Re: [PATCH] ethdev: fix push new event

[Thomas Monjalon <thomas@monjalon.net>]

02/06/2022 13:24, lihuisong (C):
> 
> 在 2022/5/30 19:10, Ferruh Yigit 写道:
> > On 5/30/2022 9:28 AM, Thomas Monjalon wrote:
> >> [CAUTION: External Email]
> >>
> >> 28/05/2022 10:53, lihuisong (C):
> >>>
> >>> 在 2022/5/23 22:36, Thomas Monjalon 写道:
> >>>> 23/05/2022 11:51, David Marchand:
> >>>>> On Sat, May 21, 2022 at 8:57 AM Min Hu 
> >>>>> (Connor)<humin29@huawei.com>  wrote:
> >>>>>> From: Huisong Li<lihuisong@huawei.com>
> >>>>>>
> >>>>>> The 'state' in struct rte_eth_dev may be used to update some 
> >>>>>> information
> >>>>>> when app receive these events. For example, when app receives a 
> >>>>>> new event,
> >>>>>> app may get the socket id of this port by calling 
> >>>>>> rte_eth_dev_socket_id to
> >>>>>> setup the attached port. The 'state' is used in 
> >>>>>> rte_eth_dev_socket_id.
> >>>>>>
> >>>>>> If the state isn't modified to RTE_ETH_DEV_ATTACHED before 
> >>>>>> pushing the new
> >>>>>> event, app will get the socket id failed. So this patch moves 
> >>>>>> pushing event
> >>>>>> operation after the state updated.
> >>>>>>
> >>>>>> Fixes: 99a2dd955fba ("lib: remove librte_ prefix from directory 
> >>>>>> names")
> >>>>> A patch moving code is unlikely to be at fault.
> >>>>>
> >>>>>
> >>>>> Looking at the patch which moved those notifications in this point of
> >>>>> the code, the state update was pushed after the notification on
> >>>>> purpose.
> >>>>> See be8cd210379a ("ethdev: fix port probing notification")
> >>>>>
> >>>>>       ethdev: fix port probing notification
> >>>>>
> >>>>>       The new device was notified as soon as it was allocated.
> >>>>>       It leads to use a device which is not yet initialized.
> >>>>>
> >>>>>       The notification must be published after the initialization 
> >>>>> is done
> >>>>>       by the PMD, but before the state is changed, in order to let
> >>>>>       notified entities taking ownership before general availability.
> >>>>>
> >>>>>
> >>>>> Do we need an intermediate state during probing?
> >>>> Possibly. Currently we have only 3 states:
> >>>>      RTE_ETH_DEV_UNUSED
> >>>>      RTE_ETH_DEV_ATTACHED
> >>>>      RTE_ETH_DEV_REMOVED
> >>>>
> >>>> We may add RTE_ETH_DEV_ALLOCATED just before calling
> >>>>      rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_NEW, NULL);
> >>>> Then we would need to check against RTE_ETH_DEV_ALLOCATED
> >>>> in some ethdev functions.
> >>>>
> >>> Hi, Thomas,
> >>>
> >>> Do you mean that we need to modify some funcions like following?
> >>>
> >>> int rte_eth_dev_is_valid_port(uint16_t port_id)
> >>> {
> >>>       if (port_id >= RTE_MAX_ETHPORTS ||
> >>>           (rte_eth_devices[port_id].state != *RTE_ETH_DEV_ALLOCATED*))
> >>>           return 0;
> >
> > Won't this mark ATTACHED devices as invalid?
> Yes, You are right.
> 
> > If the state flow will be as UNUSED -> ALLOCATED -> ATTACHED, above 
> > check should be against 'ATTACHED' I think.

It should validate both ALLOCATED and ATTACHED.


> If these check is against 'ATTACHED', it goes back to the issue this 
> patch mentioned.
> 
> The failsafe PMD applications expect sending event before device state 
> set to 'ATTACHED'.
> But other applications expect the device with 'ATTACHED' state before 
> send event.
> They are in conflict with each other. So we can't solve this issue by 
> adding an
> 'RTE_ETH_DEV_ALLOCATED' state.