case when boundary checks are first performed after the register was spilled to the stack.
Patches 12-13 (Maxim): Support narrowing fills.
Patches 14-15 (Eduard): Optimization for state pruning in stacksafe() to mitigate the verification complexity regression.
veristat -e file,prog,states -f '!states_diff<50' -f '!states_pct<10' -f '!states_a<10' -f '!states_b<10' -C ...
* Without patch 14:
File Program States (A) States (B) States (DIFF) -------------------- ------------ ---------- ---------- ---------------- bpf_xdp.o tail_lb_ipv6 3877 2936 -941 (-24.27%) pyperf180.bpf.o on_event 8422 10456 +2034 (+24.15%) pyperf600.bpf.o on_event 22259 37319 +15060 (+67.66%) pyperf600_iter.bpf.o on_event 400 540 +140 (+35.00%) strobemeta.bpf.o on_event 4702 13435 +8733 (+185.73%)
* With patch 14:
File Program States (A) States (B) States (DIFF) -------------------- ------------ ---------- ---------- -------------- bpf_xdp.o tail_lb_ipv6 3877 2937 -940 (-24.25%) pyperf600_iter.bpf.o on_event 400 500 +100 (+25.00%)
Eduard Zingerman (4): bpf: make infinite loop detection in is_state_visited() exact selftests/bpf: check if imprecise stack spills confuse infinite loop detection bpf: Optimize state pruning for spilled scalars selftests/bpf: states pruning checks for scalar vs STACK_{MISC,ZERO}
Maxim Mikityanskiy (11): selftests/bpf: Fix the u64_offset_to_skb_data test bpf: Make bpf_for_each_spilled_reg consider narrow spills selftests/bpf: Add a test case for 32-bit spill tracking bpf: Add the assign_scalar_id_before_mov function bpf: Add the get_reg_width function bpf: Assign ID to scalars on spill selftests/bpf: Test assigning ID to scalars on spill bpf: Track spilled unbounded scalars selftests/bpf: Test tracking spilled unbounded scalars bpf: Preserve boundaries and track scalars on narrowing fill selftests/bpf: Add test cases for narrowing fill
include/linux/bpf_verifier.h | 2 +- kernel/bpf/verifier.c | 160 +++++- .../bpf/progs/verifier_direct_packet_access.c | 2 +- .../selftests/bpf/progs/verifier_loops1.c | 24 + .../selftests/bpf/progs/verifier_spill_fill.c | 529 +++++++++++++++++- .../testing/selftests/bpf/verifier/precise.c | 6 +- 6 files changed, 677 insertions(+), 46 deletions(-)
From: Maxim Mikityanskiy maxim@isovalent.com
The u64_offset_to_skb_data test is supposed to make a 64-bit fill, but instead makes a 16-bit one. Fix the test according to its intention. The 16-bit fill is covered by u16_offset_to_skb_data.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- tools/testing/selftests/bpf/progs/verifier_spill_fill.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index 39fe3372e0e0..84eccab36582 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -243,7 +243,7 @@ l0_%=: r0 = 0; \
SEC("tc") __description("Spill u32 const scalars. Refill as u64. Offset to skb->data") -__failure __msg("invalid access to packet") +__failure __msg("math between pkt pointer and register with unbounded min value is not allowed") __naked void u64_offset_to_skb_data(void) { asm volatile (" \ @@ -253,7 +253,7 @@ __naked void u64_offset_to_skb_data(void) w7 = 20; \ *(u32*)(r10 - 4) = r6; \ *(u32*)(r10 - 8) = r7; \ - r4 = *(u16*)(r10 - 8); \ + r4 = *(u64*)(r10 - 8); \ r0 = r2; \ /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=umax=65535 */\ r0 += r4; \
On Wed, Dec 20, 2023 at 11:39:59PM +0200, Maxim Mikityanskiy wrote:
From: Maxim Mikityanskiy maxim@isovalent.com
The u64_offset_to_skb_data test is supposed to make a 64-bit fill, but instead makes a 16-bit one. Fix the test according to its intention. The 16-bit fill is covered by u16_offset_to_skb_data.
Cover letter mentioned
Patch 1 (Maxim): Fix for an existing test, it will matter later in the series.
However no subsequent patch touch upon u64_offset_to_skb_data(). Was the followup missing from this series?
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com [...] SEC("tc") __description("Spill u32 const scalars. Refill as u64. Offset to skb->data") -__failure __msg("invalid access to packet") +__failure __msg("math between pkt pointer and register with unbounded min value is not allowed") __naked void u64_offset_to_skb_data(void) { asm volatile (" \ @@ -253,7 +253,7 @@ __naked void u64_offset_to_skb_data(void) w7 = 20; \ *(u32*)(r10 - 4) = r6; \ *(u32*)(r10 - 8) = r7; \
- r4 = *(u16*)(r10 - 8); \
- r4 = *(u64*)(r10 - 8); \ r0 = r2; \ /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=umax=65535 */\ r0 += r4; \
On Tue, 26 Dec 2023 at 17:52:56 +0800, Shung-Hsi Yu wrote:
On Wed, Dec 20, 2023 at 11:39:59PM +0200, Maxim Mikityanskiy wrote:
From: Maxim Mikityanskiy maxim@isovalent.com
The u64_offset_to_skb_data test is supposed to make a 64-bit fill, but instead makes a 16-bit one. Fix the test according to its intention. The 16-bit fill is covered by u16_offset_to_skb_data.
Cover letter mentioned
Patch 1 (Maxim): Fix for an existing test, it will matter later in the series.
However no subsequent patch touch upon u64_offset_to_skb_data(). Was the followup missing from this series?
Thanks for your vigilance, but it's actually correct, sorry for not making it clear enough. In patch 11 ("bpf: Preserve boundaries and track scalars on narrowing fill") I modify u16_offset_to_skb_data, because it becomes a valid pattern after that change. If I didn't change and fix u64_offset_to_skb_data here, I'd need to modify it in patch 11 as well (that's what I meant when I said "it will matter later in the series", it's indeed subtle and implicit, now that I look at it), because it would also start passing, however, that's not what we want, because:
1. Both tests would essentially test the same thing: a 16-bit fill after a 32-bit spill.
2. The description of u64_offset_to_skb_data clearly says: "Refill as u64". It's a typo in the code, u16->u64 makes sense, because we spill two u32s and fill them as a single u64.
So, this patch essentially prevents wrong changes in a further patch.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com [...] SEC("tc") __description("Spill u32 const scalars. Refill as u64. Offset to skb->data") -__failure __msg("invalid access to packet") +__failure __msg("math between pkt pointer and register with unbounded min value is not allowed") __naked void u64_offset_to_skb_data(void) { asm volatile (" \ @@ -253,7 +253,7 @@ __naked void u64_offset_to_skb_data(void) w7 = 20; \ *(u32*)(r10 - 4) = r6; \ *(u32*)(r10 - 8) = r7; \
- r4 = *(u16*)(r10 - 8); \
- r4 = *(u64*)(r10 - 8); \ r0 = r2; \ /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=umax=65535 */\ r0 += r4; \
On Tue, Dec 26, 2023 at 12:38:06PM +0200, Maxim Mikityanskiy wrote:
On Tue, 26 Dec 2023 at 17:52:56 +0800, Shung-Hsi Yu wrote:
On Wed, Dec 20, 2023 at 11:39:59PM +0200, Maxim Mikityanskiy wrote:
From: Maxim Mikityanskiy maxim@isovalent.com
The u64_offset_to_skb_data test is supposed to make a 64-bit fill, but instead makes a 16-bit one. Fix the test according to its intention. The 16-bit fill is covered by u16_offset_to_skb_data.
Cover letter mentioned
Patch 1 (Maxim): Fix for an existing test, it will matter later in the series.
However no subsequent patch touch upon u64_offset_to_skb_data(). Was the followup missing from this series?
Thanks for your vigilance, but it's actually correct, sorry for not making it clear enough. In patch 11 ("bpf: Preserve boundaries and track scalars on narrowing fill") I modify u16_offset_to_skb_data, because it becomes a valid pattern after that change. If I didn't change and fix u64_offset_to_skb_data here, I'd need to modify it in patch 11 as well (that's what I meant when I said "it will matter later in the series", it's indeed subtle and implicit, now that I look at it), because it would also start passing, however, that's not what we want, because:
- Both tests would essentially test the same thing: a 16-bit fill after
a 32-bit spill.
