[SRU] Fix the session workqueue thread priority setting issue for newer Linux kernels (>=6.x)

[Mustafa Kemal Gilor]

[ Impact ]

The Linux SCSI driver uses `alloc_workqueue()` to create a kernel workqueue for session transmit work. This call would cause the kernel < 6.x to create a dedicated worker thread for the workqueue. The userspace library open-iscsi version < 2.1.10 then adjusts the workqueue thread's nice value for performance reasons when a new iSCSI session is initiated. The algorithm for that is roughly as follows (https://github.com/open-iscsi/open-iscsi/blob/2.1.9/usr/initiator.c#L1390) :

- Check if the driver in use has a write work queue. If not, abort.
- Open the /proc dir, and iterate over all dir entries:
- Run "stat" over /proc/<n>/stat
- Read the contents of "stat" file, which looks like the following:
898582 (kworker/u512:1-iscsi_q_0) I 2 0 0 0 -1 69238880 0 0 0 0 0 8 0 0 20 0 1 0 52431895 0 0 18446744073709551615 0 0 0 0 0 0 0 2147483647 0 1 0 0 17 28 0 0 0 0 0 0 0 0 0 0 0 0 0
- Try to locate "(" with strchr, starting from the beginning. Skip the process if not found.
- Try to locate ")" with strchr, starting from the position of "(". Skip the process if not found.
- Check whether the string between "(" and ")" contains the following pattern "iscsi_q_%d"
- Check the number %d matches with the host ID of the session.
- If %d matches the session id, grab the PID of the current proc entry and call `setpriority()`

So the algorithm assumes the following about the kernel workqueue thread:

- It would be present in the /proc list
- Its name would match the iscsi_q_%d pattern

Due to the changes in how Linux workqueue threads work in v6.x, the priority setting approach won't work for the following reasons:

- The `alloc_workqueue()` no longer creates a dedicated thread for the workqueue. The workqueue thread is shared between different workqueues.
- The workqueue thread is dynamically renamed to the name of the workqueue that's actively running
- The workqueue thread disappears from the /proc list when it's inactive

The algorithm as-is does the following right now:

- If the kernel workqueue thread *by luck* happens to be running the iscsi task, the name matches, and the priority is set. But that's not what the code wants to do since it also increases the priority for all the other tasks that are scheduled to the workqueue thread as well.

- If not, the open-iscsi prints the following log message, and proceeds to operate as normal:

```
iscsistart: Could not set session1 priority. READ/WRITE throughout and latency could be affected.
```

The upstream has fixed this issue with the patch (https://github.com/open-iscsi/open-iscsi/pull/445). The patch sets the default nice value for `node.session.xmit_thread_priority` to `0`, and then skips the priority adjustment algorithm altogether when the priority is set to zero.

This SRU proposes to backport this patch to the Ubuntu releases that use Linux kernel 6.8 and above by default, and have an open-iscsi version of less than (2.1.10).

[ Test Plan ]

# Launch a test VM:
$> lxc launch ubuntu:noble --vm iscsi-test-noble

# Obtain a shell from the VM:
$> lxc shell iscsi-test-noble

# Install 'tgt' and 'open-iscsi':
$> sudo apt -y update && sudo apt -y install tgt open-iscsi

# Configure 'tgt':

## Step 1: Configure a LUN

Add the following to '/etc/tgt/conf.d/iscsi.conf':

<target iqn.2020-07.example.com:lun1>
     backing-store /dev/sda
     initiator-address 127.0.0.1
     incominguser iscsi-user password
     outgoinguser iscsi-target secretpass
</target>

(change /dev/sda to an existing device's name if it's not present)

## Step 2: Restart 'tgt' to make changes effective:
$> systemctl restart tgt

## Step 3: Check if 'tgt' has started serving the LUN:
$> tgtadm --mode target --op show

(output should be similar to below)
Target 1: iqn.2020-07.example.com:lun1
    System information:
        Driver: iscsi
        State: ready
    I_T nexus information:
    LUN information:
        LUN: 0
            Type: controller
            SCSI ID: IET 00010000
            SCSI SN: beaf10
            Size: 0 MB, Block size: 1
            Online: Yes
            Removable media: No
            Prevent removal: No
            Readonly: No
            SWP: No
            Thin-provisioning: No
            Backing store type: null
            Backing store path: None
            Backing store flags:
    Account information:
        iscsi-user
        iscsi-target (outgoing)
    ACL information:
        127.0.0.1

# Configure 'open-iscsi':

## Step 1: Check whether the LUN being served by 'tgt' is discoverable:

$> iscsiadm -m discovery -t st -p 127.0.0.1

(should output the text below)
127.0.0.1:3260,1 iqn.2020-07.example.com:lun1

## Step 2: Configure open-iscsi to consume the target LUN:

Add the following line to '/etc/iscsi/initiatorname.iscsi':

```
InitiatorName=iqn.2020-07.example.com:lun1
```

## Step 3: Modify the following file '/etc/iscsi/nodes/iqn.2020-07.example.com:lun1/127.0.0.1,3260,1/default':
# (the file must already exist, it should've been automatically created after the discovery)

Append the following lines to the end of the file, and save:

```
node.session.auth.authmethod = CHAP
node.session.auth.username = iscsi-user
node.session.auth.password = password
node.session.auth.username_in = iscsi-target
node.session.auth.password_in = secretpass
node.startup = automatic
```

## Step 4: Restart open-iscsi to make changes effective:

$> systemctl restart open-iscsi.service iscsid

## Step 5: Check the outcome
(the service status should indicate that login to 'iqn.2020-07.example.com:lun1' has been successful)

