Re: [PATCH 5/6] sched_ext: idle: Per-node idle cpumasks

[Andrea Righi <arighi@nvidia.com>]

On Tue, Feb 11, 2025 at 11:38:27AM -0500, Steven Rostedt wrote:
> On Tue, 11 Feb 2025 15:45:15 +0100
> Andrea Righi <arighi@nvidia.com> wrote:
> 
> > ...which is basically this (with GFP_ATOMIC):
> > 
> > [   11.829079] =============================
> > [   11.829109] [ BUG: Invalid wait context ]
> > [   11.829146] 6.13.0-virtme #51 Not tainted
> > [   11.829185] -----------------------------
> > [   11.829243] fish/344 is trying to lock:
> > [   11.829285] ffff9659bec450b0 (&c->lock){..-.}-{3:3}, at: ___slab_alloc+0x66/0x1510
> > [   11.829380] other info that might help us debug this:
> > [   11.829450] context-{5:5}
> > [   11.829494] 8 locks held by fish/344:
> > [   11.829534]  #0: ffff965a409c70a0 (&tty->ldisc_sem){++++}-{0:0}, at: tty_ldisc_ref_wait+0x28/0x60
> > [   11.829643]  #1: ffff965a409c7130 (&tty->atomic_write_lock){+.+.}-{4:4}, at: file_tty_write.isra.0+0xa1/0x330
> > [   11.829765]  #2: ffff965a409c72e8 (&tty->termios_rwsem/1){++++}-{4:4}, at: n_tty_write+0x9e/0x510
> > [   11.829871]  #3: ffffbc6d01433380 (&ldata->output_lock){+.+.}-{4:4}, at: n_tty_write+0x1f1/0x510
> > [   11.829979]  #4: ffffffffb556b5c0 (rcu_read_lock){....}-{1:3}, at: __queue_work+0x59/0x680
> > [   11.830173]  #5: ffff9659800f0018 (&pool->lock){-.-.}-{2:2}, at: __queue_work+0xd7/0x680
> > [   11.830286]  #6: ffff9659801bcf60 (&p->pi_lock){-.-.}-{2:2}, at: try_to_wake_up+0x56/0x920
> > [   11.830396]  #7: ffffffffb556b5c0 (rcu_read_lock){....}-{1:3}, at: scx_select_cpu_dfl+0x56/0x460
> > 
> > And I think that's because:
> > 
> >  * %GFP_ATOMIC users can not sleep and need the allocation to succeed. A lower
> >  * watermark is applied to allow access to "atomic reserves".
> >  * The current implementation doesn't support NMI and few other strict
> >  * non-preemptive contexts (e.g. raw_spin_lock). The same applies to %GFP_NOWAIT.
> > 
> > So I guess we the only viable option is to preallocate nodemask_t and
> > protect it somehow, hoping that it doesn't add too much overhead...
> 
> I believe it's because you have p->pi_lock which is a raw_spin_lock() and
> you are trying to take a lock in ___slab_alloc() which I bet is a normal
> spin_lock(). In PREEMPT_RT() that turns into a mutex, and you can not take
> a spin_lock while holding a raw_spin_lock.

Exactly that, thanks Steve. I'll run some tests using per-cpu nodemask_t,
given that most of the times this is called with p->pi_lock held, it should
be safe and we shouldn't introduce any overhead.

-Andrea