From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ferruh Yigit Subject: Re: [PATCH v5 00/15] common ethdev linkstatus functions Date: Fri, 19 Jan 2018 16:35:09 +0000 Message-ID: References: <20180116183755.24542-1-stephen@networkplumber.org> <42f4f5a2-008d-c3e8-4c00-ed9ef59065c8@solarflare.com> <27f327e5-5632-69cc-feaa-10cf9384a701@intel.com> <20180117080555.5afd8c78@xeon-e3> <24375588-ec05-9e54-264d-69dcb48f7f16@intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Cc: Andrew Rybchenko , dev@dpdk.org To: Stephen Hemminger , Yuanhan Liu , Maxime Coquelin Return-path: Received: from mga02.intel.com (mga02.intel.com [134.134.136.20]) by dpdk.org (Postfix) with ESMTP id 1884B1B29B for ; Fri, 19 Jan 2018 17:35:12 +0100 (CET) In-Reply-To: <24375588-ec05-9e54-264d-69dcb48f7f16@intel.com> Content-Language: en-US List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" On 1/17/2018 4:18 PM, Ferruh Yigit wrote: > On 1/17/2018 4:05 PM, Stephen Hemminger wrote: >> On Wed, 17 Jan 2018 14:32:17 +0000 >> Ferruh Yigit wrote: >> >>> On 1/17/2018 7:56 AM, Andrew Rybchenko wrote: >>>> On 01/16/2018 09:37 PM, Stephen Hemminger wrote: >>>>> While reviewing drivers, noticed a lot of unnecessary >>>>> duplication of code in drivers for handling the eth_dev link status >>>>> information. While consolidating this, it also became obvious that >>>>> some drivers behave differently for no good reason. >>>>> >>>>> It also was a good chance to introduce atomic exchange primitives >>>>> in EAL because there are other places using cmpset where not >>>>> necessary (such as bonding). >>>>> >>>>> Mostly only compile tested only, don't have all of the hardware >>>>> available (except ixgbe and virtio) to test. >>>>> >>>>> Note: the eth_dev_link_update function return value is inconsistent >>>>> across drivers. Should be changed to be void. >>>> >>>> I would say "link_update" callback return value is inconsistent across >>>> drivers. I'm not sure which direction is right here: make it consistent >>>> or make it void. Also any changes in link information could be >>>> important. As I understand it should not happen without up/down, >>>> but bugs with loss of intermediate transitions are definitely possible. >>>> So, notifying about any changes in link information is definitely safer. >>>> May be not now. >>> >>> Again, why not return previous link status, it is simple enough to prevent >>> inconsistent usage. >>> >>> rte_eth_link_get() already discards the return value, so won't be a problem there. >>> >>> For the cases PMD would like know about link changes, they will need to >>> implement almost same link_update function with a return value, so why not use >>> existing link_update function? >>> >>> Like been in virtio, link_update() used in interrupt handler, and calls a >>> callback process if status changes. When link_update() return status changed to >>> void, I guess they will need to implement another version of the link_update >>> with return and use it. >> >> The interrupt and non-interrupt model are different. > > Yes. But for virtio specific usage: > > virtio_interrupt_handler() > virtio_dev_link_update() == 0 > _rte_eth_dev_callback_process() > > meantime same exact virtio_dev_link_update() used as: > .link_update = virtio_dev_link_update, > > so updating virtio_dev_link_update() to not return status change, will update > logic in virtio_interrupt_handler(), no? I would like to see this patch in, because it is useful and almost done. The concern I mentioned above effects virtio. Can virtio maintainers check if it is OK to get this as it is please? > >> Also the driver internally may want to do something different, this is about >> the return value for dev_ops->link_update. > > Agreed, driver may do something different. And the function needs to be > implemented will be very close to dev_ops->link_update. I thought making > dev_ops->link_update more generic can prevent duplication there. And aligns with > virtio usage.. > >> The code in rte_eth_dev never >> used the return value. The bonding driver was expecting it to work but it >> doesn't. > > Agreed. > >> Anyway drivers shouldn't in general be directly calling other >> devices eth_dev_ops > > I guess now there are a few overlay PMDs does this. >