Skip to content

Latest commit

 

History

History
261 lines (178 loc) · 11.5 KB

README.md

File metadata and controls

261 lines (178 loc) · 11.5 KB
Raw

tcmu-runner

A daemon that handles the userspace side of the LIO TCM-User backstore.

Background

LIO is the SCSI target in the Linux kernel. It is entirely kernel code, and allows exported SCSI logical units (LUNs) to be backed by regular files or block devices. But, if we want to get fancier with the capabilities of the device we're emulating, the kernel is not necessarily the right place. While there are userspace libraries for compression, encryption, and clustered storage solutions like Ceph or Gluster, these are not accessible from the kernel.

The TCMU userspace-passthrough backstore allows a userspace process to handle requests to a LUN. But since the kernel-user interface that TCMU provides must be fast and flexible, it is complex enough that we'd like to avoid each userspace handler having to write boilerplate code.

tcmu-runner handles the messy details of the TCMU interface -- UIO, netlink, pthreads, and DBus -- and exports a more friendly C plugin module API. Modules using this API are called "TCMU handlers". Handler authors can write code just to handle the SCSI commands as desired, and can also link with whatever userspace libraries they like.

Usage example

One goal of TCMU is that configuring a userspace-backed LUN should be as easy as configuring a kernel-backed LUN. We're not quite there yet. This will require cooperation with the LIO configuration tool, targetcli. targetcli should list user-backed backstores along with the built-in kernel backstores, and ensure tcmu-runner is started if a user-backed backstore is created.

Info for potential contributors and handler authors

License

tcmu-runner is LGPLv2.1 or Apache License 2.0.

Releases

Tarballs are available from https://github.com/open-iscsi/tcmu-runner/releases.

Development

We encourage pull requests and issues tracking via Github, and the target-devel mailing list (list info) may be used for discussion.

Getting started

Building tcmu-runner
  1. Clone this repo.
  2. Type ./extra/install_dep.sh to install development packages for dependencies, or you can do it manually:
    • Note: Install cmake and other packages which usually ending with "-devel" or "-dev": libnl3, libglib2 (or glib2-devel on Fedora), libpthread, libdl, libkmod, libgfapi (Gluster), librbd1 (Ceph), zlib.
  3. Type cd ./extra && ./make_runnerrpms.sh [--without (rbd|glfs|qcow|zbc|fbo)] to build the RPM packages automatically.
  4. Type cmake . [-Dwith-<rbd|glfs|qcow|zbc|fbo>=false]
    • Note: If using systemd, -DSUPPORT_SYSTEMD=ON -DCMAKE_INSTALL_PREFIX=/usr should be passed to cmake, so files are installed to the correct location.
  5. Type make
  6. Type make install
  7. Type xargs rm < install_manifest.txt to uninstall from source
Running tcmu-runner
  1. Copy tcmu-runner.conf to /etc/dbus-1/system.d/. This allows tcmu-runner to be on the system bus, which is privileged.
  2. If using systemd, copy org.kernel.TCMUService1.service to /usr/share/dbus-1/system-services/ and tcmu-runner.service to /lib/systemd/system.
  3. Or, run it from the command line as root. It should print the number of handlers and devices found.
Creating a LIO user-backed storage object with backstore specific tools
  • Ceph:

If setting up tcmu-runner in a HA configuration, the ceph-iscsi-cli (https://github.com/ceph/ceph-iscsi-cli) tool is the preferred management tool.

Bug reports should be made to the tcmu-runner github: https://github.com/open-iscsi/tcmu-runner/issues, but can be made to ceph-users@ceph.com mailing list.

  • Gluster:

Gluster management must be done with the gluster-block tools (https://github.com/gluster/gluster-block).

Bug reports must be made to the gluster-block github: https://github.com/gluster/gluster-block/issues

Creating a LIO user-backed storage object with targetcli-fb or configfs

Support for the user/tcmu backstore is supported in targetcli-fb/rtslib-fb:

https://github.com/open-iscsi/targetcli-fb https://github.com/open-iscsi/rtslib-fb

  1. Start targetcli

# targetcli

  1. Go to the user/tcmu backstore dir.

/> cd /backstores/

  1. By default, tcmu-runner installs the file, zbc, glfs, qcow and rbd tcmu-runner handlers:
/backstores> ls
o- backstores .......................................................... [...]
  o- user:glfs .......................................... [Storage Objects: 0]
  o- user:qcow .......................................... [Storage Objects: 0]
  o- user:rbd ........................................... [Storage Objects: 0]
  o- user:file .......................................... [Storage Objects: 0]
  o- user:zbc ........................................... [Storage Objects: 0]
  1. 'cd' to the handler you want to setup:

/backstores> cd user:rbd

/backstores/user:rbd> create cfgstring=pool/rbd1;osd_op_timeout=30 name=rbd0 size=1G Created user-backed storage object rbd0 size 1073741824.

The cfgstrng format is:

cfgstring=;tcmu-runner daemon arguments;handler arguments

The following tcmu-runner daemon arguments are optional:

  • tcmur_cmd_time_out: Number of seconds before logging the command as timed out, and executing a handler specific timeout handler if supported.

If passed in they must start before the handler specific arguments and each argument must start and end with a semicolon ";".

