Hi, See comments inline
-----Original Message----- From: Bala Manoharan [mailto:bala.manoha...@linaro.org] Sent: Tuesday, October 03, 2017 21:13 To: Bill Fischofer <bill.fischo...@linaro.org> Cc: Liron Himi <lir...@marvell.com>; lng-odp@lists.linaro.org Subject: Re: [EXT] Re: [lng-odp] odp_pktin_queue_param_t - classifier_enable & num_queues Yes. This is an implementation bug. Will fix the bug. [L.H.] As part of the bug ticket I already opened? Or should I open new one? Minor correction, when classifier_enabled is set the current design is to create all the queues in a single shot rather than as needed since it will create network jitter if queues are created at random. Regards, Bala On 3 October 2017 at 09:59, Bill Fischofer <bill.fischo...@linaro.org> wrote: > Thanks Liron, I've checked the code and that seems to be another > issue. The intent is that when classifier_enable is set the ODP > classifier creates queues as needed. [L.H.] Yes, this is true, when working with classifier user should create queue and attach it to cos which in the end will be related to some pktio. But the current scheduler mechanism expect that pktio will have at least one “internal” input queue that will be registered with and by doing that will be able to be triggered by the scheduler. See ‘sched_cb_pktin_poll’. So I wonder if I can trick it and create one internal input queue and use it to represent all my HW queues. Seems like I can do it, but I saw that 'odp_pktin_event_queue' can be called if PKTIO is in schedule mode, I wonder it can also be called if classifier is enable? > > +cc Bala for additional commentary. > > On Tue, Oct 3, 2017 at 8:48 AM, Liron Himi <lir...@marvell.com> wrote: >> >> Hi Bill, >> >> >> >> I opened ‘https://bugs.linaro.org/show_bug.cgi?id=3289’ for this issue. >> >> >> >> I think there is another issue related to this topic in case PKTIO >> was configure in schedule mode. >> >> The ‘start’ function assumes ‘num_in_queue’ is bigger than ‘0’ but in >> the classifier_enable case it is ‘0’ so pktio is register with no >> real indexes, actually in ‘index’ array there is garbage. (see code >> below) >> >> >> >> If I want to have a local fix, what do you think should be the number >> of queues? ‘max_input_queues’ in capabilities ? >> >> >> >> Regards, >> >> Liron >> >> >> >> >> >> int odp_pktio_start(odp_pktio_t hdl) >> >> { >> >> … >> >> mode = entry->s.param.in_mode; >> >> >> >> if (mode == ODP_PKTIN_MODE_SCHED) { >> >> unsigned i; >> >> unsigned num = entry->s.num_in_queue; >> >> int index[num]; >> >> >> >> for (i = 0; i < num; i++) { >> >> index[i] = i; >> >> >> >> if >> (entry->s.in_queue[i].queue == ODP_QUEUE_INVALID) { >> >> >> ODP_ERR("No input queue\n"); >> >> >> return -1; >> >> } >> >> } >> >> >> >> >> sched_fn->pktio_start(pktio_to_id(hdl), >> num, index); >> >> } >> >> >> >> return res; >> >> } >> >> From: Bill Fischofer [mailto:bill.fischo...@linaro.org] >> Sent: Monday, October 02, 2017 14:52 >> To: Liron Himi <lir...@marvell.com> >> Cc: lng-odp@lists.linaro.org >> Subject: [EXT] Re: [lng-odp] odp_pktin_queue_param_t - >> classifier_enable & num_queues >> >> >> >> External Email >> >> ________________________________ >> >> Thanks Liron. This appears to be a bug in the implementation. It's >> only an error for num_queues to be zero if hash_enabled is set. >> >> >> >> On Mon, Oct 2, 2017 at 6:02 AM, Liron Himi <lir...@marvell.com> wrote: >> >> Hi, >> >> According to the API if 'classifier_enable' is enabled than 'num_queues' >> and 'queue_param' should be ignored. >> But looking at linux-generic implementation those parameters are >> always being checked regardless of 'classifier_enable' value. >> 'classifier_enable' is actually ignore at linux-generic implementation. >> The main problem is if 'num_queues' is '0' than 'odp_pktin_queue_config' >> return an error although it can be '0' if 'classifier_enable' is enabled. >> >> Regards, >> Liron >> >> > >
