[PATCH bpf-next 0/3] bpf: arg tracking for imprecise/multi-offset BPF_ST/STX

[PATCH bpf-next 0/3] bpf: arg tracking for imprecise/multi-offset BPF_ST/STX

[Eduard Zingerman]

When the static arg tracking analysis encounters a store through a
pointer with imprecise or multi-offset destination, it must use weak
updates (join) instead of strong updates (overwrite) for the affected
at_stack slots. At runtime only one slot is actually written; the
others retain their old values.

Two cases are addressed:
- BPF_STX, handled by spill_to_stack(). It was gated on
  `dst_is_local_fp = (frame == depth)`, which missed ARG_IMPRECISE
  pointers entirely.
- BPF_ST, handled by clear_stack_for_all_offs(). It delegates to
  clear_overlapping_stack_slots() which unconditionally set
  `at_stack[i] = none`. Change to `at_stack[i] = join(old, none)`
  when multiple candidate slots exist (cnt != 1), so that untouched
  slots preserve their tracked values.

No veristat diff compared to current master when tested on selftests,
sched_ext, cilium and a set of Meta internal programs.

This addresses issues reported by sashiko for patch #7 in [1].

[1] https://sashiko.dev/#/patchset/20260410-patch-set-v4-0-5d4eecb343db%40gmail.com

---
Eduard Zingerman (3):
      bpf: fix arg tracking for BPF_STX through imprecise ptrs
      bpf: fix arg tracking for BPF_ST through imprecise/multi-offset ptrs
      selftests/bpf: arg tracking for imprecise/multi-offset BPF_ST/STX

 kernel/bpf/liveness.c                              |  67 ++++---
 .../selftests/bpf/progs/verifier_live_stack.c      | 194 +++++++++++++++++++++
 2 files changed, 232 insertions(+), 29 deletions(-)
---
base-commit: 71b500afd2f7336f5b6c6026f2af546fc079be26
change-id: 20260413-stacklive-fixes-42e258cf0397

[PATCH bpf-next 1/3] bpf: fix arg tracking for BPF_STX through imprecise ptrs

[Eduard Zingerman]

BPF_STX through ARG_IMPRECISE dst should be recognized as a local
spill and join at_stack with the written value. For example,
consider the following situation:

   // r1 = ARG_IMPRECISE{mask=BIT(0)|BIT(1)}
   *(u64 *)(r1 + 0) = r8

Here the analysis should produce an equivalent of

  at_stack[*] = join(old, r8)

Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
---
 kernel/bpf/liveness.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/kernel/bpf/liveness.c b/kernel/bpf/liveness.c
index 1fb4c511db5a..8069ea955023 100644
--- a/kernel/bpf/liveness.c
+++ b/kernel/bpf/liveness.c
@@ -1042,6 +1042,12 @@ static void arg_track_log(struct bpf_verifier_env *env, struct bpf_insn *insn, i
 		verbose(env, "\n");
 }
 
+static bool can_be_local_fp(int depth, int regno, struct arg_track *at)
+{
+	return regno == BPF_REG_FP || at->frame == depth ||
+	       (at->frame == ARG_IMPRECISE && (at->mask & BIT(depth)));
+}
+
 /*
  * Pure dataflow transfer function for arg_track state.
  * Updates at_out[] based on how the instruction modifies registers.
@@ -1111,8 +1117,7 @@ static void arg_track_xfer(struct bpf_verifier_env *env, struct bpf_insn *insn,
 			at_out[r] = none;
 	} else if (class == BPF_LDX) {
 		u32 sz = bpf_size_to_bytes(BPF_SIZE(insn->code));
-		bool src_is_local_fp = insn->src_reg == BPF_REG_FP || src->frame == depth ||
-				       (src->frame == ARG_IMPRECISE && (src->mask & BIT(depth)));
+		bool src_is_local_fp = can_be_local_fp(depth, insn->src_reg, src);
 
