[PATCH] futex: Canonicalize arm64 tagged user addresses

[PATCH] futex: Canonicalize arm64 tagged user addresses

[Shubhang Kaushik (Ampere)]

On arm64, a thread that enables the tagged address ABI can pass valid
top-byte-tagged user pointers to futex syscalls. Such a pointer must
have the same syscall behaviour as its untagged form.

get_futex_key() validates the user address with access_ok(), but uses
the raw address when constructing private and shared-anonymous futex
keys. A waiter using a tagged address and a waker using the untagged
form can therefore access the same futex word while hashing to distinct
keys, causing FUTEX_WAKE to miss the waiter.

Keep access_ok() on the original pointer so that arm64 retains its
per-thread tagged-address ABI enforcement. Canonicalize the address
afterwards, before it is used for futex key construction and shared
mapping lookup.

Add an arm64 64-bit selftest for a tagged private waiter woken through
the corresponding untagged address.

Fixes: 63f0c6037965 ("arm64: Introduce prctl() options to control the tagged user addresses ABI")
Cc: stable@vger.kernel.org
Signed-off-by: Shubhang Kaushik (Ampere) <sh@gentwo.org>
---
Testing:
  - futex_tagged_addr: PASS
  - Standard futex functional tests: PASS, except RT PI tests when run
    without real-time scheduling permission
  - sudo ./futex_requeue_pi: PASS (24/24)
  - sudo ./futex_requeue_pi_signal_restart: PASS
---
 kernel/futex/core.c                                |  2 +
 tools/testing/selftests/futex/functional/Makefile  |  1 +
 .../selftests/futex/functional/futex_tagged_addr.c | 47 ++++++++++++++++++++++
 tools/testing/selftests/futex/functional/run.sh    |  3 ++
 4 files changed, 53 insertions(+)

diff --git a/kernel/futex/core.c b/kernel/futex/core.c
index a240638f69e5c65bd036434a15f3df680efb4638..dcc8ffca38595a5ea69f3672375ffb9f9c6766f9 100644
--- a/kernel/futex/core.c
+++ b/kernel/futex/core.c
@@ -539,6 +539,8 @@ int get_futex_key(u32 __user *uaddr, unsigned int flags, union futex_key *key,
 	if (unlikely(!access_ok(uaddr, size)))
 		return -EFAULT;
 
+	address = untagged_addr(address);
+
 	if (unlikely(should_fail_futex(fshared)))
 		return -EFAULT;
 
