xref: /openbmc/qemu/docs/interop/bitmaps.rst (revision 542b10bd)
1..
2   Copyright 2019 John Snow <jsnow@redhat.com> and Red Hat, Inc.
3   All rights reserved.
4
5   This file is licensed via The FreeBSD Documentation License, the full
6   text of which is included at the end of this document.
7
8====================================
9Dirty Bitmaps and Incremental Backup
10====================================
11
12Dirty Bitmaps are in-memory objects that track writes to block devices. They
13can be used in conjunction with various block job operations to perform
14incremental or differential backup regimens.
15
16This document explains the conceptual mechanisms, as well as up-to-date,
17complete and comprehensive documentation on the API to manipulate them.
18(Hopefully, the "why", "what", and "how".)
19
20The intended audience for this document is developers who are adding QEMU
21backup features to management applications, or power users who run and
22administer QEMU directly via QMP.
23
24.. contents::
25
26Overview
27--------
28
29Bitmaps are bit vectors where each '1' bit in the vector indicates a modified
30("dirty") segment of the corresponding block device. The size of the segment
31that is tracked is the granularity of the bitmap. If the granularity of a
32bitmap is 64K, each '1' bit means that a 64K region as a whole may have
33changed in some way, possibly by as little as one byte.
34
35Smaller granularities mean more accurate tracking of modified disk data, but
36requires more computational overhead and larger bitmap sizes. Larger
37granularities mean smaller bitmap sizes, but less targeted backups.
38
39The size of a bitmap (in bytes) can be computed as such:
40    ``size`` = ceil(ceil(``image_size`` / ``granularity``) / 8)
41
42e.g. the size of a 64KiB granularity bitmap on a 2TiB image is:
43    ``size`` = ((2147483648K / 64K) / 8)
44         = 4194304B = 4MiB.
45
46QEMU uses these bitmaps when making incremental backups to know which sections
47of the file to copy out. They are not enabled by default and must be
48explicitly added in order to begin tracking writes.
49
50Bitmaps can be created at any time and can be attached to any arbitrary block
51node in the storage graph, but are most useful conceptually when attached to
52the root node attached to the guest's storage device model.
53
54That is to say: It's likely most useful to track the guest's writes to disk,
55but you could theoretically track things like qcow2 metadata changes by
56attaching the bitmap elsewhere in the storage graph. This is beyond the scope
57of this document.
58
59QEMU supports persisting these bitmaps to disk via the qcow2 image format.
60Bitmaps which are stored or loaded in this way are called "persistent",
61whereas bitmaps that are not are called "transient".
62
63QEMU also supports the migration of both transient bitmaps (tracking any
64arbitrary image format) or persistent bitmaps (qcow2) via live migration.
65
66Supported Image Formats
67-----------------------
68
69QEMU supports all documented features below on the qcow2 image format.
70
71However, qcow2 is only strictly necessary for the persistence feature, which
72writes bitmap data to disk upon close. If persistence is not required for a
73specific use case, all bitmap features excepting persistence are available for
74any arbitrary image format.
75
76For example, Dirty Bitmaps can be combined with the 'raw' image format, but
77any changes to the bitmap will be discarded upon exit.
78
79.. warning:: Transient bitmaps will not be saved on QEMU exit! Persistent
80             bitmaps are available only on qcow2 images.
81
82Dirty Bitmap Names
83------------------
84
85Bitmap objects need a method to reference them in the API. All API-created and
86managed bitmaps have a human-readable name chosen by the user at creation
87time.
88
89- A bitmap's name is unique to the node, but bitmaps attached to different
90  nodes can share the same name. Therefore, all bitmaps are addressed via
91  their (node, name) pair.
92
93- The name of a user-created bitmap cannot be empty ("").
94
95- Transient bitmaps can have JSON unicode names that are effectively not
96  length limited. (QMP protocol may restrict messages to less than 64MiB.)
