lock_task_sighand() can trigger a hard lockup. If NR_CPUS threads call
getrusage() at the same time and the process has NR_THREADS, spin_lock_irq
will spin with irqs disabled O(NR_CPUS * NR_THREADS) time.

Change getrusage() to use sig->stats_lock, it was specifically designed
for this type of use. This way it runs lockless in the likely case.

TODO:
	- Change do_task_stat() to use sig->stats_lock too, then we can
	  remove spin_lock_irq(siglock) in wait_task_zombie().

	- Turn sig->stats_lock into seqcount_rwlock_t, this way the
	  readers in the slow mode won't exclude each other. See
	  https://lore.kernel.org/all/20230913154907.GA26210@redhat.com/

	- stats_lock has to disable irqs because ->siglock can be taken
	  in irq context, it would be very nice to change __exit_signal()
	  to avoid the siglock->stats_lock dependency.

Reported-and-tested-by: Dylan Hatch <dylanbhatch@google.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
---
 kernel/sys.c | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/kernel/sys.c b/kernel/sys.c
index 70ad06ad852e..f8e543f1e38a 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -1788,7 +1788,9 @@ void getrusage(struct task_struct *p, int who, struct rusage *r)
 	unsigned long maxrss;
 	struct mm_struct *mm;
 	struct signal_struct *sig = p->signal;
+	unsigned int seq = 0;
 
+retry:
 	memset(r, 0, sizeof(*r));
 	utime = stime = 0;
 	maxrss = 0;
@@ -1800,8 +1802,7 @@ void getrusage(struct task_struct *p, int who, struct rusage *r)
 		goto out_thread;
 	}
 
-	if (!lock_task_sighand(p, &flags))
-		return;
+	flags = read_seqbegin_or_lock_irqsave(&sig->stats_lock, &seq);
 
 	switch (who) {
 	case RUSAGE_BOTH:
@@ -1829,14 +1830,23 @@ void getrusage(struct task_struct *p, int who, struct rusage *r)
 		r->ru_oublock += sig->oublock;
 		if (maxrss < sig->maxrss)
 			maxrss = sig->maxrss;
+
+		rcu_read_lock();
 		__for_each_thread(sig, t)
 			accumulate_thread_rusage(t, r);
+		rcu_read_unlock();
+
 		break;
 
 	default:
 		BUG();
 	}
-	unlock_task_sighand(p, &flags);
+
+	if (need_seqretry(&sig->stats_lock, seq)) {
+		seq = 1;
+		goto retry;
+	}
+	done_seqretry_irqrestore(&sig->stats_lock, seq, flags);
 
 	if (who == RUSAGE_CHILDREN)
 		goto out_children;
-- 
2.25.1.362.g51ebf55

On Mon, 22 Jan 2024 16:50:53 +0100 Oleg Nesterov <oleg@redhat.com> wrote:

> lock_task_sighand() can trigger a hard lockup. If NR_CPUS threads call
> getrusage() at the same time and the process has NR_THREADS, spin_lock_irq
> will spin with irqs disabled O(NR_CPUS * NR_THREADS) time.

It would be super interesting to see Dylan's original report.

Is it possible for carefully-crafted unprivileged userspace to
deliberately trigger this?

On 01/22, Andrew Morton wrote:
>
> On Mon, 22 Jan 2024 16:50:53 +0100 Oleg Nesterov <oleg@redhat.com> wrote:
>
> > lock_task_sighand() can trigger a hard lockup. If NR_CPUS threads call
> > getrusage() at the same time and the process has NR_THREADS, spin_lock_irq
> > will spin with irqs disabled O(NR_CPUS * NR_THREADS) time.
>
> It would be super interesting to see Dylan's original report.

from "[RFC PATCH] getrusage: Use trylock when getting sighand lock."
https://lore.kernel.org/all/20240117192534.1327608-1-dylanbhatch@google.com/

	Processes with many threads run the risk of causing a hard lockup if
	too many threads are calling getrusage() at once. This is because a
	calling thread with RUSAGE_SELF spins on the sighand lock with irq
	disabled, and the critical section of getrusage scales linearly with the
	size of the process. All cpus may end up spinning on the sighand lock
	for a long time because another thread has the lock and is busy
	iterating over 250k+ threads.

> Is it possible for carefully-crafted unprivileged userspace to
> deliberately trigger this?

Yes, just you need to create a process with a lot of threads calling
getrusage().

See mine and Dylan's test-cases in
https://lore.kernel.org/all/CADBMgpz7k=LhktfcJhSDBDWN0oLeQxPqhOVws3fq0LNpnfOSYg@mail.gmail.com/
There are very similar and simple.


And again, this is a known problem and we need more fixes.

Oleg.