Blktap2 Userspace Tools + Library
================================

Dutch Meyer
4th June 2009

Andrew Warfield and Julian Chesterfield
16th June 2006


The blktap2 userspace toolkit provides a user-level disk I/O
interface. The blktap2 mechanism involves a kernel driver that acts
similarly to the existing Xen/Linux blkback driver, and a set of
associated user-level libraries.  Using these tools, blktap2 allows
virtual block devices presented to VMs to be implemented in userspace
and to be backed by raw partitions, files, network, etc.

The key benefit of blktap2 is that it makes it easy and fast to write
arbitrary block backends, and that these user-level backends actually
perform very well.  Specifically:

- Metadata disk formats such as Copy-on-Write, encrypted disks, sparse
  formats and other compression features can be easily implemented.

- Accessing file-based images from userspace avoids problems related
  to flushing dirty pages which are present in the Linux loopback
  driver.  (Specifically, doing a large number of writes to an
  NFS-backed image don't result in the OOM killer going berserk.)

- Per-disk handler processes enable easier userspace policing of block
  resources, and process-granularity QoS techniques (disk scheduling
  and related tools) may be trivially applied to block devices.

- It's very easy to take advantage of userspace facilities such as
  networking libraries, compression utilities, peer-to-peer
  file-sharing systems and so on to build more complex block backends.

- Crashes are contained -- incremental development/debugging is very
  fast.

How it works (in one paragraph):

Working in conjunction with the kernel blktap2 driver, all disk I/O
requests from VMs are passed to the userspace deamon (using a shared
memory interface) through a character device. Each active disk is
mapped to an individual device node, allowing per-disk processes to
implement individual block devices where desired.  The userspace
drivers are implemented using asynchronous (Linux libaio),
O_DIRECT-based calls to preserve the unbuffered, batched and
asynchronous request dispatch achieved with the existing blkback
code.  We provide a simple, asynchronous virtual disk interface that
makes it quite easy to add new disk implementations.

As of June 2009 the current supported disk formats are:

 - Raw Images (both on partitions and in image files)
 - Fast sharable RAM disk between VMs (requires some form of 
   cluster-based filesystem support e.g. OCFS2 in the guest kernel)
 - VHD, including snapshots and sparse images
 - Qcow, including snapshots and sparse images


Build and Installation Instructions
===================================

Make to configure the blktap2 backend driver in your dom0 kernel.  It
will inter-operate with the existing backend and frontend drivers.  It
will also cohabitate with the original blktap driver.  However, some
formats (currently aio and qcow) will default to their blktap2
versions when specified in a vm configuration file.

To build the tools separately, "make && make install" in
tools/blktap2.


Using the Tools
===============

Preparing an image for boot:

The userspace disk agent is configured to start automatically via xend

Customize the VM config file to use the 'tap:tapdisk' handler,
followed by the driver type. e.g. for a raw image such as a file or
partition:

disk = ['tap:tapdisk:aio:<FILENAME>,sda1,w']

Alternatively, the vhd-util tool (installed with make install, or in
/blktap2/vhd) can be used to build sparse copy-on-write vhd images.

For example, to build a sparse image -
  vhd-util create -n MyVHDFile -s 1024

This creates a sparse 1GB file named "MyVHDFile" that can be mounted
and populated with data.

One can also base the image on a raw file -
  vhd-util snapshot -n MyVHDFile -p SomeRawFile -m

This creates a sparse VHD file named "MyVHDFile" using "SomeRawFile"
as a parent image.  Copy-on-write semantics ensure that writes will be
stored in "MyVHDFile" while reads will be directed to the most
recently written version of the data, either in "MyVHDFile" or
"SomeRawFile" as is appropriate.  Other options exist as well, consult
the vhd-util application for the complete set of VHD tools.

VHD files can be mounted automatically in a guest similarly to the
above AIO example simply by specifying the vhd driver.

disk = ['tap:tapdisk:vhd:<VHD FILENAME>,sda1,w']


Snapshots:

Pausing a guest will also plug the corresponding IO queue for blktap2
devices and stop blktap2 drivers.  This can be used to implement a
safe live snapshot of qcow and vhd disks.  An example script "xmsnap"
is shown in the tools/blktap2/drivers directory.  This script will
perform a live snapshot of a qcow disk.  VHD files can use the
"vhd-util snapshot" tool discussed above.  If this snapshot command is
applied to a raw file mounted with tap:tapdisk:AIO, include the -m
flag and the driver will be reloaded as VHD.  If applied to an already
mounted VHD file, omit the -m flag.


Mounting images in Dom0 using the blktap2 driver
===============================================
Tap (and blkback) disks are also mountable in Dom0 without requiring an
active VM to attach. 

The syntax is -
  tapdisk2 -n <type>:<full path to file>

For example -
  tapdisk2  -n aio:/home/images/rawFile.img

When successful the location of the new device will be provided by
tapdisk2 to stdout and tapdisk2 will terminate.  From that point
forward control of the device is provided through sysfs in the
directory-

  /sys/class/blktap2/blktap#/

Where # is a blktap2 device number present in the path that tapdisk2
printed before terminating.  The sysfs interface is largely intuitive,
for example, to remove tap device 0 one would-
  
  echo 1 > /sys/class/blktap2/blktap0/remove

Similarly, a pause control is available, which is can be used to plug
the request queue of a live running guest.