- The description of u64_offset_to_skb_data clearly says: "Refill as
u64". It's a typo in the code, u16->u64 makes sense, because we spill two u32s and fill them as a single u64.
So, this patch essentially prevents wrong changes in a further patch.
Thank for the thorough explanation. Now I can see and agree that the u16->u64 change should be made. Digging back a big, the change also aligns with what's said in commit 0be2516f865f5 ("selftests/bpf: Tests for state pruning with u32 spill/fill") that introduced the check:
... checks that a filled u64 register is marked unknown if the register spilled in the same slack slot was less than 8B.
Side note: the r4 value in comment is still "R4=umax=65535", that probably should be updated as well now that r4 is unbounded.
[...]
- r4 = *(u16*)(r10 - 8); \
- r4 = *(u64*)(r10 - 8); \ r0 = r2; \ /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=umax=65535 */\ r0 += r4; \
From: Eduard Zingerman eddyz87@gmail.com
Current infinite loops detection mechanism is speculative: - first, states_maybe_looping() check is done which simply does memcmp for R1-R10 in current frame; - second, states_equal(..., exact=false) is called. With exact=false states_equal() would compare scalars for equality only if in old state scalar has precision mark.
Such logic might be problematic if compiler makes some unlucky stack spill/fill decisions. An artificial example of a false positive looks as follows:
r0 = ... unknown scalar ... r0 &= 0xff; *(u64 *)(r10 - 8) = r0; r0 = 0; loop: r0 = *(u64 *)(r10 - 8); if r0 > 10 goto exit_; r0 += 1; *(u64 *)(r10 - 8) = r0; r0 = 0; goto loop;
This commit updates call to states_equal to use exact=true, forcing all scalar comparisons to be exact.
Signed-off-by: Eduard Zingerman eddyz87@gmail.com --- kernel/bpf/verifier.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index f13008d27f35..89f8c527ed3c 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -17008,7 +17008,7 @@ static int is_state_visited(struct bpf_verifier_env *env, int insn_idx) } /* attempt to detect infinite loop to avoid unnecessary doomed work */ if (states_maybe_looping(&sl->state, cur) && - states_equal(env, &sl->state, cur, false) && + states_equal(env, &sl->state, cur, true) && !iter_active_depths_differ(&sl->state, cur) && sl->state.callback_unroll_depth == cur->callback_unroll_depth) { verbose_linfo(env, insn_idx, "; ");
From: Eduard Zingerman eddyz87@gmail.com
Verify that infinite loop detection logic separates states with identical register states but different imprecise scalars spilled to stack.
Signed-off-by: Eduard Zingerman eddyz87@gmail.com --- .../selftests/bpf/progs/verifier_loops1.c | 24 +++++++++++++++++++ 1 file changed, 24 insertions(+)
diff --git a/tools/testing/selftests/bpf/progs/verifier_loops1.c b/tools/testing/selftests/bpf/progs/verifier_loops1.c index 71735dbf33d4..e07b43b78fd2 100644 --- a/tools/testing/selftests/bpf/progs/verifier_loops1.c +++ b/tools/testing/selftests/bpf/progs/verifier_loops1.c @@ -259,4 +259,28 @@ l0_%=: r2 += r1; \ " ::: __clobber_all); }
+SEC("xdp") +__success +__naked void not_an_inifinite_loop(void) +{ + asm volatile (" \ + call %[bpf_get_prandom_u32]; \ + r0 &= 0xff; \ + *(u64 *)(r10 - 8) = r0; \ + r0 = 0; \ +loop_%=: \ + r0 = *(u64 *)(r10 - 8); \ + if r0 > 10 goto exit_%=; \ + r0 += 1; \ + *(u64 *)(r10 - 8) = r0; \ + r0 = 0; \ + goto loop_%=; \ +exit_%=: \ + r0 = 0; \ + exit; \ +" : + : __imm(bpf_get_prandom_u32) + : __clobber_all); +} + char _license[] SEC("license") = "GPL";
From: Maxim Mikityanskiy maxim@isovalent.com
Adjust the check in bpf_get_spilled_reg to take into account spilled registers narrower than 64 bits. That allows find_equal_scalars to properly adjust the range of all spilled registers that have the same ID. Before this change, it was possible for a register and a spilled register to have the same IDs but different ranges if the spill was narrower than 64 bits and a range check was performed on the register.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- include/linux/bpf_verifier.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h index d07d857ca67f..e11baecbde68 100644 --- a/include/linux/bpf_verifier.h +++ b/include/linux/bpf_verifier.h @@ -453,7 +453,7 @@ struct bpf_verifier_state {
#define bpf_get_spilled_reg(slot, frame, mask) \ (((slot < frame->allocated_stack / BPF_REG_SIZE) && \ - ((1 << frame->stack[slot].slot_type[0]) & (mask))) \ + ((1 << frame->stack[slot].slot_type[BPF_REG_SIZE - 1]) & (mask))) \ ? &frame->stack[slot].spilled_ptr : NULL)
/* Iterate over 'frame', setting 'reg' to either NULL or a spilled register. */
From: Maxim Mikityanskiy maxim@isovalent.com
When a range check is performed on a register that was 32-bit spilled to the stack, the IDs of the two instances of the register are the same, so the range should also be the same.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- .../selftests/bpf/progs/verifier_spill_fill.c | 31 +++++++++++++++++++ 1 file changed, 31 insertions(+)
diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index 84eccab36582..f2c1fe5b1dba 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -737,4 +737,35 @@ __naked void stack_load_preserves_const_precision_subreg(void) : __clobber_common); }
+SEC("xdp") +__description("32-bit spilled reg range should be tracked") +__success __retval(0) +__naked void spill_32bit_range_track(void) +{ + asm volatile(" \ + call %[bpf_ktime_get_ns]; \ + /* Make r0 bounded. */ \ + r0 &= 65535; \ + /* Assign an ID to r0. */ \ + r1 = r0; \ + /* 32-bit spill r0 to stack. */ \ + *(u32*)(r10 - 8) = r0; \ + /* Boundary check on r0. */ \ + if r0 < 1 goto l0_%=; \ + /* 32-bit fill r1 from stack. */ \ + r1 = *(u32*)(r10 - 8); \ + /* r1 == r0 => r1 >= 1 always. */ \ + if r1 >= 1 goto l0_%=; \ + /* Dead branch: the verifier should prune it. \ + * Do an invalid memory access if the verifier \ + * follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ +l0_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_ktime_get_ns) + : __clobber_all); +} + char _license[] SEC("license") = "GPL";
From: Maxim Mikityanskiy maxim@isovalent.com
Extract the common code that generates a register ID for src_reg before MOV if needed into a new function. This function will also be used in a following commit.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- kernel/bpf/verifier.c | 33 +++++++++++++++++++-------------- 1 file changed, 19 insertions(+), 14 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 89f8c527ed3c..a703e3adedd3 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4401,6 +4401,18 @@ static bool __is_pointer_value(bool allow_ptr_leaks, return reg->type != SCALAR_VALUE; }
+static void assign_scalar_id_before_mov(struct bpf_verifier_env *env, + struct bpf_reg_state *src_reg) +{ + if (src_reg->type == SCALAR_VALUE && !src_reg->id && + !tnum_is_const(src_reg->var_off)) + /* Ensure that src_reg has a valid ID that will be copied to + * dst_reg and then will be used by find_equal_scalars() to + * propagate min/max range. + */ + src_reg->id = ++env->id_gen; +} + /* Copy src state preserving dst->parent and dst->live fields */ static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src) { @@ -13886,20 +13898,13 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn) if (BPF_SRC(insn->code) == BPF_X) { struct bpf_reg_state *src_reg = regs + insn->src_reg; struct bpf_reg_state *dst_reg = regs + insn->dst_reg; - bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id && - !tnum_is_const(src_reg->var_off);
if (BPF_CLASS(insn->code) == BPF_ALU64) { if (insn->off == 0) { /* case: R1 = R2 * copy register state to dest reg */ - if (need_id) - /* Assign src and dst registers the same ID - * that will be used by find_equal_scalars() - * to propagate min/max range. - */ - src_reg->id = ++env->id_gen; + assign_scalar_id_before_mov(env, src_reg); copy_register_state(dst_reg, src_reg); dst_reg->live |= REG_LIVE_WRITTEN; dst_reg->subreg_def = DEF_NOT_SUBREG; @@ -13914,8 +13919,8 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn) bool no_sext;