 		/*
 		 * Reload from callee stack: if src is current-frame FP-derived
@@ -1147,7 +1152,7 @@ static void arg_track_xfer(struct bpf_verifier_env *env, struct bpf_insn *insn,
 		bool dst_is_local_fp;
 
 		/* Track spills to current-frame FP-derived callee stack */
-		dst_is_local_fp = insn->dst_reg == BPF_REG_FP || dst->frame == depth;
+		dst_is_local_fp = can_be_local_fp(depth, insn->dst_reg, dst);
 		if (dst_is_local_fp && BPF_MODE(insn->code) == BPF_MEM)
 			spill_to_stack(insn, at_out, insn->dst_reg,
 				       at_stack_out, src, sz);

-- 
2.53.0

[PATCH bpf-next 2/3] bpf: fix arg tracking for BPF_ST through imprecise/multi-offset ptrs

[Eduard Zingerman]

BPF_ST through multi-offset or imprecise dst should join at_stack with
none instead of overwriting the slots. For example, consider the
following situation:

   // r1 = ARG_IMPRECISE{mask=BIT(0)|BIT(1)}
   *(u64 *)(r1 + 0) = 0

Here the analysis should produce an equivalent of

  at_stack[*r1] = join(old, none).

Move the definition clear_overlapping_stack_slots() in order to have
__arg_track_join() visible.

Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
---
 kernel/bpf/liveness.c | 56 +++++++++++++++++++++++++++------------------------
 1 file changed, 30 insertions(+), 26 deletions(-)

diff --git a/kernel/bpf/liveness.c b/kernel/bpf/liveness.c
index 8069ea955023..d5e0bd18ee06 100644
--- a/kernel/bpf/liveness.c
+++ b/kernel/bpf/liveness.c
@@ -622,28 +622,6 @@ static bool arg_is_fp(const struct arg_track *at)
 	return at->frame >= 0 || at->frame == ARG_IMPRECISE;
 }
 
-/*
- * Clear all tracked callee stack slots overlapping the byte range
- * [off, off+sz-1] where off is a negative FP-relative offset.
- */
-static void clear_overlapping_stack_slots(struct arg_track *at_stack, s16 off, u32 sz)
-{
-	struct arg_track none = { .frame = ARG_NONE };
-
-	if (off == OFF_IMPRECISE) {
-		for (int i = 0; i < MAX_ARG_SPILL_SLOTS; i++)
-			at_stack[i] = none;
-		return;
-	}
-	for (int i = 0; i < MAX_ARG_SPILL_SLOTS; i++) {
-		int slot_start = -((i + 1) * 8);
-		int slot_end = slot_start + 8;
-
-		if (slot_start < off + (int)sz && slot_end > off)
-			at_stack[i] = none;
-	}
-}
-
 static void verbose_arg_track(struct bpf_verifier_env *env, struct arg_track *at)
 {
 	int i;
@@ -980,6 +958,32 @@ static void spill_to_stack(struct bpf_insn *insn, struct arg_track *at_out,
 	}
 }
 
+/*
+ * Clear all tracked callee stack slots overlapping the byte range
+ * [off, off+sz-1] where off is a negative FP-relative offset.
+ */
+static void clear_overlapping_stack_slots(struct arg_track *at_stack, s16 off, u32 sz, int cnt)
+{
+	struct arg_track none = { .frame = ARG_NONE };
+
+	if (off == OFF_IMPRECISE) {
+		for (int i = 0; i < MAX_ARG_SPILL_SLOTS; i++)
+			at_stack[i] = __arg_track_join(at_stack[i], none);
+		return;
+	}
+	for (int i = 0; i < MAX_ARG_SPILL_SLOTS; i++) {
+		int slot_start = -((i + 1) * 8);
+		int slot_end = slot_start + 8;
+
+		if (slot_start < off + (int)sz && slot_end > off) {
+			if (cnt == 1)
+				at_stack[i] = none;
+			else
+				at_stack[i] = __arg_track_join(at_stack[i], none);
+		}
+	}
+}
+
 /*
  * Clear stack slots overlapping all possible FP offsets in @reg.
  */
@@ -990,18 +994,18 @@ static void clear_stack_for_all_offs(struct bpf_insn *insn,
 	int cnt, i;
 