97
98- Persistent storage formats may impose their own requirements on bitmap names
99  and namespaces. Presently, only qcow2 supports persistent bitmaps. See
100  docs/interop/qcow2.txt for more details on restrictions. Notably:
101
102   - qcow2 bitmap names are limited to between 1 and 1023 bytes long.
103
104   - No two bitmaps saved to the same qcow2 file may share the same name.
105
106- QEMU occasionally uses bitmaps for internal use which have no name. They are
107  hidden from API query calls, cannot be manipulated by the external API, are
108  never persistent, nor ever migrated.
109
110Bitmap Status
111-------------
112
113Dirty Bitmap objects can be queried with the QMP command `query-block
114<qemu-qmp-ref.html#index-query_002dblock>`_, and are visible via the
115`BlockDirtyInfo <qemu-qmp-ref.html#index-BlockDirtyInfo>`_ QAPI structure.
116
117This struct shows the name, granularity, and dirty byte count for each bitmap.
118Additionally, it shows several boolean status indicators:
119
120- ``recording``: This bitmap is recording writes.
121- ``busy``: This bitmap is in-use by an operation.
122- ``persistent``: This bitmap is a persistent type.
123- ``inconsistent``: This bitmap is corrupted and cannot be used.
124
125The ``+busy`` status prohibits you from deleting, clearing, or otherwise
126modifying a bitmap, and happens when the bitmap is being used for a backup
127operation or is in the process of being loaded from a migration. Many of the
128commands documented below will refuse to work on such bitmaps.
129
130The ``+inconsistent`` status similarly prohibits almost all operations,
131notably allowing only the ``block-dirty-bitmap-remove`` operation.
132
133There is also a deprecated ``status`` field of type `DirtyBitmapStatus
134<qemu-qmp-ref.html#index-DirtyBitmapStatus>`_. A bitmap historically had
135five visible states:
136
137   #. ``Frozen``: This bitmap is currently in-use by an operation and is
138      immutable. It can't be deleted, renamed, reset, etc.
139
140      (This is now ``+busy``.)
141
142   #. ``Disabled``: This bitmap is not recording new writes.
143
144      (This is now ``-recording -busy``.)
145
146   #. ``Active``: This bitmap is recording new writes.
147
148      (This is now ``+recording -busy``.)
149
150   #. ``Locked``: This bitmap is in-use by an operation, and is immutable.
151      The difference from "Frozen" was primarily implementation details.
152
153      (This is now ``+busy``.)
154
155   #. ``Inconsistent``: This persistent bitmap was not saved to disk
156      correctly, and can no longer be used. It remains in memory to serve as
157      an indicator of failure.
158
159      (This is now ``+inconsistent``.)
160
161These states are directly replaced by the status indicators and should not be
162used. The difference between ``Frozen`` and ``Locked`` is an implementation
163detail and should not be relevant to external users.
164
165Basic QMP Usage
166---------------
167
168The primary interface to manipulating bitmap objects is via the QMP
169interface. If you are not familiar, see the :doc:`qmp-spec` for the
170protocol, and :doc:`qemu-qmp-ref` for a full reference of all QMP
171commands.
172
173Supported Commands
174~~~~~~~~~~~~~~~~~~
175
176There are six primary bitmap-management API commands:
177
178- ``block-dirty-bitmap-add``
179- ``block-dirty-bitmap-remove``
180- ``block-dirty-bitmap-clear``
181- ``block-dirty-bitmap-disable``
182- ``block-dirty-bitmap-enable``
183- ``block-dirty-bitmap-merge``
184
185And one related query command:
186
187- ``query-block``
188
189Creation: block-dirty-bitmap-add
190~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
191
192`block-dirty-bitmap-add
193<qemu-qmp-ref.html#index-block_002ddirty_002dbitmap_002dadd>`_:
194
195Creates a new bitmap that tracks writes to the specified node. granularity,
196persistence, and recording state can be adjusted at creation time.
