Re: [PATCH] ethdev: fix push new event

[Ferruh Yigit <ferruh.yigit@amd.com>]

On 5/23/2022 10:51 AM, David Marchand wrote:
> [CAUTION: External Email]
> 
> 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?
> 

As David highlight above commit, sending event before setting state to 
'RTE_ETH_DEV_ATTACHED' seems done intentionally.

I guess this is done mainly because of failsafe PMD, which registers the 
'RTE_ETH_EVENT_NEW' event and take sub-device ownership.

The ethdev iterate macros, like 'RTE_ETH_FOREACH_DEV', uses the 
condition that device is NOT owned and device state is NOT UNUSED.
Sending event before device state set to 'ATTACHED' lets failsafe to own 
the device so that it won't be visible to application at all via 
mentioned macros.


But I think the need for the libraries and application can be different, 
unlike failsafe PMD application may want notification on fully probed 
device, as the commit log of this patch mentions.

Perhaps solution can be as David mentioned, to create two events, one 
for libraries (failsafe) and other for applications.