Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.35 as permitted sender) client-ip=23.128.96.35;
From:   Xiaobing Li <xiaobing.li@samsung.com>
To:     axboe@kernel.dk
Cc:     asml.silence@gmail.com, linux-kernel@vger.kernel.org,
        io-uring@vger.kernel.org, kun.dou@samsung.com,
        peiwei.li@samsung.com, joshi.k@samsung.com,
        kundan.kumar@samsung.com, wenwen.chen@samsung.com,
        ruyi.zhang@samsung.com, xiaobing.li@samsung.com
Subject: Re: [PATCH v3] io_uring/fdinfo: remove need for sqpoll lock for
 thread/pid retrieval
Date:   Sat, 18 Nov 2023 11:19:51 +0800
Message-Id: <20231118031951.21764-1-xiaobing.li@samsung.com>
In-Reply-To: <605eac76-ec47-436b-872a-f6e8b4094293@kernel.dk>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Content-Type: text/plain; charset="utf-8"
CMS-TYPE: 105P
DLP-Filter: Pass
References: <605eac76-ec47-436b-872a-f6e8b4094293@kernel.dk>
        <CGME20231118032740epcas5p20b6aad6264323376fa024bc2a56f0990@epcas5p2.samsung.com>
Precedence: bulk

On 11/15/23 6:42 AM, Jens Axboe wrote:
> 	 */
> 	has_lock = mutex_trylock(&ctx->uring_lock);
> 
>-	if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
>-		struct io_sq_data *sq = ctx->sq_data;
>-
>-		sq_pid = sq->task_pid;
>-		sq_cpu = sq->sq_cpu;
>+	if (ctx->flags & IORING_SETUP_SQPOLL) {
>+		struct io_sq_data *sq;
>+
>+		rcu_read_lock();
>+		sq = READ_ONCE(ctx->sq_data);
>+		if (sq) {
>+			sq_pid = sq->task_pid;
>+			sq_cpu = sq->sq_cpu;
>+		}
>+		rcu_read_unlock();
> 	}
> 
> 	seq_printf(m, "SqThread:\t%d\n", sq_pid);
>diff --git a/io_uring/sqpoll.c b/io_uring/sqpoll.c
>index 65b5dbe3c850..583c76945cdf 100644
>--- a/io_uring/sqpoll.c
>+++ b/io_uring/sqpoll.c
>@@ -70,7 +70,7 @@ void io_put_sq_data(struct io_sq_data *sqd)
> 		WARN_ON_ONCE(atomic_read(&sqd->park_pending));
> 
> 		io_sq_thread_stop(sqd);
>-		kfree(sqd);
>+		kfree_rcu(sqd, rcu);
> 	}
> }
> 
>@@ -313,7 +313,7 @@ static int io_sq_thread(void *data)
> 	}
> 
> 	io_uring_cancel_generic(true, sqd);
>-	sqd->thread = NULL;
>+	WRITE_ONCE(sqd->thread, NULL);
> 	list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
> 		atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags);
> 	io_run_task_work();
>@@ -411,7 +411,7 @@ __cold int io_sq_offload_create(struct io_ring_ctx *ctx,
> 			goto err_sqpoll;
> 		}
> 
>-		sqd->thread = tsk;
>+		WRITE_ONCE(sqd->thread, tsk);
> 		ret = io_uring_alloc_task_context(tsk, ctx);
> 		wake_up_new_task(tsk);
> 		if (ret)
>diff --git a/io_uring/sqpoll.h b/io_uring/sqpoll.h
>index 8df37e8c9149..0cf0c5833a27 100644
>--- a/io_uring/sqpoll.h
>+++ b/io_uring/sqpoll.h
>@@ -18,6 +18,8 @@ struct io_sq_data {
> 
> 	unsigned long		state;
> 	struct completion	exited;
>+
>+	struct rcu_head		rcu;
> };
> 
> int io_sq_offload_create(struct io_ring_ctx *ctx, struct io_uring_params *p);

I tested this and it worked after adding RCU lock.
It consistently outputs correct results.

The results of a simple test are as follows:
Every 0.5s: cat /proc/10212/fdinfo/6 | grep Sq
SqMask: 0x3
SqHead: 17422716
SqTail: 17422716
CachedSqHead:   17422716
SqThread:       10212
SqThreadCpu:    73
SqBusy: 97%
-------------------------------------------------------------
But the name of the sq thread is "iou-sqp-" + "the PID of its parent process":
    PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
  10211 root      20   0  184408   8192      0 R  99.9   0.0   4:01.42 fio
  10212 root      20   0  184408   8192      0 R  99.9   0.0   4:01.48 iou-sqp-10211
Is this the originally desired effect?