197
198.. admonition:: Example
199
200 to create a new, actively recording persistent bitmap:
201
202 .. code-block:: QMP
203
204  -> { "execute": "block-dirty-bitmap-add",
205       "arguments": {
206         "node": "drive0",
207         "name": "bitmap0",
208         "persistent": true,
209       }
210     }
211
212  <- { "return": {} }
213
214- This bitmap will have a default granularity that matches the cluster size of
215  its associated drive, if available, clamped to between [4KiB, 64KiB]. The
216  current default for qcow2 is 64KiB.
217
218.. admonition:: Example
219
220 To create a new, disabled (``-recording``), transient bitmap that tracks
221 changes in 32KiB segments:
222
223 .. code-block:: QMP
224
225  -> { "execute": "block-dirty-bitmap-add",
226       "arguments": {
227         "node": "drive0",
228         "name": "bitmap1",
229         "granularity": 32768,
230         "disabled": true
231       }
232     }
233
234  <- { "return": {} }
235
236Deletion: block-dirty-bitmap-remove
237~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
238
239`block-dirty-bitmap-remove
240<qemu-qmp-ref.html#index-block_002ddirty_002dbitmap_002dremove>`_:
241
242Deletes a bitmap. Bitmaps that are ``+busy`` cannot be removed.
243
244- Deleting a bitmap does not impact any other bitmaps attached to the same
245  node, nor does it affect any backups already created from this bitmap or
246  node.
247
248- Because bitmaps are only unique to the node to which they are attached, you
249  must specify the node/drive name here, too.
250
251- Deleting a persistent bitmap will remove it from the qcow2 file.
252
253.. admonition:: Example
254
255 Remove a bitmap named ``bitmap0`` from node ``drive0``:
256
257 .. code-block:: QMP
258
259  -> { "execute": "block-dirty-bitmap-remove",
260       "arguments": {
261         "node": "drive0",
262         "name": "bitmap0"
263       }
264     }
265
266  <- { "return": {} }
267
268Resetting: block-dirty-bitmap-clear
269~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
270
271`block-dirty-bitmap-clear
272<qemu-qmp-ref.html#index-block_002ddirty_002dbitmap_002dclear>`_:
273
274Clears all dirty bits from a bitmap. ``+busy`` bitmaps cannot be cleared.
275
276- An incremental backup created from an empty bitmap will copy no data, as if
277  nothing has changed.
278
279.. admonition:: Example
280
281 Clear all dirty bits from bitmap ``bitmap0`` on node ``drive0``:
282
283 .. code-block:: QMP
284
285  -> { "execute": "block-dirty-bitmap-clear",
286       "arguments": {
287         "node": "drive0",
288         "name": "bitmap0"
289       }
290     }
291
292  <- { "return": {} }
293
294Enabling: block-dirty-bitmap-enable
295~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
296
297`block-dirty-bitmap-enable
298<qemu-qmp-ref.html#index-block_002ddirty_002dbitmap_002denable>`_:
299
300"Enables" a bitmap, setting the ``recording`` bit to true, causing writes to
301begin being recorded. ``+busy`` bitmaps cannot be enabled.
302
303- Bitmaps default to being enabled when created, unless configured otherwise.
304
305- Persistent enabled bitmaps will remember their ``+recording`` status on
306  load.
307
308.. admonition:: Example
309
310 To set ``+recording`` on bitmap ``bitmap0`` on node ``drive0``:
311
312 .. code-block:: QMP
313
314  -> { "execute": "block-dirty-bitmap-enable",
315       "arguments": {
316         "node": "drive0",
317         "name": "bitmap0"
318       }
319     }
320
321  <- { "return": {} }
322
323Enabling: block-dirty-bitmap-disable
324~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
325
326`block-dirty-bitmap-disable
327<qemu-qmp-ref.html#index-block_002ddirty_002dbitmap_002ddisable>`_:
328
329"Disables" a bitmap, setting the ``recording`` bit to false, causing further
330writes to begin being ignored. ``+busy`` bitmaps cannot be disabled.
