selftests/bpf: Test packet range of pointers sharing an idgrafted

Add tests where two packet pointers share an id and tightening one
pointer's umax from its var_off would put it less than their constant
distance from the other's umax: with an index & 0x38 capped at 50, the
base pointer keeps umax 50, so the pointer 8 bytes further on must keep
umax 58, even though its known bits allow at most 56.

These refused a valid program or accepted an out-of-bounds access before
the fix:

- check the advanced copy, load through the base: valid, was refused;
- check the base, load the byte at base + 1 through a copy advanced by
  8: was accepted;
- check base + 4, load 4 bytes at base + 2 through base + 8: reads two
  bytes past the checked range, was accepted;
- the same as the second with data_meta pointers checked against data:
  was accepted.

These pass with and without the fix and cover nearby paths:

- subtract an unknown scalar from a checked pointer and load below it
  (the range is kept across a new id);
- reach a load through two paths whose checks cover 8 and 7 bytes after
  the loaded pointer; the second path must not be pruned by the first;
- spill a copy of a pointer, check the pointer, fill the copy and load
  one byte past the checked range: the load is refused, and the copy
  has the range of the check.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20261001145255.855630-2-alexei.starovoitov@gmail.com
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>

1 files changed, 88 insertions, 0 deletions

diff --git a/lib/math/gcd.c b/lib/math/gcd.c
new file mode 100644
index 000000000..62efca678
--- /dev/null
+++ b/lib/math/gcd.c
@@ -0,0 +1,88 @@
+// SPDX-License-Identifier: GPL-2.0-only
+#include <linux/kernel.h>
+#include <linux/gcd.h>
+#include <linux/export.h>
+
+/*
+ * This implements the binary GCD algorithm. (Often attributed to Stein,
+ * but as Knuth has noted, appears in a first-century Chinese math text.)
+ *
+ * This is faster than the division-based algorithm even on x86, which
+ * has decent hardware division.
+ */
+
+DEFINE_STATIC_KEY_TRUE(efficient_ffs_key);
+
+#if !defined(CONFIG_CPU_NO_EFFICIENT_FFS)
+
+/* If __ffs is available, the even/odd algorithm benchmarks slower. */
+
+static unsigned long binary_gcd(unsigned long a, unsigned long b)
+{
+	unsigned long r = a | b;
+
+	b >>= __ffs(b);
+	if (b == 1)
+		return r & -r;
+
+	for (;;) {
+		a >>= __ffs(a);
+		if (a == 1)
+			return r & -r;
+		if (a == b)
+			return a << __ffs(r);
+
+		if (a < b)
+			swap(a, b);
+		a -= b;
+	}
+}
+
+#endif
+
+/* If normalization is done by loops, the even/odd algorithm is a win. */
+
+/**
+ * gcd - calculate and return the greatest common divisor of 2 unsigned longs
+ * @a: first value
+ * @b: second value
+ */
+unsigned long gcd(unsigned long a, unsigned long b)
+{
+	unsigned long r = a | b;
+
+	if (!a || !b)
+		return r;
+
+#if !defined(CONFIG_CPU_NO_EFFICIENT_FFS)
+	if (static_branch_likely(&efficient_ffs_key))
+		return binary_gcd(a, b);
+#endif
+
+	/* Isolate lsbit of r */
+	r &= -r;
+
+	while (!(b & r))
+		b >>= 1;
+	if (b == r)
+		return r;
+
+	for (;;) {
+		while (!(a & r))
+			a >>= 1;
+		if (a == r)
+			return r;
+		if (a == b)
+			return a;
+
+		if (a < b)
+			swap(a, b);
+		a -= b;
+		a >>= 1;
+		if (a & r)
+			a += b;
+		a >>= 1;
+	}
+}
+
+EXPORT_SYMBOL_GPL(gcd);