Re: [dpdk-dev] [EXT] [PATCH 1/5] examples/l2fwd-event: free resources in case of error

["Jakub Grajciar -X (jgrajcia - PANTHEON TECH SRO at Cisco)" <jgrajcia@cisco.com>]

> -----Original Message-----
> From: Sunil Kumar Kori <skori@marvell.com>
> Sent: Wednesday, June 10, 2020 12:01 PM
> To: Muhammad Bilal <m.bilal@emumba.com>
> Cc: declan.doherty@intel.com; tomasz.kantecki@intel.com; Pavan Nikhilesh
> Bhagavatula <pbhagavatula@marvell.com>; dev@dpdk.org; Jakub Grajciar -X
> (jgrajcia - PANTHEON TECH SRO at Cisco) <jgrajcia@cisco.com>;
> vipin.varghese@intel.com
> Subject: RE: [EXT] [PATCH 1/5] examples/l2fwd-event: free resources in case
> of error
> 
> >-----Original Message-----
> >From: Muhammad Bilal <m.bilal@emumba.com>
> >Sent: Tuesday, June 2, 2020 5:57 PM
> >To: Sunil Kumar Kori <skori@marvell.com>
> >Cc: declan.doherty@intel.com; tomasz.kantecki@intel.com; Pavan
> >Nikhilesh Bhagavatula <pbhagavatula@marvell.com>; dev@dpdk.org;
> >jgrajcia@cisco.com; vipin.varghese@intel.com
> >Subject: Re: [EXT] [PATCH 1/5] examples/l2fwd-event: free resources in
> >case of error
> >
> >On Tue, May 19, 2020 at 2:35 PM Sunil Kumar Kori <skori@marvell.com>
> >wrote:
> >>
> >> >-----Original Message-----
> >> >From: Muhammad Bilal <m.bilal@emumba.com>
> >> >Sent: Tuesday, May 19, 2020 2:25 PM
> >> >To: declan.doherty@intel.com; tomasz.kantecki@intel.com; Pavan
> >> >Nikhilesh Bhagavatula <pbhagavatula@marvell.com>; Sunil Kumar Kori
> >> ><skori@marvell.com>
> >> >Cc: dev@dpdk.org; Muhammad Bilal <m.bilal@emumba.com>
> >> >Subject: [EXT] [PATCH 1/5] examples/l2fwd-event: free resources in
> >> >case of error
> >> >
> >> >External Email
> >> >
> >> >--------------------------------------------------------------------
> >> >-
> >> >- Freeing the resources and call rte_eal_cleanup in case of error
> >> >exit.
> >> >Signed-off-by: Muhammad Bilal <m.bilal@emumba.com>
> >> >---
> >> > examples/l2fwd-event/main.c | 43
> >> >++++++++++++++++++++++++++++++-------
> >> > 1 file changed, 35 insertions(+), 8 deletions(-)
> >> >
> >> >diff --git a/examples/l2fwd-event/main.c
> >> >b/examples/l2fwd-event/main.c index 9593ef11e..442a664e9 100644
> >> >--- a/examples/l2fwd-event/main.c
> >> >+++ b/examples/l2fwd-event/main.c
> >> >@@ -556,13 +556,26 @@ signal_handler(int signum)
> >> >       }
> >> > }
> >> >
> >> >+static void
> >> >+stop_and_close_eth_dev(uint16_t portid) {
> >> >+      RTE_ETH_FOREACH_DEV(portid) {
> >> >+              printf("Closing port %d...", portid);
> >> >+              rte_eth_dev_stop(portid);
> >> >+              rte_eth_dev_close(portid);
> >> >+              printf(" Done\n");
> >> >+      }
> >> >+      rte_eal_cleanup();
> >> >+}
> >> >+
> >> > int
> >> > main(int argc, char **argv)
> >> > {
> >> >       struct l2fwd_resources *rsrc;
> >> >       uint16_t nb_ports_available = 0;
> >> >       uint32_t nb_ports_in_mask = 0;
> >> >-      uint16_t port_id, last_port;
> >> >+      uint16_t port_id = 0;
> >> >+      uint16_t last_port;
> >> >       uint32_t nb_mbufs;
> >> >       uint16_t nb_ports;
> >> >       int i, ret;
> >> >@@ -581,20 +594,26 @@ main(int argc, char **argv)
> >> >
> >> >       /* parse application arguments (after the EAL ones) */
> >> >       ret = l2fwd_event_parse_args(argc, argv, rsrc);
> >> >-      if (ret < 0)
> >> >+      if (ret < 0) {
> >> >+              stop_and_close_eth_dev(port_id);
> >> >               rte_panic("Invalid L2FWD arguments\n");
> >> >+      }
> >> >
> >Yes you are right we should use rte_exit instead of rte_panic, as
> >rte_exit internally calls rte_eal_cleanup function.
> >> IMO, up to this point only eal_init is done so rte_eal_cleanup will
> >> be
> >sufficient for this.
> >> Also another way to handle this, use rte_exit instead rte_panic.
> >> rte_exit internally calls rte_eal_cleanup. Refer l2fwd.
> >>
> >> Also I think, it is better to release the relevant resources on error.
> >Here I'm solving the problem reported in bugzilla id 437. The bug was
> >that if we use --vdev=net_memif with l2fwd application (and with its
> >other variants) then a socket is created by memif PMD, after
> >rte_eal_init function has run successfully. And if an error occurs then
> >the application exits without freeing the resources (socket). On
> >running the application 2nd time we get an error of "socket already exists".
> >As in the previous version of patch "
> >https://urldefense.proofpoint.com/v2/url?u=http-
> >3A__patches.dpdk.org_patch_70081_&d=DwIBaQ&c=nKjWec2b6R0mOyPa
> z7x
> >tfQ&r=dXeXaAMkP5COgn1zxHMyaF1_d9IIuq6vHQO6NrIPjaE&m=XKcRI2e7
> sMv
> >Y0nGabnBQl_Q8meL03FXFAjeNGdCV91A&s=TKq1J0W3QbnkeuG4c63payD
> Wc
> >Pc4zTg4DumA95RVzwg&e=  " it was recommended to clean the resources
> when
> >an error occurs.
> >
> >Here only using rte_eal_cleanup() is not solving the problem as using
> >memif PMD is creating a socket and it's only cleaned when
> >rte_eth_dev_close(portid) function is called. so that's why using it
> >along with rte_exit or rte_panic.
> 
> 
> Understood but I am only thinking from user's perspective that user didn't do
> rte_eth_dev_configure and related APIs so closing the device using
> rte_eth_dev_close does not look good.
> May be other's (eal and memif PMD owners) can suggest something better.
> Please redirect this query to them for suggestions also.
> 