331
332.. warning::
333
334  This is potentially dangerous: QEMU makes no effort to stop any writes if
335  there are disabled bitmaps on a node, and will not mark any disabled bitmaps
336  as ``+inconsistent`` if any such writes do happen. Backups made from such
337  bitmaps will not be able to be used to reconstruct a coherent image.
338
339- Disabling a bitmap may be useful for examining which sectors of a disk
340  changed during a specific time period, or for explicit management of
341  differential backup windows.
342
343- Persistent disabled bitmaps will remember their ``-recording`` status on
344  load.
345
346.. admonition:: Example
347
348 To set ``-recording`` on bitmap ``bitmap0`` on node ``drive0``:
349
350 .. code-block:: QMP
351
352  -> { "execute": "block-dirty-bitmap-disable",
353       "arguments": {
354         "node": "drive0",
355         "name": "bitmap0"
356       }
357     }
358
359  <- { "return": {} }
360
361Merging, Copying: block-dirty-bitmap-merge
362~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
363
364`block-dirty-bitmap-merge
365<qemu-qmp-ref.html#index-block_002ddirty_002dbitmap_002dmerge>`_:
366
367Merges one or more bitmaps into a target bitmap. For any segment that is dirty
368in any one source bitmap, the target bitmap will mark that segment dirty.
369
370- Merge takes one or more bitmaps as a source and merges them together into a
371  single destination, such that any segment marked as dirty in any source
372  bitmap(s) will be marked dirty in the destination bitmap.
373
374- Merge does not create the destination bitmap if it does not exist. A blank
375  bitmap can be created beforehand to achieve the same effect.
376
377- The destination is not cleared prior to merge, so subsequent merge
378  operations will continue to cumulatively mark more segments as dirty.
379
380- If the merge operation should fail, the destination bitmap is guaranteed to
381  be unmodified. The operation may fail if the source or destination bitmaps
382  are busy, or have different granularities.
383
384- Bitmaps can only be merged on the same node. There is only one "node"
385  argument, so all bitmaps must be attached to that same node.
386
387- Copy can be achieved by merging from a single source to an empty
388  destination.
389
390.. admonition:: Example
391
392 Merge the data from ``bitmap0`` into the bitmap ``new_bitmap`` on node
393 ``drive0``. If ``new_bitmap`` was empty prior to this command, this achieves
394 a copy.
395
396 .. code-block:: QMP
397
398  -> { "execute": "block-dirty-bitmap-merge",
399       "arguments": {
400         "node": "drive0",
401         "target": "new_bitmap",
402         "bitmaps": [ "bitmap0" ]
403       }
404     }
405
406  <- { "return": {} }
407
408Querying: query-block
409~~~~~~~~~~~~~~~~~~~~~
410
411`query-block
412<qemu-qmp-ref.html#index-query_002dblock>`_:
413
414Not strictly a bitmaps command, but will return information about any bitmaps
415attached to nodes serving as the root for guest devices.
416
417- The "inconsistent" bit will not appear when it is false, appearing only when
418  the value is true to indicate there is a problem.
419
420.. admonition:: Example
421
422 Query the block sub-system of QEMU. The following json has trimmed irrelevant
423 keys from the response to highlight only the bitmap-relevant portions of the
424 API. This result highlights a bitmap ``bitmap0`` attached to the root node of
425 device ``drive0``.
426
427 .. code-block:: QMP
428
429  -> {
430       "execute": "query-block",
431       "arguments": {}
432     }
433
434  <- {
435       "return": [ {
436         "dirty-bitmaps": [ {
437           "status": "active",
438           "count": 0,
439           "busy": false,
440           "name": "bitmap0",
441           "persistent": false,
442           "recording": true,
443           "granularity": 65536
444         } ],
445         "device": "drive0",
446       } ]
447     }
448
449Bitmap Persistence
450------------------
451
452As outlined in `Supported Image Formats`_, QEMU can persist bitmaps to qcow2
453files. Demonstrated in `Creation: block-dirty-bitmap-add`_, passing
454``persistent: true`` to ``block-dirty-bitmap-add`` will persist that bitmap to
455disk.
