Re: [External] Re: [PATCH] mlx5: fix race at mlx5_dev_close

[贺鹏 <hepeng.0320@bytedance.com>]

[-- Attachment #1: Type: text/plain, Size: 1298 bytes --]

It's 1ms, not 1 second. 
It's a workaround, just to provide a fast and dirty fix for someone who needs this.


From: "Stephen Hemminger"<stephen@networkplumber.org>
> Date:  Tue, Oct 8, 2024, 01:54
> Subject:  [External] Re: [PATCH] mlx5: fix race at mlx5_dev_close
> To: "hepeng"<hepeng.0320@bytedance.com>
> Cc: <dev@dpdk.org>
> On Thu, 11 Apr 2024 14:17:40 +0800 
> hepeng <hepeng.0320@bytedance.com> wrote: 
> 
> 
> > From: "hepeng.0320" <hepeng.0320@bytedance.com> 
> > 
> > mlx5_dev_close currently will set priv->sh->port[priv->dev_port - 
> > 1].nl_ih_port_id to RTE_MAX_ETHPORTS to avoid mlx5_dev_interrupt_nl_cb 
> > to use the port's dev_private, because later the rte_eth_dev_close 
> > will free the dev_private and set the pointer to NULL. 
> > 
> > However, since mlx5_dev_interrupt_nl_cb is running in another thread, 
> > I think the race still exists. So perhaps an easy fix is to wait for 
> > 1ms to avoid this race. 
> > 
> > Signed-off-by: hepeng.0320 <hepeng.0320@bytedance.com> 
> 
> 
> Not the pest way to handle this. Adding a one second delay on shutdown 
> hurts some availability scenarios. Looks like mlx5 needs a more coordinated 
> shutdown to be safe; adding big delays is not the correct fix.
> 
> 
> 
> 
> 
> 

[-- Attachment #2: Type: text/html, Size: 8870 bytes --]