Revision cbb533f4

b/NEWS
7 7

  
8 8
*(unreleased)*
9 9

  
10
New features
11
~~~~~~~~~~~~
12
- KVM hypervisors can now access RBD storage directly without having to
13
  go through a block device.
14

  
10 15
Incompatible/important changes
11 16
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
12 17

  
b/doc/install.rst
367 367

  
368 368
      $ apt-get install ceph-common
369 369

  
370
KVM userspace access
371
~~~~~~~~~~~~~~~~~~~~
372

  
373
If your cluster uses a sufficiently new version of KVM (you will need at
374
least QEMU 0.14 with RBD support compiled in), you can take advantage of
375
KVM's native support for ceph in order to have better performance and
376
avoid potential deadlocks_ in low memory scenarios.
377

  
378
.. _deadlocks: http://tracker.ceph.com/issues/3076
379

  
380
To initialize a cluster with support for this feature, use a command
381
such as::
382

  
383
  $ gnt-cluster init \
384
      --enabled-disk-templates rbd \
385
      --ipolicy-disk-templates rbd \
386
      --enabled-hypervisors=kvm \
387
      -D rbd:access=userspace
388

  
389
(You may want to enable more templates than just ``rbd``.)
390

  
391
You can also change this setting on a live cluster by giving the same
392
switches to ``gnt-cluster modify``, or change those settings at the node
393
group level with ``gnt-group modify``.
394

  
370 395
Configuration file
371 396
~~~~~~~~~~~~~~~~~~
372 397

  
b/man/gnt-cluster.rst
469 469
    When a new RADOS cluster is deployed, the default pool to put rbd
470 470
    volumes (Images in RADOS terminology) is 'rbd'.
471 471

  
472
access
473
    If 'userspace', instances will access their disks directly without
474
    going through a block device, avoiding expensive context switches
475
    with kernel space and the potential for deadlocks_ in low memory
476
    scenarios.
477

  
478
    The default value is 'kernelspace' and it disables this behaviour.
479
    This setting may only be changed to 'userspace' if all instance
480
    disks in the affected group or cluster can be accessed in userspace.
481

  
482
    Attempts to use this feature without rbd support compiled in KVM
483
    result in a "no such file or directory" error messages.
484

  
485
.. _deadlocks: http://tracker.ceph.com/issues/3076
486

  
472 487
The option ``--maintain-node-health`` allows one to enable/disable
473 488
automatic maintenance actions on nodes. Currently these include
474 489
automatic shutdown of instances and deactivation of DRBD devices on

Also available in: Unified diff