456
457Persistent bitmaps will be automatically loaded into memory upon load, and
458will be written back to disk upon close. Their usage should be mostly
459transparent.
460
461However, if QEMU does not get a chance to close the file cleanly, the bitmap
462will be marked as ``+inconsistent`` at next load and considered unsafe to use
463for any operation. At this point, the only valid operation on such bitmaps is
464``block-dirty-bitmap-remove``.
465
466Losing a bitmap in this way does not invalidate any existing backups that have
467been made from this bitmap, but no further backups will be able to be issued
468for this chain.
469
470Transactions
471------------
472
473Transactions are a QMP feature that allows you to submit multiple QMP commands
474at once, being guaranteed that they will all succeed or fail atomically,
475together. The interaction of bitmaps and transactions are demonstrated below.
476
477See `transaction <qemu-qmp.ref.html#index-transaction>`_ in the QMP reference
478for more details.
479
480Justification
481~~~~~~~~~~~~~
482
483Bitmaps can generally be modified at any time, but certain operations often
484only make sense when paired directly with other commands. When a VM is paused,
485it's easy to ensure that no guest writes occur between individual QMP
486commands. When a VM is running, this is difficult to accomplish with
487individual QMP commands that may allow guest writes to occur between each
488command.
489
490For example, using only individual QMP commands, we could:
491
492#. Boot the VM in a paused state.
493#. Create a full drive backup of drive0.
494#. Create a new bitmap attached to drive0, confident that nothing has been
495   written to drive0 in the meantime.
496#. Resume execution of the VM.
497#. At a later point, issue incremental backups from ``bitmap0``.
498
499At this point, the bitmap and drive backup would be correctly in sync, and
500incremental backups made from this point forward would be correctly aligned to
501the full drive backup.
502
503This is not particularly useful if we decide we want to start incremental
504backups after the VM has been running for a while, for which we would want to
505perform actions such as the following:
506
507#. Boot the VM and begin execution.
508#. Using a single transaction, perform the following operations:
509
510   -  Create ``bitmap0``.
511   -  Create a full drive backup of ``drive0``.
512
513#. At a later point, issue incremental backups from ``bitmap0``.
514
515.. note:: As a consideration, if ``bitmap0`` is created prior to the full
516          drive backup, incremental backups can still be authored from this
517          bitmap, but they will copy extra segments reflecting writes that
518          occurred prior to the backup operation. Transactions allow us to
519          narrow critical points in time to reduce waste, or, in the other
520          direction, to ensure that no segments are omitted.
521
522Supported Bitmap Transactions
523~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
524
525-  ``block-dirty-bitmap-add``
526-  ``block-dirty-bitmap-clear``
527-  ``block-dirty-bitmap-enable``
528-  ``block-dirty-bitmap-disable``
529-  ``block-dirty-bitmap-merge``
530
531The usages for these commands are identical to their respective QMP commands,
532but see the sections below for concrete examples.
533
534Incremental Backups - Push Model
535--------------------------------
536
537Incremental backups are simply partial disk images that can be combined with
538other partial disk images on top of a base image to reconstruct a full backup
539from the point in time at which the incremental backup was issued.
540
541The "Push Model" here references the fact that QEMU is "pushing" the modified
542blocks out to a destination. We will be using the  `blockdev-backup
543<qemu-qmp-ref.html#index-blockdev_002dbackup>`_ QMP command to create both
544full and incremental backups.
545
546The command is a background job, which has its own QMP API for querying and
547management documented in `Background jobs
548<qemu-qmp-ref.html#Background-jobs>`_.
549
550Example: New Incremental Backup Anchor Point
551~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
552
553As outlined in the Transactions - `Justification`_ section, perhaps we want to
554create a new incremental backup chain attached to a drive.
555
556This example creates a new, full backup of "drive0" and accompanies it with a
557new, empty bitmap that records writes from this point in time forward.
