[PATCH v1] landlock/tsync: fix null-ptr-deref in cancel_tsync_works()

[Jiayuan Chen]

cancel_tsync_works() iterates over works->works[0..size-1] and calls
task_work_cancel() on each entry.  task_work_cancel() leads to
task_work_pending(), which dereferences task->task_works.  If
works->works[i]->task is NULL, this causes a null-ptr-deref:

KASAN: null-ptr-deref in range [0x00000000000009a0-0x00000000000009a7]
RIP: 0010:task_work_pending include/linux/task_work.h:26 [inline]
RIP: 0010:task_work_cancel_match+0x86/0x250 kernel/task_work.c:124
RSP: 0018:ffffc90003597ba0 EFLAGS: 00010202
RAX: 0000000000000134 RBX: 0000000000000000 RCX: ffffc900106b1000
RDX: 0000000000080000 RSI: ffffffff81d13236 RDI: 0000000000000000
RBP: 1ffff920006b2f77 R08: 0000000000000007 R09: 0000000000000000
R10: 0000000000000002 R11: 0000000000000000 R12: ffffffff81d12dd0
R13: ffff88802c045100 R14: dffffc0000000000 R15: 00000000000009a0
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 000000110c3ea90c CR3: 0000000037f63000 CR4: 00000000003526f0
DR0: 0000000000000003 DR1: 00000000000001f8 DR2: 000000000000008e
DR3: 000000000000057a DR6: 00000000ffff0ff0 DR7: 0000000000000400
Call Trace:
 <TASK>
 task_work_cancel+0x23/0x30 kernel/task_work.c:187
 cancel_tsync_works security/landlock/tsync.c:415 [inline]
 landlock_restrict_sibling_threads+0xafe/0x1280 security/landlock/tsync.c:533
 __do_sys_landlock_restrict_self+0x5c9/0x9e0 security/landlock/syscalls.c:574
 do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
 do_syscall_64+0x106/0xf80 arch/x86/entry/syscall_64.c:94
 entry_SYSCALL_64_after_hwframe+0x77/0x7f
RIP: 0033:0x7f859b39c629
RSP: 002b:00007f85991b2028 EFLAGS: 00000246 ORIG_RAX: 00000000000001be
RAX: ffffffffffffffda RBX: 00007f859b616270 RCX: 00007f859b39c629
RDX: 0000000000000000 RSI: 000000000000000a RDI: 0000000000000003
RBP: 00007f859b432b39 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
R13: 00007f859b616308 R14: 00007f859b616270 R15: 00007ffcff084488

The root cause is a race in schedule_task_work().  tsync_works_provide()
increments works->size and stores the task reference in ctx->task *before*
task_work_add() is called.  A thread can race to call do_exit() in the
window between the PF_EXITING check and task_work_add(), causing
task_work_add() to return -ESRCH.  The error path then drops the task
reference and sets ctx->task = NULL, but works->size remains incremented.
A subsequent call to cancel_tsync_works() iterates up to the stale size
and passes the NULL task pointer to task_work_cancel().

Fix this by decrementing works->size in the task_work_add() error path,
so the failed slot is rolled back and cancel_tsync_works() never iterates
over it.  The slot is naturally reused in subsequent iterations since
tsync_works_provide() always picks works->works[works->size].

As a defensive measure, also add a WARN_ONCE() guard in cancel_tsync_works()
to catch any future NULL task pointer before dereferencing it.

Fixes: 42fc7e6543f6 ("landlock: Multithreading support for landlock_restrict_self()")
Reported-by: syzbot+911d99dc200feac03ea6 at syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=911d99dc200feac03ea6
Signed-off-by: Jiayuan Chen <jiayuan.chen at linux.dev>
---
 security/landlock/tsync.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/security/landlock/tsync.c b/security/landlock/tsync.c
index 0d2b9c646030..e6d742529484 100644
--- a/security/landlock/tsync.c
+++ b/security/landlock/tsync.c
@@ -381,14 +381,14 @@ static bool schedule_task_work(struct tsync_works *works,
 		err = task_work_add(thread, &ctx->work, TWA_SIGNAL);
 		if (err) {
 			/*
-			 * task_work_add() only fails if the task is about to exit.  We
-			 * checked that earlier, but it can happen as a race.  Resume
-			 * without setting an error, as the task is probably gone in the
-			 * next loop iteration.  For consistency, remove the task from ctx
-			 * so that it does not look like we handed it a task_work.
+			 * task_work_add() only fails if the task is about to exit.
+			 * We checked PF_EXITING earlier, but the thread can race to
+			 * exit between that check and task_work_add().  Roll back the
+			 * slot so cancel_tsync_works() never sees a NULL task pointer.
 			 */
 			put_task_struct(ctx->task);
 			ctx->task = NULL;
+			works->size--;
 
 			atomic_dec(&shared_ctx->num_preparing);
 			atomic_dec(&shared_ctx->num_unfinished);
@@ -412,6 +412,11 @@ static void cancel_tsync_works(struct tsync_works *works,
 	int i;
 
 	for (i = 0; i < works->size; i++) {
+		if (WARN_ONCE(!works->works[i]->task,
+			      "landlock: unexpected NULL task in tsync slot %d\n",
+			      i))
+			continue;
+
 		if (!task_work_cancel(works->works[i]->task,
 				      &works->works[i]->work))
 			continue;
-- 
2.43.0