 	if (reg == BPF_REG_FP) {
-		clear_overlapping_stack_slots(at_stack_out, insn->off, sz);
+		clear_overlapping_stack_slots(at_stack_out, insn->off, sz, 1);
 		return;
 	}
 	cnt = at_out[reg].off_cnt;
 	if (cnt == 0) {
-		clear_overlapping_stack_slots(at_stack_out, OFF_IMPRECISE, sz);
+		clear_overlapping_stack_slots(at_stack_out, OFF_IMPRECISE, sz, cnt);
 		return;
 	}
 	for (i = 0; i < cnt; i++) {
 		s16 fp_off = arg_add(at_out[reg].off[i], insn->off);
 
-		clear_overlapping_stack_slots(at_stack_out, fp_off, sz);
+		clear_overlapping_stack_slots(at_stack_out, fp_off, sz, cnt);
 	}
 }
 
@@ -1171,7 +1175,7 @@ static void arg_track_xfer(struct bpf_verifier_env *env, struct bpf_insn *insn,
 		}
 	} else if (class == BPF_ST && BPF_MODE(insn->code) == BPF_MEM) {
 		u32 sz = bpf_size_to_bytes(BPF_SIZE(insn->code));
-		bool dst_is_local_fp = insn->dst_reg == BPF_REG_FP || dst->frame == depth;
+		bool dst_is_local_fp = can_be_local_fp(depth, insn->dst_reg, dst);
 
 		/* BPF_ST to FP-derived dst: clear overlapping stack slots */
 		if (dst_is_local_fp)

-- 
2.53.0

[PATCH bpf-next 3/3] selftests/bpf: arg tracking for imprecise/multi-offset BPF_ST/STX

[Eduard Zingerman]

Add test cases for clear_stack_for_all_offs and dst_is_local_fp
handling of multi-offset and ARG_IMPRECISE stack pointers:

- st_imm_join_with_multi_off: BPF_ST through multi-offset dst should
  join at_stack with none instead of overwriting both candidate slots.
- st_imm_join_with_imprecise_off: BPF_ST through offset-imprecise dst
  should join at_stack with none instead of clearing all slots.
- st_imm_join_with_single_off: a canary checking that BPF_ST with a
  known offset overwrites slot instead of joining.
- imprecise_dst_spill_join: BPF_STX through ARG_IMPRECISE dst should
  be recognized as a local spill and join at_stack with the written
  value.

Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
---
 .../selftests/bpf/progs/verifier_live_stack.c      | 194 +++++++++++++++++++++
 1 file changed, 194 insertions(+)