558
559The target can be created with the help of `blockdev-add
560<qemu-qmp-ref.html#index-blockdev_002dadd>`_ or `blockdev-create
561<qemu-qmp-ref.html#index-blockdev_002dcreate>`_ command.
562
563.. note:: Any new writes that happen after this command is issued, even while
564          the backup job runs, will be written locally and not to the backup
565          destination. These writes will be recorded in the bitmap
566          accordingly.
567
568.. code-block:: QMP
569
570  -> {
571       "execute": "transaction",
572       "arguments": {
573         "actions": [
574           {
575             "type": "block-dirty-bitmap-add",
576             "data": {
577               "node": "drive0",
578               "name": "bitmap0"
579             }
580           },
581           {
582             "type": "blockdev-backup",
583             "data": {
584               "device": "drive0",
585               "target": "target0",
586               "sync": "full"
587             }
588           }
589         ]
590       }
591     }
592
593  <- { "return": {} }
594
595  <- {
596       "timestamp": {
597         "seconds": 1555436945,
598         "microseconds": 179620
599       },
600       "data": {
601         "status": "created",
602         "id": "drive0"
603       },
604       "event": "JOB_STATUS_CHANGE"
605     }
606
607  ...
608
609  <- {
610       "timestamp": {...},
611       "data": {
612         "device": "drive0",
613         "type": "backup",
614         "speed": 0,
615         "len": 68719476736,
616         "offset": 68719476736
617       },
618       "event": "BLOCK_JOB_COMPLETED"
619     }
620
621  <- {
622       "timestamp": {...},
623       "data": {
624         "status": "concluded",
625         "id": "drive0"
626       },
627       "event": "JOB_STATUS_CHANGE"
628     }
629
630  <- {
631       "timestamp": {...},
632       "data": {
633         "status": "null",
634         "id": "drive0"
635       },
636       "event": "JOB_STATUS_CHANGE"
637     }
638
639A full explanation of the job transition semantics and the JOB_STATUS_CHANGE
640event are beyond the scope of this document and will be omitted in all
641subsequent examples; above, several more events have been omitted for brevity.
642
643.. note:: Subsequent examples will omit all events except BLOCK_JOB_COMPLETED
644          except where necessary to illustrate workflow differences.
645
646          Omitted events and json objects will be represented by ellipses:
647          ``...``
648
649Example: Resetting an Incremental Backup Anchor Point
650~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
651
652If we want to start a new backup chain with an existing bitmap, we can also
653use a transaction to reset the bitmap while making a new full backup:
654
655.. code-block:: QMP
656
657  -> {
658       "execute": "transaction",
659       "arguments": {
660         "actions": [
661         {
662           "type": "block-dirty-bitmap-clear",
663           "data": {
664             "node": "drive0",
665             "name": "bitmap0"
666           }
667         },
668         {
669           "type": "blockdev-backup",
670           "data": {
671             "device": "drive0",
672             "target": "target0",
673             "sync": "full"
674           }
675         }
676       ]
677     }
678   }
679
680  <- { "return": {} }
681
682  ...
683
684  <- {
685       "timestamp": {...},
686       "data": {
687         "device": "drive0",
688         "type": "backup",
689         "speed": 0,
690         "len": 68719476736,
691         "offset": 68719476736
692       },
693       "event": "BLOCK_JOB_COMPLETED"
694     }
695
696  ...
697
698The result of this example is identical to the first, but we clear an existing
699bitmap instead of adding a new one.
700
701.. tip:: In both of these examples, "bitmap0" is tied conceptually to the
702         creation of new, full backups. This relationship is not saved or
703         remembered by QEMU; it is up to the operator or management layer to
704         remember which bitmaps are associated with which backups.
705
706Example: First Incremental Backup
707~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
708
709#. Create a full backup and sync it to a dirty bitmap using any method:
710
711   - Either of the two live backup method demonstrated above,
712   - Using QMP commands with the VM paused as in the `Justification`_ section,
713     or
714   - With the VM offline, manually copy the image and start the VM in a paused
715     state, careful to add a new bitmap before the VM begins execution.