Example using the rbd cfgstring with the optional tcmur_cmd_timeout arg:

create ;tcmur_cmd_timeout=20;pool/rbd1;osd_op_timeout=30 name=rbd0 size=1G

The handler specific arguments and their formats are:

  • rbd: /pool_name/image_name[;osd_op_timeout=N;conf=N;id=N] (osd_op_timeout is optional and N is in seconds) (conf is optional and N is the path to the conf file) (id is optional and N is the id to connect to the cluster as)
  • qcow: /path_to_file
  • glfs: /volume@hostname/filename
  • file: /path_to_file
  • zbc: /[opt1[/opt2][...]@]path_to_file

For the zbc handler, the available options are shown in the table below.

Option Description Default value
model-type Device model type, HA for host aware or HM for host managed HM
lba-size (B) LBA size in bytes (512 or 4096) 512
zsize-size (MiB) Zone size in MiB 256 MiB
conv-num Number of conventional zones at LBA 0 (can be 0) Number of zones corresponding to 1% of the device capacity
open-num Optimal (for host aware) or maximum (for host managed) number of open zones 128

Example:

cfgstring=model-HM/zsize-128/conv-100@/var/local/zbc.raw

will create a host-managed disk with 128 MiB zones and 100 conventional zones, stored in the file /var/local/zbc.raw.

  1. The created backstore device can then be mapped to a LUN like traditional backstores.
Logger setting and system configuration
  • Logger setting:

There are 6 logging levels supported:

  1. CRIT
  2. ERROR
  3. WARNING
  4. INFO
  5. DEBUG
  6. DEBUG SCSI CMD

And the default logging level is 3, if you want to change the default level, uncomment the following line in /etc/tcmu/tcmu.conf and set your level number:

# log_level = 3

The precedence of the config settings is as mentioned belows:

  1. Options set through config file /etc/tcmu/tcmu.conf [Top Prio]
    eg: uncommenting and adjusting key:value at /etc/tcmu/tcmu.conf
  2. Arguments passed at daemon
    eg: -l/--tcmu-log-dir, -d/--debug
  3. Environment variable.
    eg: export TCMU_LOGDIR="/var/log/mylogdir/"
  4. Code level defaults. [Least Prio]
    eg: TCMU_LOG_DIR_DEFAULT = '/var/log/' & TCMU_CONF_LOG_INFO = INFO
  • System configuration:

The default configuration file is installed into /etc/tcmu/tcmu.conf.

Tcmu-runner's configuration systems supports dynamic reloading without restarting the daemon. To change values open /etc/tcmu/tcmu.conf, update the value, and then close the file.

If your version of targetcli/rtslib does not support tcmu, setup can be done manually through configfs:

  1. Ensure target_core_user kernel module is loaded.
  2. Create the HBA (user_1) and the storage object (test): mkdir -p /sys/kernel/config/target/core/user_1/test
  3. Go to that directory: cd /sys/kernel/config/target/core/user_1/test
  4. Set configuration values
  5. Set size (in bytes): echo -n dev_size=16777216 > control
  6. Set configstring. See tcmu-design.txt, but note that the TCMU backstore driver already knows and will prepend the "tcm-user/hba_num/device_name" part. Therefore, if we wanted our new device to be handled by the "baz" handler, we would give subtype and path by running: echo -n dev_config=baz/addl_info_for_baz_handler > control
  7. Enable the storage object: echo -n 1 > enable
  8. Verify everything worked. There should be an entry in /sys/class/uio.

To delete:

  1. rmdir /sys/kernel/config/target/core/user_1/test
  2. rmdir /sys/kernel/config/target/core/user_1

Writing a TCMU handler

libtcmu and tcmu-runner

There are two different ways to write a TCMU handler. The primary difference is who is responsible for the event loop.

tcmu-runner plugin handler

There are two different ways to write a tcmu-runner plugin handler:

  1. one can register .handle_cmd to take the full control of command handling
  2. or else one can register .{read, write, flush, ...} to handle specific operations and .handle_cmd to override the generic handling if needed.

With the option 1, tcmu-runner is in charge of the event loop for your plugin, and your handler's handle_cmd function is called repeatedly to respond to each incoming SCSI command. While your handler sees all SCSI commands, there are helper functions provided that save each handler from writing boilerplate code for mandatory SCSI commands, if desired.

The file_optical handler is an examples of this type.

With the option 2, tcmu-runner will be partially or fully in charge of the event loop and SCSI command handling for your plugin, and your handler's registered functions are called repeatedly to handle storage requests as required by the upper SCSI layer, which will handle most of the SCSI commands for you.

  • Note: If the .handle_cmd is also implemented by the handler, tcmu-runner will try to pass through the commands to the handler first. If and only when the handler won't support the commands it should return TCMU_STS_NOT_HANDLED, then the tcmu-runner will handle them in the generic handler.

The file_example handler is an example of this type.

tcmulib

If you want to add handling of TCMU devices to an existing daemon or other program that already is processing its own event loop, the best option is to use tcmulib directly. This requires your code to keep track of tcmulib's file descriptors. While tcmulib's 'master' file descriptor must be handled with tcmulib_master_fd_ready() single-threadedly, per-device fds can be handled on the main thread (with tcmulib_get_next_command and tcmulib_command_complete) or separate threads if desired. SCSI command-processing helper functions are still available for use.

tcmu-runner itself uses tcmulib in this manner and may be used as an example of multi-threaded tcmulib use. The consumer.c example demonstrates single-threaded tcmulib processing.