On 3/16/2020 9:18 PM, Laurence Oberman wrote:
On Mon, 2020-03-16 at 09:30 +0200, Leon Romanovsky wrote:On Mon, Mar 16, 2020 at 09:21:40AM +0200, Leon Romanovsky wrote:On Sun, Mar 15, 2020 at 05:56:17PM -0400, Laurence Oberman wrote:On Sun, 2020-03-15 at 17:01 -0400, Laurence Oberman wrote:On Sun, 2020-03-15 at 22:40 +0200, Max Gurtovoy wrote:On 3/15/2020 8:36 PM, Laurence Oberman wrote:On Sun, 2020-03-15 at 20:20 +0200, Max Gurtovoy wrote:On 3/15/2020 7:59 PM, Laurence Oberman wrote:On Sun, 2020-03-15 at 18:47 +0200, Max Gurtovoy wrote:On 3/14/2020 11:30 PM, Laurence Oberman wrote:Hello Bart, Leon and MaxHi Laurence, thanks for the great analysis and the fast response !!Max had reached out to me to test a new set of patches for SRQ. I had not tested upstream ib_srpt on an LIO target for quite a while, only ib_srp client tests had been run of late. During a baseline test before applying Max's patches it was apparent that something had broken ib_srpt connections within LIO target since 5.5. Note thet ib_srp client connectivity with the commit functions fine, it's just the target that breaks with this commit. After a long bisect this is the commit that seems to break it. While it's not directly code in ib_srpt, its code in mlx5 vport ethernet connectivity that then breaks ib_srpt connectivity over mlx5 IB RDMA with LIO.I was able to connect in loopback and also from remote initiator with this commit. So I'm not sure that this commit is broken. I used Bart's scripts to configure the target and to connect to it in loopback (after some modifications for the updated kernel/sysfs/configfs interface). I did see an issue to connect from remote initiator, but after reloading openibd in the initiator side I was able to connect. So I suspect you had the same issue - that also should be debugged.I will let Leon and others decide but reverting the below commit allows SRP connectivity to an LIO target to work again.I added prints to "mlx5_core_modify_hca_vport_context" function and found that we don't call it in "pure" mlx5 mode with PFs. Maybe you can try it too... I was able to check my patches on my system and I'll send them soon. Thanks again Laurence and Bart.Max, I will test your new patches once we have a decision on this. Client Linux ibclient.lab.eng.bos.redhat.com 5.6.0-rc5+ #1 SMP Thu Mar 12 16:58:19 EDT 2020 x86_64 x86_64 x86_64 GNU/Linux Server with reverted commit Linux fedstorage.bos.redhat.com 5.6.0-rc5+ #1 SMP Sat Mar 14 16:39:35 EDT 2020 x86_64 x86_64 x86_64 GNU/Linux commit ab118da4c10a70b8437f5c90ab77adae1835963e Author: Leon Romanovsky <leonro@xxxxxxxxxxxx> Date: Wed Nov 13 12:03:47 2019 +0200 net/mlx5: Don't write read-only fields in MODIFY_HCA_VPORT_CONTEXT command The MODIFY_HCA_VPORT_CONTEXT uses field_selector to mask fields needed to be written, other fields are required to be zero according to the HW specification. The supported fields are controlled by bitfield and limited to vport state, node and port GUIDs. Signed-off-by: Leon Romanovsky < leonro@xxxxxxxxxxxx> Signed-off-by: Saeed Mahameed < saeedm@xxxxxxxxxxxx diff --git a/drivers/net/ethernet/mellanox/mlx5/core/vport.c b/drivers/net/ethernet/mellanox/mlx5 index 30f7848..1faac31f 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/vport.