Hi, Thanks for reply:) Actually when configure second target, I execute targetcli /iscsi create iqn.2016-08.nokia.lab:omu-0 which will automatically create the 0.0.0.0 3260 listening socket for me. But I think we do not need it at all, because we will delete it later, after that configure own IP(like below 10.180.180.222 ), so I am wondering whether is there any other iscsi target configuration can avoid this unnecessary issue. " targetcli /iscsi/iqn.2016-08.nokia.lab:chu-0/tpg1/portals delete 0.0.0.0 3260 targetcli /iscsi/iqn.2016-08.nokia.lab:chu-0/tpg1/portals create 10.180.180.222 Using default IP port 3260 " Thanks again. Brs, Jin Ke -----Original Message----- From: Nicholas A. Bellinger [mailto:nab@xxxxxxxxxxxxxxx] Sent: Sunday, October 30, 2016 7:58 AM To: Jin, Ke (Nokia - CN/Hangzhou) <ke.jin@xxxxxxxxx> Cc: target-devel@xxxxxxxxxxxxxxx; linux-scsi@xxxxxxxxxxxxxxx; Xiong, Li (Nokia - CN/Hangzhou) <li.xiong@xxxxxxxxx>; Tom Herbert <tom@xxxxxxxxxxxxxxx>; Andy Grover <agrover@xxxxxxxxxx> Subject: Re: iscsi target kernel error log support Hi Jin, Adding Tom CC' for SO_REUSEPORT and Andy CC' for INADDR_ANY + targetcli. On Tue, 2016-10-18 at 01:25 +0000, Jin, Ke (Nokia - CN/Hangzhou) wrote: > Hi, > Sorry to disturb you. > > I am using targetcli to configure the iscsi target on my linux host. I > find one issue, could you kindly review it and give some comments to > me. > > Issue: (linux kernel 4.4 stable) > > Configure first target: > 10/14/16 17:07:00.538417942: targetcli /iscsi create iqn.2016-08.nokia.lab:chu-0 > Created target iqn.2016-08.nokia.lab:chu-0. > Created TPG 1. > Global pref auto_add_default_portal=true > Created default portal listening on all IPs (0.0.0.0), port 3260. > 10/14/16 17:07:00.650688229: targetcli /backstores/block create iblock_0 dev=/dev/sda > Created block storage object iblock_0 using /dev/sda. > 10/14/16 17:07:00.772435404: targetcli /iscsi/iqn.2016-08.nokia.lab:chu-0/tpg1/luns create /backstores/block/iblock_0 1 > Created LUN 1. > 10/14/16 17:07:00.894355883: targetcli /iscsi/iqn.2016-08.nokia.lab:chu-0/tpg1/portals delete 0.0.0.0 3260 > Deleted network portal 0.0.0.0:3260 > 10/14/16 17:07:01.012410786: targetcli /iscsi/iqn.2016-08.nokia.lab:chu-0/tpg1/portals create 10.180.180.111 > Using default IP port 3260 > Created network portal 10.180.180.111:3260. > 10/14/16 17:07:01.130478681: targetcli /iscsi/iqn.2016-08.nokia.lab:chu-0/tpg1/acls create iqn.2016-08.nokia.lab:initiator.chu-0-0 > Created Node ACL for iqn.2016-08.nokia.lab:initiator.chu-0-0 > Created mapped LUN 1. > > Then I configure second target on same host. > targetcli /iscsi create iqn.2016-08.nokia.lab:omu-0 > Created target iqn.2016-08.nokia.lab:omu. > Created TPG 1. > Default portal not created, TPGs within a target cannot share ip:port. > Then I check the kernel log with dmesg, there is one error log > “kernel_bind() failed: -98”. > > So AFAIK, the failure of INADDR_ANY portal binding when !INADDR_ANY portals exist was the original expected behavior. > > Then I begin to check the iscsi kernel module code. linux-stable > \drivers\target\iscsi\ iscsi_target_login.c > int iscsit_setup_np() > { > ret = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, > (char *)&opt, sizeof(opt)); > } > I think SO_REUSEADDR is not enough in this case (there have two > targets configured), when targetcli /iscsi create > iqn.2016-08.nokia.lab:omu-0, it will create listening socket with > 0,0,0,0:3260, because it already have 10.180.180.111:3260 listening > sockets there in first target, so this cause the kernel_bind() failed: > -98”. So I add SO_REUSEPORT also. It works. > int iscsit_setup_np() > { > ret = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, > (char *)&opt, sizeof(opt)); > ret = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEPORT, > (char *)&opt, sizeof(opt)); > } > > How do you think, do you think this is the kernel bug or just my > configuration issue, or you already have this patch? > At iscsi-target application level, we're using a iscsi_np kthread listener per ipv4 / ipv6 addr + port. (eg: one listener per iscsi network portal) The network portal can be shared across N target endpoints in /sys/kernel/config/target/iscsi/$IQN/$TPGT/np/*, but the actual listening + accept for each ipv4 / ipv6 addr + port is still done from a single kthread context. So AFAICT for your use-case, adding SO_REUSEPORT to allow concurrent INADDR_ANY + !INADDR_ANY network portals should be OK, at least from an iscsi-target perspective. That is, iscsi-target would consider each as a separate network portal, allowing INADDR_ANY + !INADDR_ANY accepts to be serviced across multiple iscsi_np kthreads. There should not be an inherent issue allowing this, but it would useful to understand if there are any other possible ramifications before enabling by default. ��.n��������+%������w��{.n�����{������ܨ}���Ơz�j:+v�����w����ޙ��&�)ߡ�a����z�ޗ���ݢj��w�f
