On 8/4/21 9:47 AM, Sebastian Andrzej Siewior wrote:
> On 2021-08-04 09:39:30 [-0600], Jens Axboe wrote:
>> I'm confused, the waitqueue locks are always IRQ disabling.
>
> spin_lock_irq() does not disable interrupts on -RT. The patch above
> produces:
>
> | BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:35
> | in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 2020, name: iou-wrk-2018
> | 1 lock held by iou-wrk-2018/2020:
> | #0: ffff888111a47de8 (&hash->wait){+.+.}-{0:0}, at: io_worker_handle_work+0x443/0x630
> | irq event stamp: 10
> | hardirqs last enabled at (9): [<ffffffff81c47818>] _raw_spin_unlock_irqrestore+0x28/0x70
> | hardirqs last disabled at (10): [<ffffffff81c4769e>] _raw_spin_lock_irq+0x3e/0x40
> | softirqs last enabled at (0): [<ffffffff81077238>] copy_process+0x8f8/0x2020
> | softirqs last disabled at (0): [<0000000000000000>] 0x0
> | CPU: 5 PID: 2020 Comm: iou-wrk-2018 Tainted: G W 5.14.0-rc4-rt4+ #97
> | Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014
> | Call Trace:
> | dump_stack_lvl+0x45/0x59
> | ___might_sleep.cold+0xa6/0xb6
> | rt_spin_lock+0x35/0xc0
> | ? io_worker_handle_work+0x443/0x630
> | io_worker_handle_work+0x443/0x630
> | io_wqe_worker+0xb4/0x340
> | ? lockdep_hardirqs_on_prepare+0xd4/0x170
> | ? _raw_spin_unlock_irqrestore+0x28/0x70
> | ? _raw_spin_unlock_irqrestore+0x28/0x70
> | ? io_worker_handle_work+0x630/0x630
> | ? rt_mutex_slowunlock+0x2ba/0x310
> | ? io_worker_handle_work+0x630/0x630
> | ret_from_fork+0x22/0x30
>
>
> But indeed, you are right, my snippet breaks non-RT. So this then maybe:
>
> diff --git a/fs/io-wq.c b/fs/io-wq.c
> index 57d3cdddcdb3e..0b931ac3c83e6 100644
> --- a/fs/io-wq.c
> +++ b/fs/io-wq.c
> @@ -384,7 +384,7 @@ static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
> {
> struct io_wq *wq = wqe->wq;
>
> - spin_lock(&wq->hash->wait.lock);
> + spin_lock_irq(&wq->hash->wait.lock);
> if (list_empty(&wqe->wait.entry)) {
> __add_wait_queue(&wq->hash->wait, &wqe->wait);
> if (!test_bit(hash, &wq->hash->map)) {
> @@ -392,7 +392,7 @@ static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
> list_del_init(&wqe->wait.entry);
> }
> }
> - spin_unlock(&wq->hash->wait.lock);
> + spin_unlock_irq(&wq->hash->wait.lock);
> }
>
> static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
> @@ -430,9 +430,9 @@ static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
> }
>
> if (stall_hash != -1U) {
> - raw_spin_unlock(&wqe->lock);
> + raw_spin_unlock_irq(&wqe->lock);
> io_wait_on_hash(wqe, stall_hash);
> - raw_spin_lock(&wqe->lock);
> + raw_spin_lock_irq(&wqe->lock);
> }
>
> return NULL;
>
> (this is on-top of the patch you sent earlier and Daniel Cc: me on after
> I checked that the problem/warning still exists).
That'd work on non-RT as well, but it makes it worse on non-RT as well with
the irq enable/disable dance. While that's not the end of the world, would
be nice to have a solution that doesn't sacrifice anything, yet doesn't
make RT unhappy.
--
Jens Axboe