no_sext = src_reg->umax_value < (1ULL << (insn->off - 1)); - if (no_sext && need_id) - src_reg->id = ++env->id_gen; + if (no_sext) + assign_scalar_id_before_mov(env, src_reg); copy_register_state(dst_reg, src_reg); if (!no_sext) dst_reg->id = 0; @@ -13937,8 +13942,8 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn) if (insn->off == 0) { bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
- if (is_src_reg_u32 && need_id) - src_reg->id = ++env->id_gen; + if (is_src_reg_u32) + assign_scalar_id_before_mov(env, src_reg); copy_register_state(dst_reg, src_reg); /* Make sure ID is cleared if src_reg is not in u32 * range otherwise dst_reg min/max could be incorrectly @@ -13952,8 +13957,8 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn) /* case: W1 = (s8, s16)W2 */ bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
- if (no_sext && need_id) - src_reg->id = ++env->id_gen; + if (no_sext) + assign_scalar_id_before_mov(env, src_reg); copy_register_state(dst_reg, src_reg); if (!no_sext) dst_reg->id = 0;
From: Maxim Mikityanskiy maxim@isovalent.com
Put calculation of the register value width into a dedicated function. This function will also be used in a following commit.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- kernel/bpf/verifier.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index a703e3adedd3..b757fdbbbdd2 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4448,6 +4448,11 @@ static bool is_bpf_st_mem(struct bpf_insn *insn) return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM; }
+static int get_reg_width(struct bpf_reg_state *reg) +{ + return fls64(reg->umax_value); +} + /* check_stack_{read,write}_fixed_off functions track spill/fill of registers, * stack boundary and alignment are checked in check_mem_access() */ @@ -4500,7 +4505,7 @@ static int check_stack_write_fixed_off(struct bpf_verifier_env *env, if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) && env->bpf_capable) { save_register_state(env, state, spi, reg, size); /* Break the relation on a narrowing spill. */ - if (fls64(reg->umax_value) > BITS_PER_BYTE * size) + if (get_reg_width(reg) > BITS_PER_BYTE * size) state->stack[spi].spilled_ptr.id = 0; } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) && insn->imm != 0 && env->bpf_capable) { @@ -13940,7 +13945,7 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn) return -EACCES; } else if (src_reg->type == SCALAR_VALUE) { if (insn->off == 0) { - bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX; + bool is_src_reg_u32 = get_reg_width(src_reg) <= 32;
if (is_src_reg_u32) assign_scalar_id_before_mov(env, src_reg);
From: Maxim Mikityanskiy maxim@isovalent.com
Currently, when a scalar bounded register is spilled to the stack, its ID is preserved, but only if was already assigned, i.e. if this register was MOVed before.
Assign an ID on spill if none is set, so that equal scalars could be tracked if a register is spilled to the stack and filled into another register.
One test is adjusted to reflect the change in register IDs.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- kernel/bpf/verifier.c | 8 +++++++- .../selftests/bpf/progs/verifier_direct_packet_access.c | 2 +- 2 files changed, 8 insertions(+), 2 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index b757fdbbbdd2..caa768f1e369 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4503,9 +4503,15 @@ static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
mark_stack_slot_scratched(env, spi); if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) && env->bpf_capable) { + bool reg_value_fits; + + reg_value_fits = get_reg_width(reg) <= BITS_PER_BYTE * size; + /* Make sure that reg had an ID to build a relation on spill. */ + if (reg_value_fits) + assign_scalar_id_before_mov(env, reg); save_register_state(env, state, spi, reg, size); /* Break the relation on a narrowing spill. */ - if (get_reg_width(reg) > BITS_PER_BYTE * size) + if (!reg_value_fits) state->stack[spi].spilled_ptr.id = 0; } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) && insn->imm != 0 && env->bpf_capable) { diff --git a/tools/testing/selftests/bpf/progs/verifier_direct_packet_access.c b/tools/testing/selftests/bpf/progs/verifier_direct_packet_access.c index be95570ab382..28b602ac9cbe 100644 --- a/tools/testing/selftests/bpf/progs/verifier_direct_packet_access.c +++ b/tools/testing/selftests/bpf/progs/verifier_direct_packet_access.c @@ -568,7 +568,7 @@ l0_%=: r0 = 0; \
SEC("tc") __description("direct packet access: test23 (x += pkt_ptr, 4)") -__failure __msg("invalid access to packet, off=0 size=8, R5(id=2,off=0,r=0)") +__failure __msg("invalid access to packet, off=0 size=8, R5(id=3,off=0,r=0)") __flag(BPF_F_ANY_ALIGNMENT) __naked void test23_x_pkt_ptr_4(void) {
On Wed, Dec 20, 2023 at 1:40 PM Maxim Mikityanskiy maxtram95@gmail.com wrote:
From: Maxim Mikityanskiy maxim@isovalent.com
Currently, when a scalar bounded register is spilled to the stack, its ID is preserved, but only if was already assigned, i.e. if this register was MOVed before.
Assign an ID on spill if none is set, so that equal scalars could be tracked if a register is spilled to the stack and filled into another register.
One test is adjusted to reflect the change in register IDs.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com
kernel/bpf/verifier.c | 8 +++++++- .../selftests/bpf/progs/verifier_direct_packet_access.c | 2 +- 2 files changed, 8 insertions(+), 2 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index b757fdbbbdd2..caa768f1e369 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4503,9 +4503,15 @@ static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
mark_stack_slot_scratched(env, spi); if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) && env->bpf_capable) {
bool reg_value_fits;reg_value_fits = get_reg_width(reg) <= BITS_PER_BYTE * size;/* Make sure that reg had an ID to build a relation on spill. */if (reg_value_fits)assign_scalar_id_before_mov(env, reg);
Thanks. I just debugged this issue as part of my bpf_cmp series.
llvm generated:
1093: (7b) *(u64 *)(r10 -96) = r0 ; R0_w=scalar(smin=smin32=-4095,smax=smax32=256) R10=fp0 fp-96_w=scalar(smin=smin32=-4095,smax=smax32=256) ; if (bpf_cmp(filepart_length, >, MAX_PATH)) 1094: (25) if r0 > 0x100 goto pc+903 ; R0_w=scalar(id=53,smin=smin32=0,smax=umax=smax32=umax32=256,var_off=(0x0; 0x1ff))
the verifier refined the range of 'r0' here, but the code just read spilled value from stack:
1116: (79) r1 = *(u64 *)(r10 -64) ; R1_w=map_value ; payload += filepart_length; 1117: (79) r2 = *(u64 *)(r10 -96) ; R2_w=scalar(smin=smin32=-4095,smax=smax32=256) R10=fp0 fp-96=scalar(smin=smin32=-4095,smax=smax32=256) 1118: (0f) r1 += r2 ; R1_w=map_value(map=data_heap,ks=4,vs=23040,off=148,smin=smin32=-4095,smax=smax32=3344)
And later errors as: "R1 min value is negative, either use unsigned index or do a if (index
=0) check."
This verifier improvement is certainly necessary.
Since you've analyzed this issue did you figure out a workaround for C code on existing and older kernels?
On Sun, 24 Dec 2023 at 19:15:42 -0800, Alexei Starovoitov wrote:
On Wed, Dec 20, 2023 at 1:40 PM Maxim Mikityanskiy maxtram95@gmail.com wrote:
From: Maxim Mikityanskiy maxim@isovalent.com
Currently, when a scalar bounded register is spilled to the stack, its ID is preserved, but only if was already assigned, i.e. if this register was MOVed before.
Assign an ID on spill if none is set, so that equal scalars could be tracked if a register is spilled to the stack and filled into another register.
