Re: [PATCH 6/6] sched_ext: Don't hold scx_tasks_lock for too long

[David Vernet]

On Wed, Oct 09, 2024 at 11:41:02AM -1000, Tejun Heo wrote:
> While enabling and disabling a BPF scheduler, every task is iterated a
> couple times by walking scx_tasks. Except for one, all iterations keep
> holding scx_tasks_lock. On multi-socket systems under heavy rq lock
> contention and high number of threads, this can can lead to RCU and other
> stalls.
> 
> The following is triggered on a 2 x AMD EPYC 7642 system (192 logical CPUs)
> running `stress-ng --workload 150 --workload-threads 10` with >400k idle
> threads and RCU stall period reduced to 5s:
> 
>   rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
>   rcu:     91-...!: (10 ticks this GP) idle=0754/1/0x4000000000000000 softirq=18204/18206 fqs=17
>   rcu:     186-...!: (17 ticks this GP) idle=ec54/1/0x4000000000000000 softirq=25863/25866 fqs=17
>   rcu:     (detected by 80, t=10042 jiffies, g=89305, q=33 ncpus=192)
>   Sending NMI from CPU 80 to CPUs 91:
>   NMI backtrace for cpu 91
>   CPU: 91 UID: 0 PID: 284038 Comm: sched_ext_ops_h Kdump: loaded Not tainted 6.12.0-rc2-work-g6bf5681f7ee2-dirty #471
>   Hardware name: Supermicro Super Server/H11DSi, BIOS 2.8 12/14/2023
>   Sched_ext: simple (disabling+all)
>   RIP: 0010:queued_spin_lock_slowpath+0x17b/0x2f0
>   Code: 02 c0 10 03 00 83 79 08 00 75 08 f3 90 83 79 08 00 74 f8 48 8b 11 48 85 d2 74 09 0f 0d 0a eb 0a 31 d2 eb 06 31 d2 eb 02 f3 90 <8b> 07 66 85 c0 75 f7 39 d8 75 0d be 01 00 00 00 89 d8 f0 0f b1 37
>   RSP: 0018:ffffc9000fadfcb8 EFLAGS: 00000002
>   RAX: 0000000001700001 RBX: 0000000001700000 RCX: ffff88bfcaaf10c0
>   RDX: 0000000000000000 RSI: 0000000000000101 RDI: ffff88bfca8f0080
>   RBP: 0000000001700000 R08: 0000000000000090 R09: ffffffffffffffff
>   R10: ffff88a74761b268 R11: 0000000000000000 R12: ffff88a6b6765460
>   R13: ffffc9000fadfd60 R14: ffff88bfca8f0080 R15: ffff88bfcaac0000
>   FS:  0000000000000000(0000) GS:ffff88bfcaac0000(0000) knlGS:0000000000000000
>   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>   CR2: 00007f5c55f526a0 CR3: 0000000afd474000 CR4: 0000000000350eb0
>   Call Trace:
>    <NMI>
>    </NMI>
>    <TASK>
>    do_raw_spin_lock+0x9c/0xb0
>    task_rq_lock+0x50/0x190
>    scx_task_iter_next_locked+0x157/0x170
>    scx_ops_disable_workfn+0x2c2/0xbf0
>    kthread_worker_fn+0x108/0x2a0
>    kthread+0xeb/0x110
>    ret_from_fork+0x36/0x40
>    ret_from_fork_asm+0x1a/0x30
>    </TASK>
>   Sending NMI from CPU 80 to CPUs 186:
>   NMI backtrace for cpu 186
>   CPU: 186 UID: 0 PID: 51248 Comm: fish Kdump: loaded Not tainted 6.12.0-rc2-work-g6bf5681f7ee2-dirty #471
> 
> scx_task_iter can safely drop locks while iterating. Make
> scx_task_iter_next() drop scx_tasks_lock every 32 iterations to avoid
> stalls.
> 
> Signed-off-by: Tejun Heo <tj@xxxxxxxxxx>

LG, just had one question below.

Acked-by: David Vernet <void@xxxxxxxxxxxxx>

> ---
>  kernel/sched/ext.c | 19 +++++++++++++++++--
>  1 file changed, 17 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> index d53c7a365fec..b44946198ea5 100644
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -18,6 +18,12 @@ enum scx_consts {
>  	SCX_EXIT_DUMP_DFL_LEN		= 32768,
>  
>  	SCX_CPUPERF_ONE			= SCHED_CAPACITY_SCALE,
> +
> +	/*
> +	 * Iterating all tasks may take a while. Periodically drop
> +	 * scx_tasks_lock to avoid causing e.g. CSD and RCU stalls.
> +	 */
> +	SCX_OPS_TASK_ITER_BATCH		= 32,
>  };
>  
>  enum scx_exit_kind {
> @@ -1273,6 +1279,7 @@ struct scx_task_iter {
>  	struct task_struct		*locked;
>  	struct rq			*rq;
>  	struct rq_flags			rf;
> +	u32				cnt;
>  };
>  
>  /**
> @@ -1301,6 +1308,7 @@ static void scx_task_iter_start(struct scx_task_iter *iter)
>  	iter->cursor = (struct sched_ext_entity){ .flags = SCX_TASK_CURSOR };
>  	list_add(&iter->cursor.tasks_node, &scx_tasks);
>  	iter->locked = NULL;
> +	iter->cnt = 0;
>  }
>  
>  static void __scx_task_iter_rq_unlock(struct scx_task_iter *iter)
> @@ -1355,14 +1363,21 @@ static void scx_task_iter_stop(struct scx_task_iter *iter)
>   * scx_task_iter_next - Next task
>   * @iter: iterator to walk
>   *
> - * Visit the next task. See scx_task_iter_start() for details.
> + * Visit the next task. See scx_task_iter_start() for details. Locks are dropped
> + * and re-acquired every %SCX_OPS_TASK_ITER_BATCH iterations to avoid causing
> + * stalls by holding scx_tasks_lock for too long.
>   */
>  static struct task_struct *scx_task_iter_next(struct scx_task_iter *iter)
>  {
>  	struct list_head *cursor = &iter->cursor.tasks_node;
>  	struct sched_ext_entity *pos;
>  
> -	lockdep_assert_held(&scx_tasks_lock);
> +	if (!(++iter->cnt % SCX_OPS_TASK_ITER_BATCH)) {
> +		scx_task_iter_unlock(iter);
> +		cpu_relax();

Could you explain why we need this cpu_relax()? I thought it was only
necessary for busy-wait loops.

> +		cond_resched();
> +		scx_task_iter_relock(iter);
> +	}
>  
>  	list_for_each_entry(pos, cursor, tasks_node) {
>  		if (&pos->tasks_node == &scx_tasks)
> -- 
> 2.46.2
>

Attachment: signature.asc
Description: PGP signature