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/vport.c @@ -1064,26 +1064,13 @@ int mlx5_core_modify_hca_vport_context(struct mlx5_core_dev *dev, ctx = MLX5_ADDR_OF(modify_hca_vport_context_in, in, hca_vport_context); MLX5_SET(hca_vport_context, ctx, field_select, req-field_select);- MLX5_SET(hca_vport_context, ctx, sm_virt_aware, req-sm_virt_aware);- MLX5_SET(hca_vport_context, ctx, has_smi, req-has_smi);- MLX5_SET(hca_vport_context, ctx, has_raw, req-has_raw);- MLX5_SET(hca_vport_context, ctx, vport_state_policy, req-policy);- MLX5_SET(hca_vport_context, ctx, port_physical_state, req-phys_state);- MLX5_SET(hca_vport_context, ctx, vport_state, req-vport_state);- MLX5_SET64(hca_vport_context, ctx, port_guid, req-port_guid);- MLX5_SET64(hca_vport_context, ctx, node_guid, req-node_guid);- MLX5_SET(hca_vport_context, ctx, cap_mask1, req-cap_mask1);- MLX5_SET(hca_vport_context, ctx, cap_mask1_field_select, req-cap_mask1_perm);- MLX5_SET(hca_vport_context, ctx, cap_mask2, req-cap_mask2);- MLX5_SET(hca_vport_context, ctx, cap_mask2_field_select, req-cap_mask2_perm);- MLX5_SET(hca_vport_context, ctx, lid, req-lid);- MLX5_SET(hca_vport_context, ctx, init_type_reply, req-init_type_reply);- MLX5_SET(hca_vport_context, ctx, lmc, req-lmc);- MLX5_SET(hca_vport_context, ctx, subnet_timeout, req-subnet_timeout);- MLX5_SET(hca_vport_context, ctx, sm_lid, req-sm_lid);- MLX5_SET(hca_vport_context, ctx, sm_sl, req-sm_sl);- MLX5_SET(hca_vport_context, ctx, qkey_violation_counter, req-qkey_violation_counter);- MLX5_SET(hca_vport_context, ctx, pkey_violation_counter, req-pkey_violation_counter);+ if (req->field_select & MLX5_HCA_VPORT_SEL_STATE_POLICY) + MLX5_SET(hca_vport_context, ctx, vport_state_policy, + req->policy); + if (req->field_select & MLX5_HCA_VPORT_SEL_PORT_GUID) + MLX5_SET64(hca_vport_context, ctx, port_guid, req-port_guid);+ if (req->field_select & MLX5_HCA_VPORT_SEL_NODE_GUID) + MLX5_SET64(hca_vport_context, ctx, node_guid, req-node_guid);err = mlx5_cmd_exec(dev, in, in_sz, out, sizeof(out)); ex: kfree(in);Hi Max Re: " So I'm not sure that this commit is broken. .. .. I added prints to "mlx5_core_modify_hca_vport_context" function and found that we don't call it in "pure" mlx5 mode with PFs. Maybe you can try it too... " The thing is without this commit we connect immediately, no delay no issue and I am changing nothing else other than reverting here. So this clearly has a bearing directly on the functionality. I will look at adding more debug, but with this commit in there is nor evidence even of an attempt to connect and fail. Its silently faling.please send me all the configuration steps you run after booting the target + steps in the initiator (can be also in attached file). I'll try to follow this. Btw, did you try loopback initiator ? -Max.Regards LaurenceHi Max Did not try loopback because here we have actual physical connectity as that is what our customers use. Connected back to back with MLX5 cx4 HCA dual ports at EDR 100 Thi sis my standard configuration used for all upstream and Red Hat kernel testing. Reboot server and client and then first prepare server Server ---------- the prepare.sh script run is after boot on the server #!/bin/bash ./load_modules.sh ./create_ramdisk.sh targetcli restoreconfig # Set the srp_sq_size for i in /sys/kernel/config/target/srpt/0xfe800000000000007cfe900300 726e4e /sys/kernel/config/target/srpt/0xfe800000000000007cfe900300 726e4f do echo 16384 > $i/tpgt_1/attrib/srp_sq_size done [root@fedstorage ~]# cat load_modules.sh #!/bin/bash modprobe mlx5_ib modprobe ib_srpt modprobe ib_srp modprobe ib_umad [root@fedstorage ~]# cat ./create_ramdisk.sh #!