diff --git a/tools/testing/selftests/bpf/progs/verifier_live_stack.c b/tools/testing/selftests/bpf/progs/verifier_live_stack.c
index b7a9fa10e84d..536e214d0376 100644
--- a/tools/testing/selftests/bpf/progs/verifier_live_stack.c
+++ b/tools/testing/selftests/bpf/progs/verifier_live_stack.c
@@ -2647,3 +2647,197 @@ __naked void spill_join_with_imprecise_off(void)
 	"exit;"
 	::: __clobber_all);
 }
+
+/*
+ * Same as spill_join_with_multi_off but the write is BPF_ST (store
+ * immediate) instead of BPF_STX. BPF_ST goes through
+ * clear_stack_for_all_offs() rather than spill_to_stack(), and that
+ * path also needs to join instead of overwriting.
+ *
+ *   fp-8  = &fp-24
+ *   fp-16 = &fp-32
+ *   r1 = fp-8 or fp-16 (two offsets from branch)
+ *   *(u64 *)(r1 + 0) = 0        -- BPF_ST with immediate
+ *   r0 = *(u64 *)(r10 - 16)     -- fill from fp-16
+ *   r0 = *(u64 *)(r0 + 0)       -- deref: should produce use
+ */
+SEC("socket")
+__log_level(2)
+__failure
+__msg("15: (7a) *(u64 *)(r1 +0) = 0	fp-8: fp0-24 -> fp0-24|fp0+0	fp-16: fp0-32 -> fp0-32|fp0+0")
+__msg("17: (79) r0 = *(u64 *)(r0 +0)         ; use: fp0-32")
+__naked void st_imm_join_with_multi_off(void)
+{
+	asm volatile (
+	"*(u64 *)(r10 - 24) = 0;"
+	"*(u64 *)(r10 - 32) = 0;"
+	"r1 = r10;"
+	"r1 += -24;"
+	"*(u64 *)(r10 - 8) = r1;"
+	"r1 = r10;"
+	"r1 += -32;"
+	"*(u64 *)(r10 - 16) = r1;"
+	/* create r1 with two candidate offsets: fp-8 or fp-16 */
+	"call %[bpf_get_prandom_u32];"
+	"if r0 == 0 goto 1f;"
+	"r1 = r10;"
+	"r1 += -8;"
+	"goto 2f;"
+"1:"
+	"r1 = r10;"
+	"r1 += -16;"
+"2:"
+	/* BPF_ST: store immediate through multi-offset r1 */
+	"*(u64 *)(r1 + 0) = 0;"
+	/* read back fp-16 and deref */
+	"r0 = *(u64 *)(r10 - 16);"
+	"r0 = *(u64 *)(r0 + 0);"
+	"r0 = 0;"
+	"exit;"
+	:: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
+/*
+ * Check that BPF_ST with a known offset fully overwrites stack slot
+ * from the arg tracking point of view.
+ */
+SEC("socket")
+__log_level(2)
+__success
+__msg("5: (7a) *(u64 *)(r1 +0) = 0	fp-8: fp0-16 -> _{{$}}")
+__naked void st_imm_join_with_single_off(void)
+{
+	asm volatile (
+	"r2 = r10;"
+	"r2 += -16;"
+	"*(u64 *)(r10 - 8) = r2;"
+	"r1 = r10;"
+	"r1 += -8;"
+	"*(u64 *)(r1 + 0) = 0;"
+	"r0 = 0;"
+	"exit;"
+	:: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}
+
+/*
+ * Same as spill_join_with_imprecise_off but the write is BPF_ST.
+ * Use "r2 = -8; r1 += r2" to make arg tracking lose offset
+ * precision while the main verifier keeps r1 as fixed-offset.
+ *
+ *   fp-8  = &fp-24
+ *   fp-16 = &fp-32
+ *   r1 = fp-8 (imprecise to arg tracking)
+ *   *(u64 *)(r1 + 0) = 0        -- BPF_ST with immediate
+ *   r0 = *(u64 *)(r10 - 16)     -- fill from fp-16
+ *   r0 = *(u64 *)(r0 + 0)       -- deref: should produce use
+ */
+SEC("socket")
+__log_level(2)
+__success
+__msg("13: (79) r0 = *(u64 *)(r0 +0)         ; use: fp0-32")