One test is adjusted to reflect the change in register IDs.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com
kernel/bpf/verifier.c | 8 +++++++- .../selftests/bpf/progs/verifier_direct_packet_access.c | 2 +- 2 files changed, 8 insertions(+), 2 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index b757fdbbbdd2..caa768f1e369 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4503,9 +4503,15 @@ static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
mark_stack_slot_scratched(env, spi); if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) && env->bpf_capable) {
bool reg_value_fits;reg_value_fits = get_reg_width(reg) <= BITS_PER_BYTE * size;/* Make sure that reg had an ID to build a relation on spill. */if (reg_value_fits)assign_scalar_id_before_mov(env, reg);Thanks. I just debugged this issue as part of my bpf_cmp series.
llvm generated:
1093: (7b) *(u64 *)(r10 -96) = r0 ; R0_w=scalar(smin=smin32=-4095,smax=smax32=256) R10=fp0 fp-96_w=scalar(smin=smin32=-4095,smax=smax32=256) ; if (bpf_cmp(filepart_length, >, MAX_PATH)) 1094: (25) if r0 > 0x100 goto pc+903 ; R0_w=scalar(id=53,smin=smin32=0,smax=umax=smax32=umax32=256,var_off=(0x0; 0x1ff))
the verifier refined the range of 'r0' here, but the code just read spilled value from stack:
1116: (79) r1 = *(u64 *)(r10 -64) ; R1_w=map_value ; payload += filepart_length; 1117: (79) r2 = *(u64 *)(r10 -96) ; R2_w=scalar(smin=smin32=-4095,smax=smax32=256) R10=fp0 fp-96=scalar(smin=smin32=-4095,smax=smax32=256) 1118: (0f) r1 += r2 ; R1_w=map_value(map=data_heap,ks=4,vs=23040,off=148,smin=smin32=-4095,smax=smax32=3344)
And later errors as: "R1 min value is negative, either use unsigned index or do a if (index
=0) check."
This verifier improvement is certainly necessary.
Glad that you found it useful!
Since you've analyzed this issue did you figure out a workaround for C code on existing and older kernels?
Uhm... in my case (Cilium, it was a while ago) I did some big change (reorganized function calls and revalidate_data() calls) that changed codegen significantly, and the problematic pattern disappeared.
I can suggest trying to play with volatile, e.g., declare filepart_length as volatile; if it doesn't help, create another volatile variable and copy filepart_length to it before doing bpf_cmp (copying reg->reg will assign an ID, but I'm not sure if they'll still be in registers after being declared as volatile).
Unfortunately, I couldn't reproduce your issue locally, so I couldn't try these suggestions myself. Is this the right code, or should I take it from elsewhere?
https://patchwork.kernel.org/project/netdevbpf/list/?series=812010
What LLVM version do you see the issue on? I can try to look for a specific C workaround if I reproduce it locally.
BTW, the asm workaround is obvious (copy reg to another reg to assign an ID), so maybe an inline asm like this would do the thing?
asm volatile("r8 = %0" :: "r"(filepart_length) : "r8");
On Mon, Dec 25, 2023 at 1:11 PM Maxim Mikityanskiy maxtram95@gmail.com wrote:
On Sun, 24 Dec 2023 at 19:15:42 -0800, Alexei Starovoitov wrote:
On Wed, Dec 20, 2023 at 1:40 PM Maxim Mikityanskiy maxtram95@gmail.com wrote:
From: Maxim Mikityanskiy maxim@isovalent.com
Currently, when a scalar bounded register is spilled to the stack, its ID is preserved, but only if was already assigned, i.e. if this register was MOVed before.
Assign an ID on spill if none is set, so that equal scalars could be tracked if a register is spilled to the stack and filled into another register.
One test is adjusted to reflect the change in register IDs.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com
kernel/bpf/verifier.c | 8 +++++++- .../selftests/bpf/progs/verifier_direct_packet_access.c | 2 +- 2 files changed, 8 insertions(+), 2 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index b757fdbbbdd2..caa768f1e369 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4503,9 +4503,15 @@ static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
mark_stack_slot_scratched(env, spi); if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) && env->bpf_capable) {
bool reg_value_fits;reg_value_fits = get_reg_width(reg) <= BITS_PER_BYTE * size;/* Make sure that reg had an ID to build a relation on spill. */if (reg_value_fits)assign_scalar_id_before_mov(env, reg);Thanks. I just debugged this issue as part of my bpf_cmp series.
llvm generated:
1093: (7b) *(u64 *)(r10 -96) = r0 ; R0_w=scalar(smin=smin32=-4095,smax=smax32=256) R10=fp0 fp-96_w=scalar(smin=smin32=-4095,smax=smax32=256) ; if (bpf_cmp(filepart_length, >, MAX_PATH)) 1094: (25) if r0 > 0x100 goto pc+903 ; R0_w=scalar(id=53,smin=smin32=0,smax=umax=smax32=umax32=256,var_off=(0x0; 0x1ff))
the verifier refined the range of 'r0' here, but the code just read spilled value from stack:
1116: (79) r1 = *(u64 *)(r10 -64) ; R1_w=map_value ; payload += filepart_length; 1117: (79) r2 = *(u64 *)(r10 -96) ; R2_w=scalar(smin=smin32=-4095,smax=smax32=256) R10=fp0 fp-96=scalar(smin=smin32=-4095,smax=smax32=256) 1118: (0f) r1 += r2 ; R1_w=map_value(map=data_heap,ks=4,vs=23040,off=148,smin=smin32=-4095,smax=smax32=3344)
And later errors as: "R1 min value is negative, either use unsigned index or do a if (index
=0) check."
This verifier improvement is certainly necessary.
Glad that you found it useful!
Since you've analyzed this issue did you figure out a workaround for C code on existing and older kernels?
Uhm... in my case (Cilium, it was a while ago) I did some big change (reorganized function calls and revalidate_data() calls) that changed codegen significantly, and the problematic pattern disappeared.
I can suggest trying to play with volatile, e.g., declare filepart_length as volatile; if it doesn't help, create another volatile variable and copy filepart_length to it before doing bpf_cmp (copying reg->reg will assign an ID, but I'm not sure if they'll still be in registers after being declared as volatile).
Unfortunately, I couldn't reproduce your issue locally, so I couldn't try these suggestions myself.
No worries.
What LLVM version do you see the issue on? I can try to look for a specific C workaround if I reproduce it locally.
BTW, the asm workaround is obvious (copy reg to another reg to assign an ID), so maybe an inline asm like this would do the thing?
asm volatile("r8 = %0" :: "r"(filepart_length) : "r8");
Right. I tried: asm volatile("%[reg]=%[reg]"::[reg]"r"((short)filepart_length));
and it forces ID assignment, but depending on the code it might still be too late. I've seen the pattern: call ... *(u64 *)(r10 -96) = r0 r0 = r0 // asm trick above if r0 > 0x100 goto pc+903
So it may or may not help, but it was good to understand this issue.
