On 13.08.2018 05:50, Lu, Wenzhuo wrote:
Hi Thomas,
-----Original Message-----
From: Thomas Monjalon [mailto:tho...@monjalon.net]
Sent: Wednesday, August 1, 2018 11:37 PM
To: Lu, Wenzhuo <wenzhuo...@intel.com>; Andrew Rybchenko
<arybche...@solarflare.com>; Yigit, Ferruh <ferruh.yi...@intel.com>
Cc: dev@dpdk.org
Subject: Re: [dpdk-dev] [PATCH v2] ethdev: fix device info getting
16/07/2018 03:58, Lu, Wenzhuo:
Hi Andrew,
-----Original Message-----
From: dev [mailto:dev-boun...@dpdk.org] On Behalf Of Lu, Wenzhuo
Sent: Monday, July 16, 2018 9:08 AM
To: Andrew Rybchenko <arybche...@solarflare.com>; dev@dpdk.org
Cc: Yigit, Ferruh <ferruh.yi...@intel.com>; Thomas Monjalon
<tho...@monjalon.net>
Subject: Re: [dpdk-dev] [PATCH v2] ethdev: fix device info getting
Hi Andrew,
-----Original Message-----
From: Andrew Rybchenko [mailto:arybche...@solarflare.com]
Sent: Friday, July 13, 2018 4:03 PM
To: Lu, Wenzhuo <wenzhuo...@intel.com>; dev@dpdk.org
Cc: Yigit, Ferruh <ferruh.yi...@intel.com>; Thomas Monjalon
<tho...@monjalon.net>
Subject: Re: [dpdk-dev] [PATCH v2] ethdev: fix device info getting
Hi, Wenzhuo,
I'm sorry, but I have more even harder questions than the previous one.
This questions are rather generic and mainly to ethdev maintainers.
On 13.07.2018 05:42, Wenzhuo Lu wrote:
The device information cannot be gotten correctly before the
configuration is set. Because on some NICs the information has
dependence on the configuration.
Thinking about it I have the following question. Is it valid
behaviour of the dev_info if it changes after configuration?
I always thought that the primary goal of the dev_info is to
provide information to app about device capabilities to allow app
configure device and queues correctly. Now we see the case when
dev_info changes on configure. May be it is acceptable, but it is
really suspicious. If we accept it, it should be documented.
May be dev_info should be split into parts: part which is
persistent and part which may depend on device configuration.
As I remember, the similar discussion has happened :) I've raised
the similar suggestion like this. But we don’t make it happen.
The reason is, you see, this is the rte layer's behavior. So the
user doesn't have to know it. From APP's PoV, it inputs the
configuration, it calls this API "rte_eth_dev_configure". It doesn't
know the configuration is copied before getting the info or not.
So, to my opinion, we can still keep the behavior. We only need to
split it into parts when we do see the case that cannot make it.
Maybe I talked too much about the patch. Think about it again. Your
comments is about how to use the APIs, rte_eth_dev_info_get,
rte_eth_dev_configure. To my opinion, rte_eth_dev_info_get is just to get
the info. It can be called anywhere, before configuration or after. It's
reasonable the info changes with the configuration changing.
But we do have something missing, like, rte_eth_dev_capability_get which
should be stable. APP can use this API to get the necessary info before
configuration.
A question, maybe a little divergent thinking, that APP should have some
intelligence to handle the capability automatically. So getting the capability
is not so good and effective, looks like we still need the human involvement.
Maybe that the reason currently we suppose APP know the capability from
the paper copies, examples...
I am not sure to understand all the sentences.
But I agree that we should take a decision about the stability of these infos.
Either infos cannot change after probing, or we must document that the app
must request infos regularly (when?).
Sorry, I missed this mail.
I have the concern that different NICs have different behavior. One info can be
stable on a NIC but dynamic on another. Considering this, we may better not
splitting the rte_eth_dev_info_get to 2 APIs. And comparing with handling this
in rte layer, maybe we can let every NIC has its own decision.
I have an idea. Maybe we can add a parameter for potential dynamic fields. Like,
Changing
uint16_t nb_rx_queues;
to
struct nb_rx_queues {
uint16_t value;
bool stable;
}
May be it is just very bad example, but as I understand nb_rx_queues is
mainly required to configure the device properly. Or should app
configure, get new value, reconfigure again, get new value and so on and
stop when previous is equal to the new one. Yes, I dramatise and it
sounds really bad. In any case it would over-complicate interface and no
single app will do it correctly.
Stable dev_info is simple. If there are real cases when something can't
be stable (and may be recommended Rx/Tx ring sizes is good example, it
should at least documented in dev_info structure description or may be
moved to separate API.
By default, the stable is false. Then every NIC can maintain its own behavior.
Some fileds that must be stable can be left unchanged, like, driver_name,
max_rx_queues.
As this patch is just reversing a bad commit to fix a bug, if my idea sounds
good or worth discussing, I can send another RFC mail for it.