diff --git a/tools/testing/selftests/futex/functional/Makefile b/tools/testing/selftests/futex/functional/Makefile
index a03bd5acba50319f08b04ff0904ef155c000fee3..8f7f6730285b725b483787362db4626c459d65c9 100644
--- a/tools/testing/selftests/futex/functional/Makefile
+++ b/tools/testing/selftests/futex/functional/Makefile
@@ -19,6 +19,7 @@ TEST_GEN_PROGS := \
 	futex_requeue_pi \
 	futex_requeue_pi_signal_restart \
 	futex_requeue_pi_mismatched_ops \
+	futex_tagged_addr \
 	futex_wait_uninitialized_heap \
 	futex_wait_private_mapped_file \
 	futex_wait \
diff --git a/tools/testing/selftests/futex/functional/futex_tagged_addr.c b/tools/testing/selftests/futex/functional/futex_tagged_addr.c
new file mode 100644
index 0000000000000000000000000000000000000000..81224f024a375d08e1dd4fd12ccf548f23126f27
--- /dev/null
+++ b/tools/testing/selftests/futex/functional/futex_tagged_addr.c
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <stdint.h>
+#include <sys/prctl.h>
+
+#include <linux/prctl.h>
+
+#include "futextest.h"
+#include "futex_thread.h"
+#include "kselftest_harness.h"
+
+static futex_t futex;
+static futex_t *tagged_futex;
+
+static int waiterfn(void *arg)
+{
+	struct timespec timeout = { .tv_sec = 1 };
+
+	return futex_wait(tagged_futex, 0, &timeout, FUTEX_PRIVATE_FLAG);
+}
+
+TEST(tagged_addr)
+{
+	struct futex_thread waiter;
+	int ret;
+
+#if !defined(__aarch64__) || __SIZEOF_POINTER__ != 8
+	SKIP(return, "arm64 64-bit tagged address ABI not available");
+#endif
+
+	ret = prctl(PR_SET_TAGGED_ADDR_CTRL, PR_TAGGED_ADDR_ENABLE, 0, 0, 0);
+	if (ret && errno == EINVAL)
+		SKIP(return, "tagged address ABI not available");
+	ASSERT_EQ(ret, 0)
+		TH_LOG("PR_SET_TAGGED_ADDR_CTRL failed: %s", strerror(errno));
+
+	tagged_futex = (futex_t *)((uintptr_t)&futex | (0x42ULL << 56));
+	ASSERT_EQ(futex_thread_create(&waiter, waiterfn, NULL), 0)
+		TH_LOG("pthread_create failed");
+	ASSERT_EQ(futex_wait_for_thread(&waiter, _metadata), 0)
+		TH_LOG("waiter did not enter futex wait");
+
+	EXPECT_EQ(futex_wake(&futex, 1, FUTEX_PRIVATE_FLAG), 1);
+	EXPECT_EQ(futex_thread_destroy(&waiter), 0);
+}
+
+TEST_HARNESS_MAIN
diff --git a/tools/testing/selftests/futex/functional/run.sh b/tools/testing/selftests/futex/functional/run.sh
index d1a681b798bd385570710e7c5f9bb28cca64de3e..620d13ee7171e7cc6061053750c9ff4cf525ff72 100755
--- a/tools/testing/selftests/futex/functional/run.sh
+++ b/tools/testing/selftests/futex/functional/run.sh
@@ -46,6 +46,9 @@ echo
 echo
 ./futex_waitv
 
+echo
+./futex_tagged_addr
+
 echo
 ./futex_priv_hash
 

---
base-commit: 3ac032ef6ca8e47bb06d1e189dfcf88fa0a09c75
change-id: 20261022-futex-arm64-tagged-uaddr-1cf56a42b2c1

Best regards,
-- 
Shubhang Kaushik (Ampere) <sh@gentwo.org>

Re: [PATCH] futex: Canonicalize arm64 tagged user addresses

[Catalin Marinas]

On Thu, Oct 22, 2026 at 07:06:02PM -0700, Shubhang Kaushik (Ampere) wrote:
> On arm64, a thread that enables the tagged address ABI can pass valid
> top-byte-tagged user pointers to futex syscalls. Such a pointer must
> have the same syscall behaviour as its untagged form.
> 
> get_futex_key() validates the user address with access_ok(), but uses
> the raw address when constructing private and shared-anonymous futex
> keys. A waiter using a tagged address and a waker using the untagged
> form can therefore access the same futex word while hashing to distinct
> keys, causing FUTEX_WAKE to miss the waiter.
> 
> Keep access_ok() on the original pointer so that arm64 retains its
> per-thread tagged-address ABI enforcement. Canonicalize the address
> afterwards, before it is used for futex key construction and shared
> mapping lookup.
> 
> Add an arm64 64-bit selftest for a tagged private waiter woken through
> the corresponding untagged address.
> 
> Fixes: 63f0c6037965 ("arm64: Introduce prctl() options to control the tagged user addresses ABI")
> Cc: stable@vger.kernel.org
> Signed-off-by: Shubhang Kaushik (Ampere) <sh@gentwo.org>

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>

The patch is fairly generic but I guess at the moment it only affects
arm64 (on x86, access_ok() on tagged pointers still fails).

Thomas, do you have any preference for this patch? Happy for it to go
via tip.

Thanks.

-- 
Catalin