On 10/18/2019 2:12 PM, Bruce Richardson wrote: > On Fri, Oct 18, 2019 at 12:38:53PM +0100, Ferruh Yigit wrote: >> On 10/18/2019 11:13 AM, Bruce Richardson wrote: >>> On Thu, Oct 17, 2019 at 04:33:59PM +0100, Ferruh Yigit wrote: >>>> On 10/17/2019 2:43 PM, Bruce Richardson wrote: >>>>> On Thu, Oct 17, 2019 at 12:05:56PM +0100, Ferruh Yigit wrote: >>>>>> On 10/17/2019 11:51 AM, Andrew Rybchenko wrote: >>>>>>> On 10/17/19 1:47 PM, Ferruh Yigit wrote: >>>>>>>> On 10/17/2019 11:37 AM, Andrew Rybchenko wrote: >>>>>>>>> On 10/16/19 9:07 PM, Ferruh Yigit wrote: >>>>>>>>>> On 10/16/2019 4:46 PM, Ciara Power wrote: >>>>>>>>>>> Adding promiscuous functions prevents sample applications failing >>>>>>>>>>> when run >>>>>>>>>>> on this virtual PMD. The sample applications call promiscuous >>>>>>>>>>> functions, >>>>>>>>>>> and fail if this function call returns an error, which occurs when >>>>>>>>>>> the >>>>>>>>>>> virtual PMD does not support the promiscuous function being called. >>>>>>>>>>> >>>>>>>>>>> This change will be implemented for all virtual PMDs that currently >>>>>>>>>>> do not >>>>>>>>>>> have existing promiscuous functions. Multicast functions will also >>>>>>>>>>> be >>>>>>>>>>> added for virtual PMDs to prevent sample application breakages here >>>>>>>>>>> also. >>>>>>>>>> +Andrew >>>>>>>>>> >>>>>>>>>> With the some ethdev APIs returning error code, some sample >>>>>>>>>> applications stop >>>>>>>>>> working with virtual interfaces, >>>>>>>>>> >>>>>>>>>> We can, >>>>>>>>>> 1- update sample applications to ignore the errors >>>>>>>>>> 2- Add dummy dev_ops support to (almost all) virtual PMDs (what this >>>>>>>>>> RFC suggests) >>>>>>>>>> >>>>>>>>>> (1) puts us back to before the ethdev APIs updated status, and this >>>>>>>>>> may be wrong >>>>>>>>>> for the physical devices case, so I am for this RFC. >>>>>>>>>> >>>>>>>>>> Only perhaps we can have some common empty function and keep >>>>>>>>>> assigning that one >>>>>>>>>> to reduce the dummy code, what do you think? >>>>>>>>> I don't like the idea to have common empty/dummy functions. >>>>>>>>> If virtual PMD behaves in accordance with enabled promiscuous mode, >>>>>>>>> it should initialize it properly on init: >>>>>>>>> eth_dev->data->promiscuous = 1; >>>>>>>>> If so, if application requires promiscuous mode, attempt to enable >>>>>>>>> will >>>>>>>>> do nothing. If application requires non-promiscuous mode, disable will >>>>>>>>> fail and it is good. >>>>>>>> It is technically correct that we can't disable promiscuous mode in >>>>>>>> virtual PMDs >>>>>>>> but I think mainly we don't really care so it returning error may make >>>>>>>> the >>>>>>>> applications fail/exit unnecessarily with virtual PMDs. >>>>>>> >>>>>>> If I test virtual PMD promiscuous mode, I would prefer enable/disable >>>>>>> callback to say me truth. >>>>>>> >>>>>>> If application really does not care, it should be in the application >>>>>>> code. >>>>>> >>>>>> Application can't change this because they may be caring return result >>>>>> for the >>>>>> physical devices. >>>>>> >>>>>> Up until this release these missing dev_ops in virtual PMDs were silently >>>>>> ignored, now APIs are more strict on this (which is good) but to get >>>>>> close the >>>>>> previous behavior for virtual PMDs we need to relax on these features >>>>>> (like >>>>>> saying success on promiscuous disable although it didn't). >>>>>> >>>>> The other variable here is how often an app is going to request >>>>> promiscuous >>>>> disabling? Given that most ports generally come up in that state anyway, >>>>> and one needs to request enabling it, surely the disable case is >>>>> relatively >>>>> rare? In that case I'd tend to agree with having disabling it returning >>>>> error for vpmds. >>>>> >>>> >>>> Yes disabling most probably rare, but still it will generate an error and >>>> application is failing because of ring PMD promiscuous disable doesn't look >>>> right to me. >>> >>> Well, if an app needs promiscuous mode disabled then having it fail is the >>> right thing to do. If the app doesn't care about promiscuous mode failing, >>> why is it checking the return value at all? >>> >>>> >>>> Perhaps application should differentiate between -ENOTSUP error and >>>> operation >>>> fail error, but that looks to me adding unnecessary complexity to the app. >>>> >>> Again, does the app care or not? It's probably still better to return >>> correct info to the app in all cases, and then let the app decide how best >>> to handle it. >> >> My thinking is, application "cares" about the ethdev API return values, but >> to >> check/test quickly with null/ring PMD perhaps not really care that much abut >> promisc/allmuticast support of these PMDs and let's relax these support of >> virtual PMDs to make life easy. >> >> But eventually the main target is to fix sample applications to run with >> virtual >> PMDs which has been broken in this release. >> Both approach works, >> a) Implement dummy dev_ops to virtual PMDs to report success >> b) Update ethdev APIs to not call dev_ops if the requested configuration is >> already satisfied and change virtual PMDs to report promisc & allmulticast >> enabled by default. (disable still will have same issue) >> >> Is the consensus option (b)? >> >> >> >> btw, the problem exists in high level for the offload support, if the >> application is requesting a specific offload support it fails to run with the >> virtual PMDs since virtual PMDs doesn't support any offloads. Indeed I have >> same >> suggestion for this case too, relax the virtual PMD by claiming it supports >> all >> offloads. Because at least for me when I use those virtual PMDs I don't >> really >> would like to test offloads or the procmisc/allmulticast features ... >> > I really dislike having the drivers lying. It may work in some cases, but > eventually you will hit a problem where an app really does need a feature > and then breaks for the user in mysterious ways when run with a virtual > PMD. Much better to have the vPMD always report the truth to the app, and > let the app worry about whether the app can continue on error or not. > > Final option I'd throw out there, is to allow a vdev parameter to tell the > vpmd it's allowed to lie. That gives an override in case of an app that > can't handle a non-fatal failure.
vdev devarg can work for the offload case, only issue I can see at first glance is code duplication among the PMDs but perhaps that can be solved when issue checked in-dept.