$> systemctl status open-iscsi.service

● open-iscsi.service - Login to default iSCSI targets
     Loaded: loaded (/usr/lib/systemd/system/open-iscsi.service; enabled; preset: enabled)
     Active: active (exited) since Mon 2024-07-22 13:36:15 UTC; 4s ago
       Docs: man:iscsiadm(8)
             man:iscsid(8)
    Process: 3049 ExecStart=/usr/sbin/iscsiadm -m node --loginall=automatic (code=exited, status=0/SUCCESS)
    Process: 3065 ExecStart=/usr/lib/open-iscsi/activate-storage.sh (code=exited, status=0/SUCCESS)
   Main PID: 3065 (code=exited, status=0/SUCCESS)
        CPU: 4ms

Jul 22 13:36:15 welcomed-bluebird systemd[1]: Starting open-iscsi.service - Login to default iSCSI targets...
Jul 22 13:36:15 welcomed-bluebird iscsiadm[3049]: Logging in to [iface: default, target: iqn.2020-07.example.com:lun1, portal: 127.0.0.1,3260]
Jul 22 13:36:15 welcomed-bluebird iscsiadm[3049]: Login to [iface: default, target: iqn.2020-07.example.com:lun1, portal: 127.0.0.1,3260] successful.
Jul 22 13:36:15 welcomed-bluebird systemd[1]: Finished open-iscsi.service - Login to default iSCSI targets

# (the command should list an active connection to the 'iqn.2020-07.example.com:lun1')

$> iscsiadm -m session -o show

tcp: [1] 127.0.0.1:3260,1 iqn.2020-07.example.com:lun1 (non-flash)

# Observe iscsid is complaining about priority:

$> cat /var/log/syslog | grep "Could not set"

2024-07-22T13:36:16.874243+00:00 welcomed-bluebird iscsid: Could not set session1 priority. READ/WRITE throughout and latency could be affected.
2024-07-22T13:38:31.002732+00:00 welcomed-bluebird iscsid: Could not set session1 priority. READ/WRITE throughout and latency could be affected.

## VERIFICATION OF THE FIX ##

# Add the PPA that includes the fix, and update the open-iscsi package:

$> add-apt-repository ppa:mustafakemalgilor/lp-2073846
$> apt update
$> apt -y install open-iscsi

# Edit the '/etc/iscsi/nodes/iqn.2020-07.example.com:lun1/127.0.0.1,3260,1/default' file to set the priority to 0:
(this is needed because we've created it before the update so the priority is explicitly set to "-20". The new ones should have the "node.session.xmit_thread_priority" value of "0".)

$> sed -E -i 's/^(node.session.xmit_thread_priority[[:blank:]]*=[[:blank:]]*).*/\10/' /etc/iscsi/nodes/iqn.2020-07.example.com\:lun1/127.0.0.1\,3260\,1/default

# Verify that "node.session.xmit_thread_priority" is indeed set to "0":

$> grep "node.session.xmit_thread_priority" /etc/iscsi/nodes/iqn.2020-07.example.com\:lun1/127.0.0.1\,3260\,1/default

(should output the following):
node.session.xmit_thread_priority = 0

# Truncate the syslog
$> truncate -s 0 /var/log/syslog

# Restart the service

$> systemctl restart open-iscsi.service

# Observe that the priority warning has disappeared:

$> grep "iscsi\(adm\|d\).*:" /var/log/syslog

# Remove the node altogether:

$> rm -rf /etc/iscsi/nodes/iqn.2020-07.example.com

# Re-discover the node:

$> iscsiadm -m discovery -t st -p 127.0.0.1

# Confirm that the priority is set to "0":

$> grep "node.session.xmit_thread_priority" /etc/iscsi/nodes/iqn.2020-07.example.com\:lun1/127.0.0.1\,3260\,1/default

(should output the following):
node.session.xmit_thread_priority = 0

# Confirm that the /etc/iscsi/iscsid.conf priority is set to "0":

$> grep "node.session.xmit_thread_priority" /etc/iscsi/iscsid.conf

(should output the following):
node.session.xmit_thread_priority = 0

[ Where problems could occur ]

The change prevents a priority change that shouldn't happen in the first place. That might affect some workloads unknowingly depending on it. On the other hand, the nice setting happens intermittently (i.e. by luck) so the behavior right now can't be depended on anyway. The patch only touches the priority setting code so I wouldn't expect any serious breakage.

[ Other Info ]

The other releases that is running a 6.x kernel installed with other means (e.g. hw-enablement, availability) may set the `node.session.xmit_thread_priority` from `-20` to `0` in `/etc/iscsid.conf` as a workaround:

node.session.xmit_thread_priority = 0

which is the default priority for the workqueue threads.

Also, the existing configuration for the nodes under `/etc/iscsi/nodes` path will not be altered, so they must be manually set to "0".

[Mustafa Kemal Gilor (mustafakemalgilor)]

Removed 'oracular' from targets as it seems to be updated to 2.1.10 already.

[Heitor Alves de Siqueira (halves)]

Thanks for the detailed bug report, Mustafa! I've left a review for you MR, feel free to take a look when you have the time.

For future reference, please refrain from removing series from an LP bug. It's preferable to mark a series as "Fix Released" or "Won't Fix" appropriately, so that the bug doesn't disappear from any searches or queries in Launchpad. I've re-added Oracular as Fix Released, according to your comment on version 2.1.10.