/bin/bash mount -t tmpfs -o size=130g tmpfs /mnt cd /mnt for i in `seq 1 30`; do dd if=/dev/zero of=block-$i bs=1024k count=4000 ; done Client -------- Once server is ready Run ./start_opensm.sh on client (I sont use the SM on a switch as we are back to back) [root@ibclient ~]# cat ./start_opensm.sh #!/bin/bash rmmod ib_srpt opensm -F opensm.1.conf & opensm -F opensm.2.conf & I will semail the conf only to you as well as the targecli config as th eout is long. Then run start_srp.sh [root@ibclient ~]# cat ./start_srp.sh run_srp_daemon -V -f /etc/ddn/srp_daemon.conf -R 30 -T 10 -t 7000 -ance -i mlx5_0 -p 1 1>/root/srp1.log 2>&1 & run_srp_daemon -V -f /etc/ddn/srp_daemon.conf -R 30 -T 10 -t 7000 -ance -i mlx5_1 -p 1 1>/root/srp2.log 2>&1 & [root@ibclient ~]# cat /etc/ddn/srp_daemon.conf a queue_size=128,max_cmd_per_lun=32,max_sect=32768I see that you're link is IB. I was working on RoCE link layer with rdma_cm. I'll try to find some free IB setup tomorrow in my lab.. Can you try login using rdma_cm ? need to load ib_ipoib for that in IB network. I'm trying to understand whether it's related to the link layer. p.s. I think the target configuration file didn't arrive.Max, Yes, I am, working primarily with SCSI over RDMA Protocol with Infiniband HCA's in IB mode. I am not using ROCE. Also lets not forget this is a target only issue, the latest 5.6 kernel runs untouched with no issues on the initiator when the target runs either 5.4 or 5.6 with the revert. It would run fine with the target on 5.5 as well if I reverted the commit on 5.5 too. I am not able at this time to switch these adapters to Ethernet mode and ROCE The weird thing is the failure is completely silent so something in the Link layer with IB has to failing early. Thje other strange observation is that the IB interfaces come up with no issue. I will try add some debug after reboot into the failing kernel. Regards LaurenceMax, Rupesh in cc here will be testing latest upstream on a client/server configuration with ROCE and report back here on if he sees a similar issue with the LIO target with that commit. I will continue working on the IB srpt issue by adding some debug. If you think about anything related to the commit let me know.Laurence, As I said above, the most chances are that I removed some RW initialization that wasn't protected by field_select and wasn't marked in our PRM as RW field. The question is which one.I think that I know what is missing. Can you please try this patch? diff --git a/drivers/net/ethernet/mellanox/mlx5/core/vport.c b/drivers/net/ethernet/mellanox/mlx5/core/vport.c index 1faac31f74d0..23f879da9104 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/vport.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/vport.c @@ -1071,6 +1071,9 @@ int mlx5_core_modify_hca_vport_context(struct mlx5_core_dev *dev, MLX5_SET64(hca_vport_context, ctx, port_guid, req-port_guid);if (req->field_select & MLX5_HCA_VPORT_SEL_NODE_GUID) MLX5_SET64(hca_vport_context, ctx, node_guid, req-node_guid);+ MLX5_SET(hca_vport_context, ctx, cap_mask1, req->cap_mask1); + MLX5_SET(hca_vport_context, ctx, cap_mask1_field_select, + req->cap_mask1_perm); err = mlx5_cmd_exec(dev, in, in_sz, out, sizeof(out)); ex: kfree(in);ThanksRegards LaurenceLeon, That patch does resolve the issue. Tested with CX4 with mlx5 with SRP over IB to LIO target. Please add a fixes to that commit tag with this one when you send it.
Great Laurence ! thanks for the effort. Eventually it was helpful that we both run on different link layers :)
Reviewed-by: Laurence Oberman <loberman@xxxxxxxxxx> Tested-by: Laurence Oberman <loberman@xxxxxxxxxx> Thanks very much Laurence