From: Maxim Mikityanskiy maxim@isovalent.com
The previous commit implemented assigning IDs to registers holding scalars before spill. Add the test cases to check the new functionality.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- .../selftests/bpf/progs/verifier_spill_fill.c | 133 ++++++++++++++++++ 1 file changed, 133 insertions(+)
diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index f2c1fe5b1dba..86881eaab4e2 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -768,4 +768,137 @@ l0_%=: r0 = 0; \ : __clobber_all); }
+SEC("xdp") +__description("64-bit spill of 64-bit reg should assign ID") +__success __retval(0) +__naked void spill_64bit_of_64bit_ok(void) +{ + asm volatile (" \ + /* Roll one bit to make the register inexact. */\ + call %[bpf_get_prandom_u32]; \ + r0 &= 0x80000000; \ + r0 <<= 32; \ + /* 64-bit spill r0 to stack - should assign an ID. */\ + *(u64*)(r10 - 8) = r0; \ + /* 64-bit fill r1 from stack - should preserve the ID. */\ + r1 = *(u64*)(r10 - 8); \ + /* Compare r1 with another register to trigger find_equal_scalars.\ + * Having one random bit is important here, otherwise the verifier cuts\ + * the corners. \ + */ \ + r2 = 0; \ + if r1 != r2 goto l0_%=; \ + /* The result of this comparison is predefined. */\ + if r0 == r2 goto l0_%=; \ + /* Dead branch: the verifier should prune it. Do an invalid memory\ + * access if the verifier follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ + exit; \ +l0_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_get_prandom_u32) + : __clobber_all); +} + +SEC("xdp") +__description("32-bit spill of 32-bit reg should assign ID") +__success __retval(0) +__naked void spill_32bit_of_32bit_ok(void) +{ + asm volatile (" \ + /* Roll one bit to make the register inexact. */\ + call %[bpf_get_prandom_u32]; \ + w0 &= 0x80000000; \ + /* 32-bit spill r0 to stack - should assign an ID. */\ + *(u32*)(r10 - 8) = r0; \ + /* 32-bit fill r1 from stack - should preserve the ID. */\ + r1 = *(u32*)(r10 - 8); \ + /* Compare r1 with another register to trigger find_equal_scalars.\ + * Having one random bit is important here, otherwise the verifier cuts\ + * the corners. \ + */ \ + r2 = 0; \ + if r1 != r2 goto l0_%=; \ + /* The result of this comparison is predefined. */\ + if r0 == r2 goto l0_%=; \ + /* Dead branch: the verifier should prune it. Do an invalid memory\ + * access if the verifier follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ + exit; \ +l0_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_get_prandom_u32) + : __clobber_all); +} + +SEC("xdp") +__description("16-bit spill of 16-bit reg should assign ID") +__success __retval(0) +__naked void spill_16bit_of_16bit_ok(void) +{ + asm volatile (" \ + /* Roll one bit to make the register inexact. */\ + call %[bpf_get_prandom_u32]; \ + r0 &= 0x8000; \ + /* 16-bit spill r0 to stack - should assign an ID. */\ + *(u16*)(r10 - 8) = r0; \ + /* 16-bit fill r1 from stack - should preserve the ID. */\ + r1 = *(u16*)(r10 - 8); \ + /* Compare r1 with another register to trigger find_equal_scalars.\ + * Having one random bit is important here, otherwise the verifier cuts\ + * the corners. \ + */ \ + r2 = 0; \ + if r1 != r2 goto l0_%=; \ + /* The result of this comparison is predefined. */\ + if r0 == r2 goto l0_%=; \ + /* Dead branch: the verifier should prune it. Do an invalid memory\ + * access if the verifier follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ + exit; \ +l0_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_get_prandom_u32) + : __clobber_all); +} + +SEC("xdp") +__description("8-bit spill of 8-bit reg should assign ID") +__success __retval(0) +__naked void spill_8bit_of_8bit_ok(void) +{ + asm volatile (" \ + /* Roll one bit to make the register inexact. */\ + call %[bpf_get_prandom_u32]; \ + r0 &= 0x80; \ + /* 8-bit spill r0 to stack - should assign an ID. */\ + *(u8*)(r10 - 8) = r0; \ + /* 8-bit fill r1 from stack - should preserve the ID. */\ + r1 = *(u8*)(r10 - 8); \ + /* Compare r1 with another register to trigger find_equal_scalars.\ + * Having one random bit is important here, otherwise the verifier cuts\ + * the corners. \ + */ \ + r2 = 0; \ + if r1 != r2 goto l0_%=; \ + /* The result of this comparison is predefined. */\ + if r0 == r2 goto l0_%=; \ + /* Dead branch: the verifier should prune it. Do an invalid memory\ + * access if the verifier follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ + exit; \ +l0_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_get_prandom_u32) + : __clobber_all); +} + char _license[] SEC("license") = "GPL";
From: Maxim Mikityanskiy maxim@isovalent.com
Support the pattern where an unbounded scalar is spilled to the stack, then boundary checks are performed on the src register, after which the stack frame slot is refilled into a register.
Before this commit, the verifier didn't treat the src register and the stack slot as related if the src register was an unbounded scalar. The register state wasn't copied, the id wasn't preserved, and the stack slot was marked as STACK_MISC. Subsequent boundary checks on the src register wouldn't result in updating the boundaries of the spilled variable on the stack.
After this commit, the verifier will preserve the bond between src and dst even if src is unbounded, which permits to do boundary checks on src and refill dst later, still remembering its boundaries. Such a pattern is sometimes generated by clang when compiling complex long functions.
One test is adjusted to reflect the fact that an untracked register is marked as precise at an earlier stage, and one more test is adjusted to reflect that now unbounded scalars are tracked.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- kernel/bpf/verifier.c | 7 +------ tools/testing/selftests/bpf/progs/verifier_spill_fill.c | 6 +++--- tools/testing/selftests/bpf/verifier/precise.c | 6 +++--- 3 files changed, 7 insertions(+), 12 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index caa768f1e369..9b5053389739 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4387,11 +4387,6 @@ static bool __is_scalar_unbounded(struct bpf_reg_state *reg) reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX; }
-static bool register_is_bounded(struct bpf_reg_state *reg) -{ - return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg); -} - static bool __is_pointer_value(bool allow_ptr_leaks, const struct bpf_reg_state *reg) { @@ -4502,7 +4497,7 @@ static int check_stack_write_fixed_off(struct bpf_verifier_env *env, return err;
mark_stack_slot_scratched(env, spi); - if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) && env->bpf_capable) { + if (reg && !(off % BPF_REG_SIZE) && reg->type == SCALAR_VALUE && env->bpf_capable) { bool reg_value_fits;
reg_value_fits = get_reg_width(reg) <= BITS_PER_BYTE * size; diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index 86881eaab4e2..92e446b18e10 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -454,9 +454,9 @@ l0_%=: r1 >>= 16; \ SEC("raw_tp") __log_level(2) __success -__msg("fp-8=0m??mmmm") -__msg("fp-16=00mm??mm") -__msg("fp-24=00mm???m") +__msg("fp-8=0m??scalar()") +__msg("fp-16=00mm??scalar()") +__msg("fp-24=00mm???scalar()") __naked void spill_subregs_preserve_stack_zero(void) { asm volatile ( diff --git a/tools/testing/selftests/bpf/verifier/precise.c b/tools/testing/selftests/bpf/verifier/precise.c index 8a2ff81d8350..0a9293a57211 100644 --- a/tools/testing/selftests/bpf/verifier/precise.c +++ b/tools/testing/selftests/bpf/verifier/precise.c @@ -183,10 +183,10 @@ .prog_type = BPF_PROG_TYPE_XDP, .flags = BPF_F_TEST_STATE_FREQ, .errstr = "mark_precise: frame0: last_idx 7 first_idx 7\ - mark_precise: frame0: parent state regs=r4 stack=:\ + mark_precise: frame0: parent state regs=r4 stack=-8:\ mark_precise: frame0: last_idx 6 first_idx 4\ - mark_precise: frame0: regs=r4 stack= before 6: (b7) r0 = -1\ - mark_precise: frame0: regs=r4 stack= before 5: (79) r4 = *(u64 *)(r10 -8)\ + mark_precise: frame0: regs=r4 stack=-8 before 6: (b7) r0 = -1\ + mark_precise: frame0: regs=r4 stack=-8 before 5: (79) r4 = *(u64 *)(r10 -8)\ mark_precise: frame0: regs= stack=-8 before 4: (7b) *(u64 *)(r3 -8) = r0\ mark_precise: frame0: parent state regs=r0 stack=:\ mark_precise: frame0: last_idx 3 first_idx 3\
From: Maxim Mikityanskiy maxim@isovalent.com
The previous commit added tracking for unbounded scalars on spill. Add the test case to check the new functionality.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- .../selftests/bpf/progs/verifier_spill_fill.c | 27 +++++++++++++++++++ 1 file changed, 27 insertions(+)
diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index 92e446b18e10..809a09732168 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -901,4 +901,31 @@ l0_%=: r0 = 0; \ : __clobber_all); }
+SEC("xdp") +__description("spill unbounded reg, then range check src") +__success __retval(0) +__naked void spill_unbounded(void) +{ + asm volatile (" \ + /* Produce an unbounded scalar. */ \ + call %[bpf_get_prandom_u32]; \ + /* Spill r0 to stack. */ \ + *(u64*)(r10 - 8) = r0; \ + /* Boundary check on r0. */ \ + if r0 > 16 goto l0_%=; \ + /* Fill r0 from stack. */ \ + r0 = *(u64*)(r10 - 8); \ + /* Boundary check on r0 with predetermined result. */\ + if r0 <= 16 goto l0_%=; \ + /* Dead branch: the verifier should prune it. Do an invalid memory\ + * access if the verifier follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ +l0_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_get_prandom_u32) + : __clobber_all); +} + char _license[] SEC("license") = "GPL";
From: Maxim Mikityanskiy maxim@isovalent.com
When the width of a fill is smaller than the width of the preceding spill, the information about scalar boundaries can still be preserved, as long as it's coerced to the right width (done by coerce_reg_to_size). Even further, if the actual value fits into the fill width, the ID can be preserved as well for further tracking of equal scalars.
