Lee and Manish and others, Don't review this patchset. I'm hitting some issues with the code before my patchset. It will be easier to test/review if I fix them first. Lee, I'll send patches for the ep_disconnect/iscsi_conn_stop issue. Manish, I found some bugs in qedi that we might be hitting: - it shouldn't use iscsi_block_session in the tmf path - libiscsi and qedi can get out of sync in the tmf paths and cleanup the wrong cmds. On 4/3/21 6:22 PM, Mike Christie wrote: > The following patches apply over Linus's tree or Martin's staging branch. > They fix up the locking and refcount handling in the iscsi code so for > software iscsi we longer need a lock when going from queuecommand to the > xmit thread and no longer need a common iscsi level lock between the xmit > thread and completion paths. > > For simple throughput workloads like > > fio --filename=/dev/sdb --direct=1 --rw=randwrite --bs=256k \ > --ioengine=libaio --iodepth=128 --numjobs=1 --time_based \ > --group_reporting --name=throughput --runtime=120 > > I'm able to get throughput from 24 Gb/s to 28 where I then hit a > bottleneck on the target side. > > IOPs might increase by around 10% in some cases with: > > fio --filename=/dev/sdb --direct=1 --rw=randwrite --bs=4k \ > --ioengine=libaio --iodepth=128 --numjobs=1 --time_based \ > --group_reporting --name=throughput --runtime=120 > > I'm still debugging some target side issues. > > A bigger advantage I'm seeing with the patches is that for setups where > you have software iscsi sharing CPUs with other subsystems like vhost > IOPs can increase by up to 20%. > > Notes: > - I've tested iscsi_tcp, ib_iser, be2iscsi and qedi. I don't have cxgbi > or bnx2i hardware, but cxbgi changes were API only. > > - Lee, the first 2 patches are new bug fixes. The first half are then > similar to what you saw before. I was not sure how far through them you > were. The second half was the part that removed the back lock and frwd > lock from iscsi_queuecommand are new. > > > >