Previous versions of blktap mounted devices in dom0 by using blkfront
in dom0 and the xm block-attach command.  This approach is still
available, though slightly more cumbersome.


Tapdisk Development
===============================================

People regularly ask how to develop their own tapdisk drivers, and
while it has not yet been well documented, the process is relatively
easy.  Here I will provide a brief overview.  The best reference, of
course, comes from the existing drivers.  Specifically,
blktap2/drivers/block-ram.c and blktap2/drivers/block-aio.c provide
the clearest examples of simple drivers.
 

Setup:

First you need to register your new driver with blktap. This is done
in disktypes.h.  There are five things that you must do.  To
demonstrate, I will create a disk called "mynewdisk", you can name
yours freely.

1) Forward declare an instance of struct tap_disk.

e.g. -  
  extern struct tap_disk tapdisk_mynewdisk;

2) Claim one of the unused disk type numbers, take care to observe the
MAX_DISK_TYPES macro, increasing the number if necessary.

e.g. -
  #define DISK_TYPE_MYNEWDISK         10

3) Create an instance of disk_info_t.  The bulk of this file contains examples of these.

e.g. -
  static disk_info_t mynewdisk_disk = {
          DISK_TYPE_MYNEWDISK,
          "My New Disk (mynewdisk)",
          "mynewdisk",
          0,
  #ifdef TAPDISK
          &tapdisk_mynewdisk,
  #endif
  };

A few words about what these mean.  The first field must be the disk
type number you claimed in step (2).  The second field is a string
describing your disk, and may contain any relevant info.  The third
field is the name of your disk as will be used by the tapdisk2 utility
and xend (for example tapdisk2 -n mynewdisk:/path/to/disk.image, or in
your xm create config file).  The forth is binary and determines
whether you will have one instance of your driver, or many.  Here, a 1
means that your driver is a singleton and will coordinate access to
any number of tap devices.  0 is more common, meaning that you will
have one driver for each device that is created.  The final field
should contain a reference to the struct tap_disk you created in step
(1).

4) Add a reference to your disk info structure (from step (3)) to the
dtypes array.  Take care here - you need to place it in the position
corresponding to the device type number you claimed in step (2).  So
we would place &mynewdisk_disk in dtypes[10].  Look at the other
devices in this array and pad with "&null_disk," as necessary.

5) Modify the xend python scripts.  You need to add your disk name to
the list of disks that xend recognizes.

edit:
  tools/python/xen/xend/server/BlktapController.py

And add your disk to the "blktap_disk_types" array near the top of
your file.  Use the same name you specified in the third field of step
(3).  The order of this list is not important.


Now your driver is ready to be written.  Create a block-mynewdisk.c in
tools/blktap2/drivers and add it to the Makefile.


Development:

Copying block-aio.c and block-ram.c would be a good place to start.
Read those files as you go through this, I will be assisting by
commenting on a few useful functions and structures.

struct tap_disk:

Remember the forward declaration in step (1) of the setup phase above?
Now is the time to make that structure a reality.  This structure
contains a list of function pointers for all the routines that will be
asked of your driver.  Currently the required functions are open,
close, read, write, get_parent_id, validate_parent, and debug.

e.g. -
  struct tap_disk tapdisk_mynewdisk = {
          .disk_type          = "tapdisk_mynewdisk",
          .flags              = 0,
          .private_data_size  = sizeof(struct tdmynewdisk_state),
          .td_open            = tdmynewdisk_open,
                 ....

The private_data_size field is used to provide a structure to store
the state of your device.  It is very likely that you will want
something here, but you are free to design whatever structure you
want.  Blktap will allocate this space for you, you just need to tell
it how much space you want.


tdmynewdisk_open:

This is the open routine.  The first argument is a structure
representing your driver.  Two fields in this array are
interesting. 

driver->data will contain a block of memory of the size your requested
in in the .private_data_size field of your struct tap_disk (above).

driver->info contains a structure that details information about your
disk.  You need to fill this out.  By convention this is done with a
_get_image_info() function.  Assign a size (the total number of
sectors), sector_size (the size of each sector in bytes, and set
driver->info->info to 0.

The second parameter contains the name that was specified in the
creation of your device, either through xend, or on the command line
with tapdisk2.  Usually this specifies a file that you will open in
this routine.  The final parameter, flags, contains one of a number of
flags specified in tapdisk.h that may change the way you treat the
disk.


_queue_read/write:

These are your read and write operations.  What you do here will
depend on your disk, but you should do exactly one of- 

1) call td_complete_request with either error or success code.

2) Call td_forward_request, which will forward the request to the next
driver in the stack.

3) Queue the request for asynchronous processing with
td_prep_read/write.  In doing so, you will also register a callback
for request completion.  When the request completes you must do one of
options (1) or (2) above.  Finally, call td_queue_tiocb to submit the
request to a wait queue.

The above functions are defined in tapdisk-interface.c.  If you don't
use them as specified you will run into problems as your driver will
fail to inform blktap of the state of requests that have been
submitted.  Blktap keeps track of all requests and does not like losing track.


_close, _get_parent_id, _validate_parent:

These last few tend to be very routine.  _close is called when the
device is closed, and also when it is paused (in this case, open will
also be called later).  The other functions are used in stacking
drivers.  Most often drivers will return TD_NO_PARENT and -EINVAL,
respectively.