Possibly we are missing a public API for looking up a queue (or perhaps specifically an IPC queue) by name.
If you have two threads (in same or different processes) that first try to look up an (IPC or global) queue by name and if this fails then creates said IPC queue, you have introduced a race condition. Possibly the second IPC queue create will fail because the named queue exists (EEXIST) and the caller can then understand it does not have to create the queue. So I think we have some synchronization issues to understand and solve when it comes to IPC. Because the primary use case for IPC will be for communication between different programs (processes) so they have no means for synchronization (I would rather not have to use other Linux mechanisms for setting up an IPC channel). On 26 August 2014 11:48, Maxim Uvarov <[email protected]> wrote: > On 08/26/2014 12:56 PM, Taras Kondratiuk wrote: > >> On 08/25/2014 05:37 PM, Maxim Uvarov wrote: >> >>> I pushed here some raw code: >>> https://git.linaro.org/people/maxim.uvarov/odp.git/shortlog/ >>> refs/heads/odp_ipc2 >>> >>> >>> Example here app here: >>> https://git.linaro.org/people/maxim.uvarov/odp.git/blob/ >>> refs/heads/odp_ipc2:/example/fork/odp_fork.c >>> >>> >>> pktio_queue_thread() takes packets from inq queue and put them to ipc >>> queue. >>> >>> ring_thread() takes packets from ipc queue and free this buffer. >>> >> >> I have a few things I didn't understand: >> 1. It is not clear how IPC pktio and its pool is used. Especially on >> 'sender' side. >> > > We create pool somewhere. Then do loop up for it: > pkt_pool = odp_buffer_pool_lookup("packet_pool"); > > Then we want that IPC queue can work with that pool. So we open pktio: > > pktio_ipc_params.type = ODP_PKTIO_TYPE_IPC; > pktio_ipc = odp_pktio_open(NULL, pkt_pool, &pktio_ipc_params); > > And then create queue: > > ipcq_def = odp_queue_create("shared-queue", ODP_QUEUE_TYPE_IPC, &qparam); > > when we place packet to this queue with: > odp_queue_enq(ipcq_def, buf); > > Than everything depend on implementation. So linux-generic (software), > packet data > still will be in pool. odp_buffer_t (uint32_t) value will be added to ring. > > Other process gets odp_buffer_t, and find pointer to that buffer > (odp_buffer_hdr_t*). > > On hw implementation you just place odp_buffer_t to the queue. And if ipc > queue has the same > pool than packet can be delivered to software. If there are different > pools, then packet should be > copied to ipc pool, and after that delivered to hardware. > > The main goal here is we can say that ipc pool can be the same as ingress > pool or it can be different. > Depends on hw configuration. > > > 2. On both sides IPC queue is created. My impression was that it should >> be looked up on one side via pktio or directly via queue name. >> > > Is there way to ask hardware which queues were already created? > > Actually I do look up, but it's hidden in implementation. Idea is that if > queue type is IPC, > then I do loop up if that queue already exist. If exist then I connect to > that queue. If it does > not exist - I create it. > > static void queue_init(queue_entry_t *queue, const char *name, > odp_queue_type_t type, odp_queue_param_t *param) > > case ODP_QUEUE_TYPE_IPC: > queue->s.r = odp_ring_lookup(name); > > if (!queue->s.r) > { > printf("creating new odp ring: %s\n", name); > queue->s.r = odp_ring_create(name, RING_SIZE, > ODP_RING_SHM_PROC); > if (queue->s.r == NULL) { > ODP_ERR("ring create failed\n"); > } > } else > printf("odp ring found!!!\n"); > > Because of we can say that several processes can connect to one queue, > there is no difference > which exactly process will create this queue. So my idea was to stay with > original odp_queue_create() app API. > > Thanks, > Maxim. > > > > > > > _______________________________________________ > lng-odp mailing list > [email protected] > http://lists.linaro.org/mailman/listinfo/lng-odp >
_______________________________________________ lng-odp mailing list [email protected] http://lists.linaro.org/mailman/listinfo/lng-odp