Implement the above improvements, which makes narrowing fills behave the same as narrowing spills and MOVs between registers.
Two tests are adjusted to accommodate for endianness differences and to take into account that it's now allowed to do a narrowing fill from the least significant bits.
reg_bounds_sync is added to coerce_reg_to_size to correctly adjust umin/umax boundaries after the var_off truncation, for example, a 64-bit value 0xXXXXXXXX00000000, when read as a 32-bit, gets umin = 0, umax = 0xFFFFFFFF, var_off = (0x0; 0xffffffff00000000), which needs to be synced down to umax = 0, otherwise reg_bounds_sanity_check doesn't pass.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- kernel/bpf/verifier.c | 20 ++++++++++--- .../selftests/bpf/progs/verifier_spill_fill.c | 28 +++++++++++++------ 2 files changed, 35 insertions(+), 13 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 9b5053389739..b6e252539e52 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4772,7 +4772,13 @@ static int check_stack_read_fixed_off(struct bpf_verifier_env *env, if (dst_regno < 0) return 0;
- if (!(off % BPF_REG_SIZE) && size == spill_size) { + if (size <= spill_size && +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + !(off % BPF_REG_SIZE) +#else + !((off + size - spill_size) % BPF_REG_SIZE) +#endif + ) { /* The earlier check_reg_arg() has decided the * subreg_def for this insn. Save it first. */ @@ -4780,6 +4786,12 @@ static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
copy_register_state(&state->regs[dst_regno], reg); state->regs[dst_regno].subreg_def = subreg_def; + + /* Break the relation on a narrowing fill. + * coerce_reg_to_size will adjust the boundaries. + */ + if (get_reg_width(reg) > size * BITS_PER_BYTE) + state->regs[dst_regno].id = 0; } else { int spill_cnt = 0, zero_cnt = 0;
@@ -6055,10 +6067,10 @@ static void coerce_reg_to_size(struct bpf_reg_state *reg, int size) * values are also truncated so we push 64-bit bounds into * 32-bit bounds. Above were truncated < 32-bits already. */ - if (size < 4) { + if (size < 4) __mark_reg32_unbounded(reg); - reg_bounds_sync(reg); - } + + reg_bounds_sync(reg); }
static void set_sext64_default_val(struct bpf_reg_state *reg, int size) diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index 809a09732168..de03e72e07a9 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -217,7 +217,7 @@ __naked void uninit_u32_from_the_stack(void)
SEC("tc") __description("Spill a u32 const scalar. Refill as u16. Offset to skb->data") -__failure __msg("invalid access to packet") +__success __retval(0) __naked void u16_offset_to_skb_data(void) { asm volatile (" \ @@ -225,19 +225,24 @@ __naked void u16_offset_to_skb_data(void) r3 = *(u32*)(r1 + %[__sk_buff_data_end]); \ w4 = 20; \ *(u32*)(r10 - 8) = r4; \ - r4 = *(u16*)(r10 - 8); \ + r4 = *(u16*)(r10 - %[offset]); \ r0 = r2; \ - /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=umax=65535 */\ + /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=20 */\ r0 += r4; \ - /* if (r0 > r3) R0=pkt,umax=65535 R2=pkt R3=pkt_end R4=umax=65535 */\ + /* if (r0 > r3) R0=pkt,off=20 R2=pkt R3=pkt_end R4=20 */\ if r0 > r3 goto l0_%=; \ - /* r0 = *(u32 *)r2 R0=pkt,umax=65535 R2=pkt R3=pkt_end R4=20 */\ + /* r0 = *(u32 *)r2 R0=pkt,off=20 R2=pkt R3=pkt_end R4=20 */\ r0 = *(u32*)(r2 + 0); \ l0_%=: r0 = 0; \ exit; \ " : : __imm_const(__sk_buff_data, offsetof(struct __sk_buff, data)), - __imm_const(__sk_buff_data_end, offsetof(struct __sk_buff, data_end)) + __imm_const(__sk_buff_data_end, offsetof(struct __sk_buff, data_end)), +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + __imm_const(offset, 8) +#else + __imm_const(offset, 6) +#endif : __clobber_all); }
@@ -270,7 +275,7 @@ l0_%=: r0 = 0; \ }
SEC("tc") -__description("Spill a u32 const scalar. Refill as u16 from fp-6. Offset to skb->data") +__description("Spill a u32 const scalar. Refill as u16 from MSB. Offset to skb->data") __failure __msg("invalid access to packet") __naked void _6_offset_to_skb_data(void) { @@ -279,7 +284,7 @@ __naked void _6_offset_to_skb_data(void) r3 = *(u32*)(r1 + %[__sk_buff_data_end]); \ w4 = 20; \ *(u32*)(r10 - 8) = r4; \ - r4 = *(u16*)(r10 - 6); \ + r4 = *(u16*)(r10 - %[offset]); \ r0 = r2; \ /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=umax=65535 */\ r0 += r4; \ @@ -291,7 +296,12 @@ l0_%=: r0 = 0; \ exit; \ " : : __imm_const(__sk_buff_data, offsetof(struct __sk_buff, data)), - __imm_const(__sk_buff_data_end, offsetof(struct __sk_buff, data_end)) + __imm_const(__sk_buff_data_end, offsetof(struct __sk_buff, data_end)), +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + __imm_const(offset, 6) +#else + __imm_const(offset, 8) +#endif : __clobber_all); }
On Wed, Dec 20, 2023 at 11:40:10PM +0200, Maxim Mikityanskiy wrote:
When the width of a fill is smaller than the width of the preceding spill, the information about scalar boundaries can still be preserved, as long as it's coerced to the right width (done by coerce_reg_to_size). Even further, if the actual value fits into the fill width, the ID can be preserved as well for further tracking of equal scalars.
Implement the above improvements, which makes narrowing fills behave the same as narrowing spills and MOVs between registers.
Two tests are adjusted to accommodate for endianness differences and to take into account that it's now allowed to do a narrowing fill from the least significant bits.
reg_bounds_sync is added to coerce_reg_to_size to correctly adjust umin/umax boundaries after the var_off truncation, for example, a 64-bit value 0xXXXXXXXX00000000, when read as a 32-bit, gets umin = 0, umax = 0xFFFFFFFF, var_off = (0x0; 0xffffffff00000000), which needs to be synced down to umax = 0, otherwise reg_bounds_sanity_check doesn't pass.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com
kernel/bpf/verifier.c | 20 ++++++++++--- .../selftests/bpf/progs/verifier_spill_fill.c | 28 +++++++++++++------ 2 files changed, 35 insertions(+), 13 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 9b5053389739..b6e252539e52 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4772,7 +4772,13 @@ static int check_stack_read_fixed_off(struct bpf_verifier_env *env, if (dst_regno < 0) return 0;
if (!(off % BPF_REG_SIZE) && size == spill_size) {
if (size <= spill_size &&+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
!(off % BPF_REG_SIZE)+#else
!((off + size - spill_size) % BPF_REG_SIZE)+#endif
If I understand correctly, it is preferred to keep endianess checking macro out of verfier.c and have helper function handle them instead.
E.g. See bpf_ctx_narrow_access_offset() from include/linux/filter.h
[...]
From: Maxim Mikityanskiy maxim@isovalent.com
The previous commit allowed to preserve boundaries and track IDs of scalars on narrowing fills. Add test cases for that pattern.
