We are having problems with the reproducer on RHEL8 due to bz1675071. By the way, <rmetrich>, this can affect your customer as well in case the problem occurs with any SCSI driver (I don't know that yet). Basically the problem is that on systems with such a large number of SCSI devices the kernel consumes about a gigabyte of memory.

We could try with real SCSI, for example tons of iSCSI luns.

That has been my idea as well, but I need to work on other stuff now....

Not promising anything, but will try to setup something.

That would be great. My idea was to setup one machine with lots of scsi_debug LUNs (with enough RAM, using the workaround or using RHEL 7 to not suffer from the bug), exporting them via iSCSI target, and let another machine connect to it using a iSCSI initiator.

I cannot reproduce the old good slowness with iscsi, but still we can see a difference.

Server setup:
------------

# yum -y install targetcli

# cat iscsi.sh

#!/bin/bash

(
for vol in $(seq 1 100); do
	echo "backstores/fileio create file_or_dev=/root/iscsi/vol${vol}.img size=1M name=vol${vol}"
done

for tgt in $(seq 1 4); do
	echo "iscsi/ create iqn.2019-02.com.bz1673978:tgt${tgt}"
	echo "iscsi/iqn.2019-02.com.bz1673978:tgt${tgt}/tpg1/acls create iqn.2019-02.com.bz1673978:client"
	for vol in $(seq 1 100); do
		echo "iscsi/iqn.2019-02.com.bz1673978:tgt${tgt}/tpg1/luns create /backstores/fileio/vol${vol}"
	done
done
) | targetcli

# chmod +x iscsi.sh; ./iscsi.sh

--> creates 100 luns and 4 targets

# systemctl start target


Client setup:
------------

# yum -y install iscsi-initiator-utils device-mapper-multipath

# echo "InitiatorName=iqn.2019-02.com.bz1673978:client" > /etc/iscsi/initiatorname.iscsi
# iscsiadm -m discovery -t st -p <server_ipaddr>
# for i in $(seq 1 4); do iscsiadm -m node -T iqn.2019-02.com.bz1673978:tgt$i -p <server_ipaddr> -l; done



WITHOUT THE FIX:

# time rear mkrescue

real	4m14.741s
user	1m56.594s
sys	1m57.890s


WITH THE FIX:

# time rear mkrescue

real	1m38.588s
user	0m50.723s
sys	0m31.869s

I suspect you are not actually using multipath. How many /dev/dm-* devices do you have? Use 

multipath -l 

to verify.

In my test case, I am enabling multipath with

mpathconf --enable --with_multipathd y

I don't think that installing device-mapper-multipath was enough.

However, the question regarding iscsi was mainly about the memory consumption of such a large number of iSCSI devices at the initiator - whether you are also seeing this memory consumption on the order of GBs like we do with the scsi_debug driver.

Multipath is activated on my system (I used a VM already with multipath disks).
And I don't see any memory impact.

> creates 100 luns and 4 targets

The number of paths per LUN is equal to the number of targets, right? i.e. 4 paths per LUN.

I suspect that 100 luns is not enough to see the problem, the old code IIUC has a quadratic complexity in the number of devices. So my test which uses 1200 devices in total (400 LUNs x 3 paths per LUN) should be 9x slower than your case, if it uses 400 devices.

> And I don't see any memory impact.

Is that on RHEL8?

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://access.redhat.com/errata/RHBA-2019:3413