So if calling ret_eth_dev_close() makes sense after rte_eth_dev_configure() the socket initialization can be moved to the configure function. One more thought: struct rte_vdev_driver holds .probe and .remove functions. Shouldn't .remove be called before exiting if .probe was called? This would close the socket in case of memif PMD.

> >>
> >> >       printf("MAC updating %s\n", rsrc->mac_updating ? "enabled" :
> >> >                       "disabled");
> >> >
> >> >       nb_ports = rte_eth_dev_count_avail();
> >> >-      if (nb_ports == 0)
> >> >+      if (nb_ports == 0) {
> >> >+              stop_and_close_eth_dev(port_id);
> >> >               rte_panic("No Ethernet ports - bye\n");
> >> >+      }
> >> >
> >> Same as above.
> >>
> >> >       /* check port mask to possible port mask */
> >> >-      if (rsrc->enabled_port_mask & ~((1 << nb_ports) - 1))
> >> >+      if (rsrc->enabled_port_mask & ~((1 << nb_ports) - 1)) {
> >> >+              stop_and_close_eth_dev(port_id);
> >> >               rte_panic("Invalid portmask; possible (0x%x)\n",
> >> >                       (1 << nb_ports) - 1);
> >> >+      }
> >> >
> >> Same as above.
> >>
> >> >       if (!rsrc->port_pairs) {
> >> >               last_port = 0;
> >> >@@ -621,8 +640,10 @@ main(int argc, char **argv)
> >> >                       rsrc->dst_ports[last_port] = last_port;
> >> >               }
> >> >       } else {
> >> >-              if (check_port_pair_config(rsrc) < 0)
> >> >+              if (check_port_pair_config(rsrc) < 0) {
> >> >+                      stop_and_close_eth_dev(port_id);
> >> >                       rte_panic("Invalid port pair config\n");
> >> >+              }
> >> >       }
> >> >
> >> >       nb_mbufs = RTE_MAX(nb_ports * (RTE_TEST_RX_DESC_DEFAULT +
> >@@
> >> >-634,12 +655,16 @@ main(int argc, char **argv)
> >> >       rsrc->pktmbuf_pool = rte_pktmbuf_pool_create("mbuf_pool",
> >> >                       nb_mbufs, MEMPOOL_CACHE_SIZE, 0,
> >> >                       RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id());
> >> >-      if (rsrc->pktmbuf_pool == NULL)
> >> >+      if (rsrc->pktmbuf_pool == NULL) {
> >> >+              stop_and_close_eth_dev(port_id);
> >> >               rte_panic("Cannot init mbuf pool\n");
> >> >+      }
> >> >
> >> >       nb_ports_available = l2fwd_event_init_ports(rsrc);
> >> >-      if (!nb_ports_available)
> >> >+      if (!nb_ports_available) {
> >> >+              stop_and_close_eth_dev(port_id);
> >> >               rte_panic("All available ports are disabled. Please
> >> >set portmask.\n");
> >> >+      }
> >> >
> >> >       /* Configure eventdev parameters if required */
> >> >       if (rsrc->event_mode)
> >> >@@ -659,9 +684,11 @@ main(int argc, char **argv)
> >> >                       continue;
> >> >
> >> >               ret = rte_eth_dev_start(port_id);
> >> >-              if (ret < 0)
> >> >+              if (ret < 0) {
> >> >+                      stop_and_close_eth_dev(port_id);
> >> >                       rte_panic("rte_eth_dev_start:err=%d,
> >> >port=%u\n", ret,
> >> >                                 port_id);
> >> >+              }
> >> >       }
> >> >
> >> >       if (rsrc->event_mode)
> >> >--
> >> >2.17.1
> >>