Signed-off-by: Maxim Mikityanskiy maxim@isovalent.com --- .../selftests/bpf/progs/verifier_spill_fill.c | 108 ++++++++++++++++++ 1 file changed, 108 insertions(+)
diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index de03e72e07a9..df195cf5c77b 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -938,4 +938,112 @@ l0_%=: r0 = 0; \ : __clobber_all); }
+SEC("xdp") +__description("32-bit fill after 64-bit spill") +__success __retval(0) +__naked void fill_32bit_after_spill_64bit(void) +{ + asm volatile(" \ + /* Randomize the upper 32 bits. */ \ + call %[bpf_get_prandom_u32]; \ + r0 <<= 32; \ + /* 64-bit spill r0 to stack. */ \ + *(u64*)(r10 - 8) = r0; \ + /* 32-bit fill r0 from stack. */ \ + r0 = *(u32*)(r10 - %[offset]); \ + /* Boundary check on r0 with predetermined result. */\ + if r0 == 0 goto l0_%=; \ + /* Dead branch: the verifier should prune it. Do an invalid memory\ + * access if the verifier follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ +l0_%=: exit; \ +" : + : __imm(bpf_get_prandom_u32), +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + __imm_const(offset, 8) +#else + __imm_const(offset, 4) +#endif + : __clobber_all); +} + +SEC("xdp") +__description("32-bit fill after 64-bit spill of 32-bit value should preserve ID") +__success __retval(0) +__naked void fill_32bit_after_spill_64bit_preserve_id(void) +{ + asm volatile (" \ + /* Randomize the lower 32 bits. */ \ + call %[bpf_get_prandom_u32]; \ + w0 &= 0xffffffff; \ + /* 64-bit spill r0 to stack - should assign an ID. */\ + *(u64*)(r10 - 8) = r0; \ + /* 32-bit fill r1 from stack - should preserve the ID. */\ + r1 = *(u32*)(r10 - %[offset]); \ + /* Compare r1 with another register to trigger find_equal_scalars. */\ + r2 = 0; \ + if r1 != r2 goto l0_%=; \ + /* The result of this comparison is predefined. */\ + if r0 == r2 goto l0_%=; \ + /* Dead branch: the verifier should prune it. Do an invalid memory\ + * access if the verifier follows it. \ + */ \ + r0 = *(u64*)(r9 + 0); \ + exit; \ +l0_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_get_prandom_u32), +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + __imm_const(offset, 8) +#else + __imm_const(offset, 4) +#endif + : __clobber_all); +} + +SEC("xdp") +__description("32-bit fill after 64-bit spill should clear ID") +__failure __msg("math between ctx pointer and 4294967295 is not allowed") +__naked void fill_32bit_after_spill_64bit_clear_id(void) +{ + asm volatile (" \ + r6 = r1; \ + /* Roll one bit to force the verifier to track both branches. */\ + call %[bpf_get_prandom_u32]; \ + r0 &= 0x8; \ + /* Put a large number into r1. */ \ + r1 = 0xffffffff; \ + r1 <<= 32; \ + r1 += r0; \ + /* 64-bit spill r1 to stack - should assign an ID. */\ + *(u64*)(r10 - 8) = r1; \ + /* 32-bit fill r2 from stack - should clear the ID. */\ + r2 = *(u32*)(r10 - %[offset]); \ + /* Compare r2 with another register to trigger find_equal_scalars.\ + * Having one random bit is important here, otherwise the verifier cuts\ + * the corners. If the ID was mistakenly preserved on fill, this would\ + * cause the verifier to think that r1 is also equal to zero in one of\ + * the branches, and equal to eight on the other branch.\ + */ \ + r3 = 0; \ + if r2 != r3 goto l0_%=; \ +l0_%=: r1 >>= 32; \ + /* The verifier shouldn't propagate r2's range to r1, so it should\ + * still remember r1 = 0xffffffff and reject the below.\ + */ \ + r6 += r1; \ + r0 = *(u32*)(r6 + 0); \ + exit; \ +" : + : __imm(bpf_get_prandom_u32), +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + __imm_const(offset, 8) +#else + __imm_const(offset, 4) +#endif + : __clobber_all); +} + char _license[] SEC("license") = "GPL";
From: Eduard Zingerman eddyz87@gmail.com
Changes for scalar ID tracking of spilled unbound scalars lead to certain verification performance regression. This commit mitigates the regression by exploiting the following properties maintained by check_stack_read_fixed_off(): - a mix of STACK_MISC, STACK_ZERO and STACK_INVALID marks is read as unbounded scalar register; - spi with all slots marked STACK_ZERO is read as scalar register with value zero.
This commit modifies stacksafe() to consider situations above equivalent.
Veristat results after this patch show significant gains:
$ ./veristat -e file,prog,states -f '!states_pct<10' -f '!states_b<10' -C not-opt after File Program States (A) States (B) States (DIFF) ---------------- -------- ---------- ---------- ---------------- pyperf180.bpf.o on_event 10456 8422 -2034 (-19.45%) pyperf600.bpf.o on_event 37319 22519 -14800 (-39.66%) strobemeta.bpf.o on_event 13435 4703 -8732 (-64.99%)
Signed-off-by: Eduard Zingerman eddyz87@gmail.com --- kernel/bpf/verifier.c | 83 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 83 insertions(+)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index b6e252539e52..a020d4d83524 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -1168,6 +1168,12 @@ static void mark_stack_slot_misc(struct bpf_verifier_env *env, u8 *stype) *stype = STACK_MISC; }
+static bool is_spilled_scalar_reg64(const struct bpf_stack_state *stack) +{ + return stack->slot_type[0] == STACK_SPILL && + stack->spilled_ptr.type == SCALAR_VALUE; +} + static void scrub_spilled_slot(u8 *stype) { if (*stype != STACK_INVALID) @@ -16449,11 +16455,45 @@ static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold, } }
+static bool is_stack_zero64(struct bpf_stack_state *stack) +{ + u32 i; + + for (i = 0; i < ARRAY_SIZE(stack->slot_type); ++i) + if (stack->slot_type[i] != STACK_ZERO) + return false; + return true; +} + +static bool is_stack_unbound_slot64(struct bpf_verifier_env *env, + struct bpf_stack_state *stack) +{ + u32 i; + + for (i = 0; i < ARRAY_SIZE(stack->slot_type); ++i) + if (stack->slot_type[i] != STACK_ZERO && + stack->slot_type[i] != STACK_MISC && + (!env->allow_uninit_stack || stack->slot_type[i] != STACK_INVALID)) + return false; + return true; +} + +static bool is_spilled_unbound_scalar_reg64(struct bpf_stack_state *stack) +{ + return is_spilled_scalar_reg64(stack) && __is_scalar_unbounded(&stack->spilled_ptr); +} + static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old, struct bpf_func_state *cur, struct bpf_idmap *idmap, bool exact) { + struct bpf_reg_state unbound_reg = {}; + struct bpf_reg_state zero_reg = {}; int i, spi;
+ __mark_reg_unknown(env, &unbound_reg); + __mark_reg_const_zero(env, &zero_reg); + zero_reg.precise = true; + /* walk slots of the explored stack and ignore any additional * slots in the current stack, since explored(safe) state * didn't use them @@ -16474,6 +16514,49 @@ static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old, continue; }
