Re: [PATCH-tip] sched: Don't call kfree() in do_set_cpus_allowed()
From: Paul E. McKenney
Date: Tue Nov 22 2022 - 11:34:12 EST
On Tue, Nov 22, 2022 at 10:23:43AM -0500, Waiman Long wrote:
> On 11/22/22 07:37, Peter Zijlstra wrote:
> > On Mon, Nov 21, 2022 at 10:04:33AM -0500, Waiman Long wrote:
> > > On 11/21/22 05:38, Peter Zijlstra wrote:
> > > > On Fri, Nov 18, 2022 at 02:33:02PM -0500, Waiman Long wrote:
> > > > > Commit 851a723e45d1 ("sched: Always clear user_cpus_ptr in
> > > > > do_set_cpus_allowed()") may call kfree() if user_cpus_ptr was previously
> > > > > set. Unfortunately, some of the callers of do_set_cpus_allowed()
> > > > 'some' ? There's only 3 or so, which one triggers this?
> > > It happenned at __kthread_bind_mask() where do_set_cpus_allowed() is called
> > > with pi_lock held.
> > >
> > > [ 1084.820105] <TASK>
> > > [ 1084.820110] dump_stack_lvl+0x57/0x81
> > > [ 1084.820117] check_noncircular+0x103/0x120
> > > [ 10[ 1084.820160] lock_acquire+0xba/0x230
> > > [ 1084.820164] ? kfree+0x10f/0x380
> > > [ 1084.820172] ? do_set_cpus_allowed+0x40/0x60
> > > [ 1084.820181] rt_spin_lock+0x27/0xe0
> > > [ 1084.820184] ? kfree+0x10f/0x380
> > > [ 1084.820188] kfree+0x10f/0x380
> > > [ 1084.820195] do_set_cpus_allowed+0x40/0x60
> > > [ 1084.820203] kthread_bind_mask+0x4a/0x70
> > > [ 1084.820211] create_worker+0xfb/0x1a0
> > > [ 1084.820220] worker_thread+0x2e3/0x3c0
> > > [ 1084.820226] ? process_one_work+0x450/0x450
> > > [ 1084.820230] kthread+0x111/0x130
> > > [ 1084.820236] ? kthread_complete_and_exit+0x20/0x20
> > > [ 1084.820244] ret_from_fork+0x22/0x30
> > > [ 1084.820258] </TASK>
> > > [ 1084.820260] BUG: sleeping function called from invalid context at
> > > kernel/locking/spinlock_rt.c:46
> > >
> > > It shows up with PREEMPT_RT kernel.
> > Oh, I see ..
> >
> > > Maybe. One thing that I am not clear about is why user_cpus_ptr is set in
> > > the first place.
> > Perhaps someone set an affinity on kthreadd ?
> >
> > But I'm thinking this exact problem is also possible (rather more likely
> > even) with select_fallback_rq() that too holds pi_lock (which account
> > for both other users of this function).
> >
> > Bah.
> >
> > And the allocation is just the one long in size (for small configs)
> > which is just enough space for a single linked list like you had.
> That is exactly the reason why I use lockless list.
> >
> > Urgh.
> >
> > The below is yuck too, and I'm not sure Paul wants us to use
> > kvfree_call_rcu() without its wrapper.
> >
> > ---
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 78b2d5cabcc5..0d0af0fc7fcf 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -2606,7 +2606,12 @@ void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
> > };
> > __do_set_cpus_allowed(p, &ac);
> > - kfree(ac.user_mask);
> > + /*
> > + * Because this is called with p->pi_lock held, it is not possible
> > + * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
> > + * kfree_rcu().
> > + */
> > + kvfree_call_rcu((struct rcu_head *)ac.user_mask, (rcu_callback_t)0);
> > }
>
> I guess you need to do a NULL check before calling kvfree_call_rcu() as I
> don't think kvfree_call_rcu() does that. Also it is unlikely that we need to
> call it.
Indeed, the NULL check is in kvfree_rcu_arg_2(). By the time you get
to kvfree_call_rcu, the pointer is assumed to be non-NULL.
> > int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
> > @@ -8196,7 +8201,7 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
> > struct affinity_context ac;
> > struct cpumask *user_mask;
> > struct task_struct *p;
> > - int retval;
> > + int retval, size;
> > rcu_read_lock();
> > @@ -8229,7 +8234,11 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
> > if (retval)
> > goto out_put_task;
> > - user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
> > + /*
> > + * See do_set_cpus_allowed() for the rcu_head usage.
> > + */
> > + size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
> > + user_mask = kmalloc(size, GFP_KERNEL);
> > if (!user_mask) {
> > retval = -ENOMEM;
> > goto out_put_task;
>
> I guess that will work too. Just like you, I am a bit uneasy to call into
> kvfree_call_rcu() directly as it may change in the future. How about
>
> iff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 78b2d5cabcc5..5fac4aa6ac7f 100644
> --- a/kernel/sched/core.c
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 78b2d5cabcc5..5fac4aa6ac7f 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2593,6 +2593,11 @@ __do_set_cpus_allowed(struct task_struct *p, struct
> affinity_context *ctx)
> set_next_task(rq, p);
> }
>
> +union cpumask_rcuhead {
> + void *cpumask;
> + struct rcu_head rcu;
> +};
> +
> /*
> * Used for kthread_bind() and select_fallback_rq(), in both cases the user
> * affinity (if any) should be destroyed too.
> @@ -2606,7 +2611,12 @@ void do_set_cpus_allowed(struct task_struct *p, const
> struct cpumask *new_mask)
> };
>
> __do_set_cpus_allowed(p, &ac);
> - kfree(ac.user_mask);
> + /*
> + * Because this is called with p->pi_lock held, it is not possible
> + * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
> + * kfree_rcu().
> + */
> + kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
This looks plausible to me.
Thanx, Paul
> }
>
> int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
> @@ -8196,7 +8206,7 @@ long sched_setaffinity(pid_t pid, const struct cpumask
> *in_mask)
> struct affinity_context ac;
> struct cpumask *user_mask;
> struct task_struct *p;
> - int retval;
> + int retval, size;
>
> rcu_read_lock();
>
> @@ -8229,7 +8239,11 @@ long sched_setaffinity(pid_t pid, const struct
> cpumask *in_mask)
> if (retval)
> goto out_put_task;
>
> - user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
> + /*
> + * See do_set_cpus_allowed() for the rcu_head usage.
> + */
> + size = max_t(int, cpumask_size(), sizeof(union cpumask_rcuhead));
> + user_mask = kmalloc(size, GFP_KERNEL);
> if (!user_mask) {
> retval = -ENOMEM;
> goto out_put_task;
>
> Cheers,
> Longman
>
>