bpf: Preserve boundaries and track scalars on narrowing fill
authorMaxim Mikityanskiy <maxim@isovalent.com>
Sat, 27 Jan 2024 17:52:34 +0000 (19:52 +0200)
committerAndrii Nakryiko <andrii@kernel.org>
Fri, 2 Feb 2024 21:22:14 +0000 (13:22 -0800)
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>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20240127175237.526726-4-maxtram95@gmail.com
include/linux/bpf_verifier.h
kernel/bpf/verifier.c
tools/testing/selftests/bpf/progs/verifier_spill_fill.c

index 0dcde339dc7e2ff55d3d74532b7d42c0595ff16e..84365e6dd85d5f1ce3303c25c5f27bb58a775fd4 100644 (file)
@@ -919,6 +919,15 @@ static inline void mark_verifier_state_scratched(struct bpf_verifier_env *env)
        env->scratched_stack_slots = ~0ULL;
 }
 
+static inline bool bpf_stack_narrow_access_ok(int off, int fill_size, int spill_size)
+{
+#ifdef __BIG_ENDIAN
+       off -= spill_size - fill_size;
+#endif
+
+       return !(off % BPF_REG_SIZE);
+}
+
 const char *reg_type_str(struct bpf_verifier_env *env, enum bpf_reg_type type);
 const char *dynptr_type_str(enum bpf_dynptr_type type);
 const char *iter_type_str(const struct btf *btf, u32 btf_id);
index 28f62f24da7e9b0f9735d31a5a3f404a634c64e6..82af971926ac09433c465a747099af15a086e0f3 100644 (file)
@@ -4778,7 +4778,8 @@ 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 &&
+                           bpf_stack_narrow_access_ok(off, size, spill_size)) {
                                /* The earlier check_reg_arg() has decided the
                                 * subreg_def for this insn.  Save it first.
                                 */
@@ -4786,6 +4787,12 @@ 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;
 
@@ -6061,10 +6068,10 @@ 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)
index f9803005e1c0b6300999ebeab2332159c4a49653..3e5d063ea7e8239ebf08c1a686607576517e3f77 100644 (file)
@@ -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 ("                                 \
@@ -225,13 +225,19 @@ __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);                          \
+       "
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+       "r4 = *(u16*)(r10 - 8);"
+#else
+       "r4 = *(u16*)(r10 - 6);"
+#endif
+       "                                               \
        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;                                           \
@@ -268,7 +274,7 @@ 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)
 {
@@ -277,7 +283,13 @@ __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);                          \
+       "
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+       "r4 = *(u16*)(r10 - 6);"
+#else
+       "r4 = *(u16*)(r10 - 8);"
+#endif
+       "                                               \
        r0 = r2;                                        \
        /* r0 += r4 R0=pkt R2=pkt R3=pkt_end R4=umax=65535 */\
        r0 += r4;                                       \