+ /* load of stack value with all MISC and ZERO slots produces unbounded + * scalar value, call regsafe to ensure scalar ids are compared. + */ + if (is_spilled_unbound_scalar_reg64(&old->stack[spi]) && + is_stack_unbound_slot64(env, &cur->stack[spi])) { + i += BPF_REG_SIZE - 1; + if (!regsafe(env, &old->stack[spi].spilled_ptr, &unbound_reg, + idmap, exact)) + return false; + continue; + } + + if (is_stack_unbound_slot64(env, &old->stack[spi]) && + is_spilled_unbound_scalar_reg64(&cur->stack[spi])) { + i += BPF_REG_SIZE - 1; + if (!regsafe(env, &unbound_reg, &cur->stack[spi].spilled_ptr, + idmap, exact)) + return false; + continue; + } + + /* load of stack value with all ZERO slots produces scalar value 0, + * call regsafe to ensure scalar ids are compared and precision + * flags are taken into account. + */ + if (is_spilled_scalar_reg64(&old->stack[spi]) && + is_stack_zero64(&cur->stack[spi])) { + if (!regsafe(env, &old->stack[spi].spilled_ptr, &zero_reg, + idmap, exact)) + return false; + i += BPF_REG_SIZE - 1; + continue; + } + + if (is_stack_zero64(&old->stack[spi]) && + is_spilled_scalar_reg64(&cur->stack[spi])) { + if (!regsafe(env, &zero_reg, &cur->stack[spi].spilled_ptr, + idmap, exact)) + return false; + i += BPF_REG_SIZE - 1; + continue; + } + if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID) continue;
From: Eduard Zingerman eddyz87@gmail.com
Check that stacksafe() considers the following old vs cur stack spill state combinations equivalent: - spill of unbound scalar vs combination of STACK_{MISC,ZERO,INVALID} - STACK_MISC vs spill of unbound scalar - spill of scalar 0 vs STACK_ZERO - STACK_ZERO vs spill of scalar 0
Signed-off-by: Eduard Zingerman eddyz87@gmail.com --- .../selftests/bpf/progs/verifier_spill_fill.c | 192 ++++++++++++++++++ 1 file changed, 192 insertions(+)
diff --git a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c index df195cf5c77b..e2acc4fc3d10 100644 --- a/tools/testing/selftests/bpf/progs/verifier_spill_fill.c +++ b/tools/testing/selftests/bpf/progs/verifier_spill_fill.c @@ -1046,4 +1046,196 @@ l0_%=: r1 >>= 32; \ : __clobber_all); }
+/* stacksafe(): check if spill of unbound scalar in old state is + * considered equivalent to any state of the spill in the current state. + * + * On the first verification path an unbound scalar is written for + * fp-8 and later marked precise. + * On the second verification path a mix of STACK_MISC/ZERO/INVALID is + * written to fp-8. These should be considered equivalent. + */ +SEC("socket") +__success __log_level(2) +__msg("10: (79) r0 = *(u64 *)(r10 -8)") +__msg("10: safe") +__msg("processed 16 insns") +__flag(BPF_F_TEST_STATE_FREQ) +__naked void old_unbound_scalar_vs_cur_anything(void) +{ + asm volatile( + /* get a random value for branching */ + "call %[bpf_ktime_get_ns];" + "r7 = r0;" + /* get a random value for storing at fp-8 */ + "call %[bpf_ktime_get_ns];" + "if r7 == 0 goto 1f;" + /* unbound scalar written to fp-8 */ + "*(u64*)(r10 - 8) = r0;" + "goto 2f;" +"1:" + /* mark fp-8 as mix of STACK_MISC/ZERO/INVALID */ + "r1 = 0;" + "*(u8*)(r10 - 8) = r0;" + "*(u8*)(r10 - 7) = r1;" + /* fp-2..fp-6 remain STACK_INVALID */ + "*(u8*)(r10 - 1) = r0;" +"2:" + /* read fp-8 and force it precise, should be considered safe + * on second visit + */ + "r0 = *(u64*)(r10 - 8);" + "r0 &= 0xff;" + "r1 = r10;" + "r1 += r0;" + "exit;" + : + : __imm(bpf_ktime_get_ns) + : __clobber_all); +} + +/* stacksafe(): check if STACK_MISC in old state is considered + * equivalent to stack spill of unbound scalar in cur state. + */ +SEC("socket") +__success __log_level(2) +__msg("8: (79) r0 = *(u64 *)(r10 -8) ; R0_w=scalar(id=1) R10=fp0 fp-8=scalar(id=1)") +__msg("8: safe") +__msg("processed 11 insns") +__flag(BPF_F_TEST_STATE_FREQ) +__naked void old_unbound_scalar_vs_cur_stack_misc(void) +{ + asm volatile( + /* get a random value for branching */ + "call %[bpf_ktime_get_ns];" + "if r0 == 0 goto 1f;" + /* conjure unbound scalar at fp-8 */ + "call %[bpf_ktime_get_ns];" + "*(u64*)(r10 - 8) = r0;" + "goto 2f;" +"1:" + /* conjure STACK_MISC at fp-8 */ + "call %[bpf_ktime_get_ns];" + "*(u64*)(r10 - 8) = r0;" + "*(u32*)(r10 - 4) = r0;" +"2:" + /* read fp-8, should be considered safe on second visit */ + "r0 = *(u64*)(r10 - 8);" + "exit;" + : + : __imm(bpf_ktime_get_ns) + : __clobber_all); +} + +/* stacksafe(): check if stack spill of unbound scalar in old state is + * considered equivalent to STACK_MISC in cur state. + */ +SEC("socket") +__success __log_level(2) +__msg("8: (79) r0 = *(u64 *)(r10 -8) ; R0_w=scalar() R10=fp0 fp-8=mmmmmmmm") +__msg("8: safe") +__msg("processed 11 insns") +__flag(BPF_F_TEST_STATE_FREQ) +__naked void old_stack_misc_vs_cur_unbound_scalar(void) +{ + asm volatile( + /* get a random value for branching */ + "call %[bpf_ktime_get_ns];" + "if r0 == 0 goto 1f;" + /* conjure STACK_MISC at fp-8 */ + "call %[bpf_ktime_get_ns];" + "*(u64*)(r10 - 8) = r0;" + "*(u32*)(r10 - 4) = r0;" + "goto 2f;" +"1:" + /* conjure unbound scalar at fp-8 */ + "call %[bpf_ktime_get_ns];" + "*(u64*)(r10 - 8) = r0;" +"2:" + /* read fp-8, should be considered safe on second visit */ + "r0 = *(u64*)(r10 - 8);" + "exit;" + : + : __imm(bpf_ktime_get_ns) + : __clobber_all); +} + +/* stacksafe(): check if spill of register with value 0 in old state + * is considered equivalent to STACK_ZERO. + */ +SEC("socket") +__success __log_level(2) +__msg("9: (79) r0 = *(u64 *)(r10 -8)") +__msg("9: safe") +__msg("processed 15 insns") +__flag(BPF_F_TEST_STATE_FREQ) +__naked void old_spill_zero_vs_stack_zero(void) +{ + asm volatile( + /* get a random value for branching */ + "call %[bpf_ktime_get_ns];" + "r7 = r0;" + /* get a random value for storing at fp-8 */ + "call %[bpf_ktime_get_ns];" + "if r7 == 0 goto 1f;" + /* conjure spilled register with value 0 at fp-8 */ + "*(u64*)(r10 - 8) = r0;" + "if r0 != 0 goto 3f;" + "goto 2f;" +"1:" + /* conjure STACK_ZERO at fp-8 */ + "r1 = 0;" + "*(u64*)(r10 - 8) = r1;" +"2:" + /* read fp-8 and force it precise, should be considered safe + * on second visit + */ + "r0 = *(u64*)(r10 - 8);" + "r1 = r10;" + "r1 += r0;" +"3:" + "exit;" + : + : __imm(bpf_ktime_get_ns) + : __clobber_all); +} + +/* stacksafe(): similar to old_spill_zero_vs_stack_zero() but the + * other way around: check if STACK_ZERO is considered equivalent to + * spill of register with value 0. + */ +SEC("socket") +__success __log_level(2) +__msg("8: (79) r0 = *(u64 *)(r10 -8)") +__msg("8: safe") +__msg("processed 14 insns") +__flag(BPF_F_TEST_STATE_FREQ) +__naked void old_stack_zero_vs_spill_zero(void) +{ + asm volatile( + /* get a random value for branching */ + "call %[bpf_ktime_get_ns];" + "if r0 == 0 goto 1f;" + /* conjure STACK_ZERO at fp-8 */ + "r1 = 0;" + "*(u64*)(r10 - 8) = r1;" + "goto 2f;" +"1:" + /* conjure spilled register with value 0 at fp-8 */ + "call %[bpf_ktime_get_ns];" + "*(u64*)(r10 - 8) = r0;" + "if r0 != 0 goto 3f;" +"2:" + /* read fp-8 and force it precise, should be considered safe + * on second visit + */ + "r0 = *(u64*)(r10 - 8);" + "r1 = r10;" + "r1 += r0;" +"3:" + "exit;" + : + : __imm(bpf_ktime_get_ns) + : __clobber_all); +} + char _license[] SEC("license") = "GPL";
linux-kselftest-mirror@lists.linaro.org
-
Alexei Starovoitov -
Maxim Mikityanskiy -
Shung-Hsi Yu