+__naked void st_imm_join_with_imprecise_off(void)
+{
+	asm volatile (
+	"*(u64 *)(r10 - 24) = 0;"
+	"*(u64 *)(r10 - 32) = 0;"
+	"r1 = r10;"
+	"r1 += -24;"
+	"*(u64 *)(r10 - 8) = r1;"
+	"r1 = r10;"
+	"r1 += -32;"
+	"*(u64 *)(r10 - 16) = r1;"
+	/* r1 = fp-8 but arg tracking sees off_cnt == 0 */
+	"r1 = r10;"
+	"r2 = -8;"
+	"r1 += r2;"
+	/* store immediate through imprecise r1 */
+	"*(u64 *)(r1 + 0) = 0;"
+	/* read back fp-16 */
+	"r0 = *(u64 *)(r10 - 16);"
+	/* deref: should produce use */
+	"r0 = *(u64 *)(r0 + 0);"
+	"r0 = 0;"
+	"exit;"
+	::: __clobber_all);
+}
+
+/*
+ * Test that spilling through an ARG_IMPRECISE pointer joins with
+ * existing at_stack values. Subprog receives r1 = fp0-24 and
+ * r2 = map_value, creates an ARG_IMPRECISE pointer by joining caller
+ * and callee FP on two branches.
+ *
+ * Setup: callee spills &fp1-16 to fp1-8 (precise, tracked).
+ * Then writes map_value through ARG_IMPRECISE r1 — on path A
+ * this hits fp1-8, on path B it hits caller stack.
+ * Since spill_to_stack is skipped for ARG_IMPRECISE dst,
+ * fp1-8 tracking isn't joined with none.
+ *
+ * Expected after the imprecise write:
+ * - arg tracking should show fp1-8 = fp1-16|fp1+0 (joined with none)
+ * - read from fp1-8 and deref should produce use for fp1-16
+ * - write through it should NOT produce def for fp1-16
+ */
+SEC("socket")
+__log_level(2)
+__success
+__msg("26: (79) r0 = *(u64 *)(r10 -8) // r1=IMP3 r6=fp0-24 r7=fp1-16 fp-8=fp1-16|fp1+0")
+__naked void imprecise_dst_spill_join(void)
+{
+	asm volatile (
+	"*(u64 *)(r10 - 24) = 0;"
+	/* map lookup for a valid non-FP pointer */
+	"*(u32 *)(r10 - 32) = 0;"
+	"r1 = %[map] ll;"
+	"r2 = r10;"
+	"r2 += -32;"
+	"call %[bpf_map_lookup_elem];"
+	"if r0 == 0 goto 1f;"
+	/* r1 = &caller_fp-24, r2 = map_value */
+	"r1 = r10;"
+	"r1 += -24;"
+	"r2 = r0;"
+	"call imprecise_dst_spill_join_sub;"
+"1:"
+	"r0 = 0;"
+	"exit;"
+	:: __imm_addr(map),
+	   __imm(bpf_map_lookup_elem)
+	: __clobber_all);
+}
+
+static __used __naked void imprecise_dst_spill_join_sub(void)
+{
+	asm volatile (
+	/* r6 = &caller_fp-24 (frame=0), r8 = map_value */
+	"r6 = r1;"
+	"r8 = r2;"
+	/* spill &fp1-16 to fp1-8: at_stack[0] = fp1-16 */
+	"*(u64 *)(r10 - 16) = 0;"
+	"r7 = r10;"
+	"r7 += -16;"
+	"*(u64 *)(r10 - 8) = r7;"
+	/* branch to create ARG_IMPRECISE pointer */
+	"call %[bpf_get_prandom_u32];"
+	/* path B: r1 = caller fp-24 (frame=0) */
+	"r1 = r6;"
+	"if r0 == 0 goto 1f;"
+	/* path A: r1 = callee fp-8 (frame=1) */
+	"r1 = r10;"
+	"r1 += -8;"
+"1:"
+	/* r1 = ARG_IMPRECISE{mask=BIT(0)|BIT(1)}.
+	 * Write map_value (non-FP) through r1. On path A this overwrites fp1-8.
+	 * Should join at_stack[0] with none: fp1-16|fp1+0.
+	 */
+	"*(u64 *)(r1 + 0) = r8;"
+	/* read fp1-8: should be fp1-16|fp1+0 (joined) */
+	"r0 = *(u64 *)(r10 - 8);"
+	"*(u64 *)(r0 + 0) = 42;"
+	"r0 = 0;"
+	"exit;"
+	:: __imm(bpf_get_prandom_u32)
+	: __clobber_all);
+}

-- 
2.53.0