[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <0d7deb6e-f12d-4cd9-97f5-2661ddaab25a@nvidia.com>
Date: Thu, 19 Sep 2024 20:48:20 +0300
From: Moshe Shemesh <moshe@...dia.com>
To: Mohamed Khalfella <mkhalfella@...estorage.com>, Tariq Toukan
<tariqt@...dia.com>, Alexander Lobakin <aleksander.lobakin@...el.com>, "Saeed
Mahameed" <saeedm@...dia.com>, Leon Romanovsky <leon@...nel.org>, "David S.
Miller" <davem@...emloft.net>, Eric Dumazet <edumazet@...gle.com>, "Jakub
Kicinski" <kuba@...nel.org>, Paolo Abeni <pabeni@...hat.com>
CC: <yzhong@...estorage.com>, Przemek Kitszel <przemyslaw.kitszel@...el.com>,
Shay Drori <shayd@...dia.com>, <netdev@...r.kernel.org>,
<linux-rdma@...r.kernel.org>, <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v3 1/1] net/mlx5: Added cond_resched() to crdump
collection
On 9/5/2024 7:02 AM, Mohamed Khalfella wrote:
> Collecting crdump involves reading vsc registers from pci config space
> of mlx device, which can take long time to complete. This might result
> in starving other threads waiting to run on the cpu.
>
> Numbers I got from testing ConnectX-5 Ex MCX516A-CDAT in the lab:
>
> - mlx5_vsc_gw_read_block_fast() was called with length = 1310716.
> - mlx5_vsc_gw_read_fast() reads 4 bytes at a time. It was not used to
> read the entire 1310716 bytes. It was called 53813 times because
> there are jumps in read_addr.
> - On average mlx5_vsc_gw_read_fast() took 35284.4ns.
> - In total mlx5_vsc_wait_on_flag() called vsc_read() 54707 times.
> The average time for each call was 17548.3ns. In some instances
> vsc_read() was called more than one time when the flag was not set.
> As expected the thread released the cpu after 16 iterations in
> mlx5_vsc_wait_on_flag().
> - Total time to read crdump was 35284.4ns * 53813 ~= 1.898s.
>
> It was seen in the field that crdump can take more than 5 seconds to
> complete. During that time mlx5_vsc_wait_on_flag() did not release the
> cpu because it did not complete 16 iterations. It is believed that pci
> config reads were slow. Adding cond_resched() every 128 register read
> improves the situation. In the common case the, crdump takes ~1.8989s,
> the thread yields the cpu every ~4.51ms. If crdump takes ~5s, the thread
> yields the cpu every ~18.0ms.
>
> Fixes: 8b9d8baae1de ("net/mlx5: Add Crdump support")
> Reviewed-by: Yuanyuan Zhong<yzhong@...estorage.com>
> Signed-off-by: Mohamed Khalfella<mkhalfella@...estorage.com>
> ---
Reviewed-by: Moshe Shemesh <moshe@...dia.com>
Powered by blists - more mailing lists