On the Linux 6.9-rc1 kernel there is a performance regression for NFS file transfers when Connected IPoIB mode is enabled. The network switch is OPA (Omnipath Architecture).

The most recent good commit in my bisection was the v6.8 mainline kernel (e8f897f4afef0031fe618a8e94127a0934896aba). Bisecting from v6.8 to v6.9-rc1 showed me that "[e084ee673c77cade06ab4c2e36b5624c82608b8c] svcrdma: Add Write chunk WRs to the RPC's Send WR chain" was indeed the culprit of the regression.


Here are the steps I ran to reproduce the issue:
1. Establish IPoIB Connected Mode on both client and host nodes:
"echo connected > /sys/class/net/ibs785/mode"


2. Start an NFS server on the host node:
"systemctl start opafm
sleep 10
systemctl start nfs-server
modprobe svcrdma
echo "rdma 20049" > /proc/fs/nfsd/portlist
mkdir -p /mnt/nfs_test
mount -t tmpfs -o size=4096M tmpfs /mnt/nfs_test
sudo exportfs -o fsid=0,rw,async,insecure,no_root_squash 192.168.2.0/255.255.255.0:/mnt/nfs_test_testrun/"


3. Ready the client node:
"mkdir -p /mnt/nfs_test
mount -o rdma,port=20049 192.168.2.1:/mnt/nfs_test_testrun /mnt/nfs_test_testrun/"


4. Run the actual test from the client node:
"
#!/bin/bash

fsize=268435456
jfile=/dev/shm/run_nfs_test2.junk
tfile=/dev/shm/run_nfs_test2.tmp
nfsfile=/mnt/nfs_test_testrun/run_nfs_test2.junk
rm -r -f /mnt/nfs_test_testrun/
rm -f ${tfile}
rm -f ${jfile}

dd if=/dev/urandom iflag=fullblock of=${jfile} bs=1024 count=$((fsize/1024));

for i in {1..100}; do
          cp ${jfile} ${nfsfile}; # Bottleneck 1

          cp ${nfsfile} ${tfile}; # Bottleneck 2

         cmp ${jfile} ${tfile};

          rm -f ${tfile};
        echo "DONE with iter ${i}"
done;

rm -f ${jfile};
rm -f ${tfile};
echo "Done";
"


On v6.8 I was seeing this test taking about 1m50s to complete, for all 10 iterations. On v6.9-rc1 it takes 3-7 minutes, and I also see these kernel traces printed continuously in dmesg during this regression:

[23720.243905] INFO: task kworker/61:1:556 blocked for more than 122 seconds.
[23720.251709]       Not tainted 6.9.0-rc1 #1
[23720.256387] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[23720.265268] task:kworker/61:1    state:D stack:0     pid:556   tgid:556   ppid:2      flags:0x00004000
[23720.275822] Workqueue: events __svc_rdma_free [rpcrdma]
[23720.281803] Call Trace:
[23720.284630]  <TASK>
[23720.287067]  __schedule+0x210/0x660
[23720.291063]  schedule+0x2c/0xb0
[23720.294668]  schedule_timeout+0x146/0x160
[23720.299249]  __wait_for_common+0x92/0x1d0
[23720.303828]  ? __pfx_schedule_timeout+0x10/0x10
[23720.308987]  __ib_drain_sq+0xfa/0x170 [ib_core]
[23720.314190]  ? __pfx_ib_drain_qp_done+0x10/0x10 [ib_core]
[23720.320343]  ib_drain_qp+0x71/0x80 [ib_core]
[23720.325232]  __svc_rdma_free+0x28/0x100 [rpcrdma]
[23720.330604]  process_one_work+0x196/0x3d0
[23720.335185]  worker_thread+0x2fc/0x410
[23720.339470]  ? __pfx_worker_thread+0x10/0x10
[23720.344336]  kthread+0xdf/0x110
[23720.347941]  ? __pfx_kthread+0x10/0x10
[23720.352225]  ret_from_fork+0x30/0x50
[23720.356317]  ? __pfx_kthread+0x10/0x10
[23720.360602]  ret_from_fork_asm+0x1a/0x30
[23720.365083]  </TASK>