mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
* [PATCH] futex/pi: Reject owners that have no mm
@ 2026-11-16 13:26 Daniel Wade
  0 siblings, 0 replies; only message in thread
From: Daniel Wade @ 2026-11-16 13:26 UTC (permalink / raw)
  To: Thomas Gleixner, Ingo Molnar, Peter Zijlstra, Darren Hart
  Cc: Davidlohr Bueso, Andre Almeida, Sebastian Andrzej Siewior,
	Kyle Zeng, linux-kernel

attach_to_pi_owner() attaches PI state to the task named in the futex word
and rejects kernel threads, which cannot own a user space futex.

A usermode helper thread runs with PF_KTHREAD cleared and acquires an mm at
exec(). While it is in that state, any task that can see its TID can name
it as the owner of a shared PI futex. Commit 59b3732f95dd ("futex/pi:
Reject cross-mm private futex owners") compares the owner mm with the
waiter mm, which applies to private keys only.

The helper then execs or exits, exit_pi_state_list() takes a reference on
the private hash of current->mm, and futex_private_hash() dereferences a
NULL mm. The oops path calls do_exit(), which runs the futex cleanup again
and blocks on the futex exit_mutex, so the thread stays there and with
preemptible RCU grace periods stop completing. Module auto-loading starts
usermode helpers on behalf of ordinary users, so an unprivileged task
reaches this.

Reject an owner that has no mm with -EPERM, as for a kernel thread, after
the futex state check so that an exited task still goes through
handle_exit_race().

Fixes: 80367ad01d93 ("futex: Add basic infrastructure for local task local hash")
Signed-off-by: Daniel Wade <danjwade95@gmail.com>
Cc: stable@vger.kernel.org
---
Tested on mainline b3dfbde62cf1 (kernel/futex identical to v7.4-rc3),
x86-64 with KASAN and lockdep, 2 vCPUs under KVM, as an ordinary user.

Without the patch a usermode helper accepted PI ownership 130 times, with
no EPERM refusal, and the kernel oopsed 22 s in:

  Oops: stack segment: 0000 [#1] SMP KASAN NOPTI
  RIP: 0010:futex_private_hash+0x4c/0x2b0
  Call Trace:
   futex_exit_exec_release
   mm_exit_exec_release
   begin_new_exec
   call_usermodehelper_exec_async

after which the kworker never exited and RCU stall warnings repeated. With
the patch: 201472 refusals, 5822 owners that do have an mm still accepted,
no oops and no stall.

Not tested on other architectures.

 kernel/futex/pi.c |   12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/kernel/futex/pi.c b/kernel/futex/pi.c
index 98f1b962e59a..77e88dd5dc22 100644
--- a/kernel/futex/pi.c
+++ b/kernel/futex/pi.c
@@ -515,6 +515,18 @@ static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
 		return ret;
 	}
 
+	/*
+	 * A usermode helper thread runs with PF_KTHREAD cleared and acquires
+	 * an mm at exec(), so the kernel thread test above lets it through.
+	 * Its exec() or exit() cleanup runs exit_pi_state_list(), which uses
+	 * current->mm.
+	 */
+	if (unlikely(!p->mm)) {
+		raw_spin_unlock_irq(&p->pi_lock);
+		put_task_struct(p);
+		return -EPERM;
+	}
+
 	if (IS_ENABLED(CONFIG_MMU) && futex_key_is_private(key)) {
 		/*
 		 * A private futex key holds a pointer to the waiter's mm

^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2026-11-16 13:26 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-16 13:26 [PATCH] futex/pi: Reject owners that have no mm Daniel Wade

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®