716
717   Whichever method is chosen, let's assume that at the end of this step:
718
719   - The full backup is named ``drive0.full.qcow2``.
720   - The bitmap we created is named ``bitmap0``, attached to ``drive0``.
721
722#. Create a destination image for the incremental backup that utilizes the
723   full backup as a backing image.
724
725   - Let's assume the new incremental image is named ``drive0.inc0.qcow2``:
726
727   .. code:: bash
728
729       $ qemu-img create -f qcow2 drive0.inc0.qcow2 \
730         -b drive0.full.qcow2 -F qcow2
731
732#. Add target block node:
733
734   .. code-block:: QMP
735
736    -> {
737         "execute": "blockdev-add",
738         "arguments": {
739           "node-name": "target0",
740           "driver": "qcow2",
741           "file": {
742             "driver": "file",
743             "filename": "drive0.inc0.qcow2"
744           }
745         }
746       }
747
748    <- { "return": {} }
749
750#. Issue an incremental backup command:
751
752   .. code-block:: QMP
753
754    -> {
755         "execute": "blockdev-backup",
756         "arguments": {
757           "device": "drive0",
758           "bitmap": "bitmap0",
759           "target": "target0",
760           "sync": "incremental"
761         }
762       }
763
764    <- { "return": {} }
765
766    ...
767
768    <- {
769         "timestamp": {...},
770         "data": {
771           "device": "drive0",
772           "type": "backup",
773           "speed": 0,
774           "len": 68719476736,
775           "offset": 68719476736
776         },
777         "event": "BLOCK_JOB_COMPLETED"
778       }
779
780    ...
781
782This copies any blocks modified since the full backup was created into the
783``drive0.inc0.qcow2`` file. During the operation, ``bitmap0`` is marked
784``+busy``. If the operation is successful, ``bitmap0`` will be cleared to
785reflect the "incremental" backup regimen, which only copies out new changes
786from each incremental backup.
787
788.. note:: Any new writes that occur after the backup operation starts do not
789          get copied to the destination. The backup's "point in time" is when
790          the backup starts, not when it ends. These writes are recorded in a
791          special bitmap that gets re-added to bitmap0 when the backup ends so
792          that the next incremental backup can copy them out.
793
794Example: Second Incremental Backup
795~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
796
797#. Create a new destination image for the incremental backup that points to
798   the previous one, e.g.: ``drive0.inc1.qcow2``
799
800   .. code:: bash
801
802       $ qemu-img create -f qcow2 drive0.inc1.qcow2 \
803         -b drive0.inc0.qcow2 -F qcow2
804
805#. Add target block node:
806
807   .. code-block:: QMP
808
809    -> {
810         "execute": "blockdev-add",
811         "arguments": {
812           "node-name": "target0",
813           "driver": "qcow2",
814           "file": {
815             "driver": "file",
816             "filename": "drive0.inc1.qcow2"
817           }
818         }
819       }
820
821    <- { "return": {} }
822
823#. Issue a new incremental backup command. The only difference here is that we
824   have changed the target image below.
825
826   .. code-block:: QMP
827
828    -> {
829         "execute": "blockdev-backup",
830         "arguments": {
831           "device": "drive0",
832           "bitmap": "bitmap0",
833           "target": "target0",
834           "sync": "incremental"
835         }
836       }
837
838    <- { "return": {} }
839
840    ...
841
842    <- {
843         "timestamp": {...},
844         "data": {
845           "device": "drive0",
846           "type": "backup",
847           "speed": 0,
848           "len": 68719476736,
849           "offset": 68719476736
850         },
851         "event": "BLOCK_JOB_COMPLETED"
852       }
853
854    ...
855
856Because the first incremental backup from the previous example completed
857successfully, ``bitmap0`` was synchronized with ``drive0.inc0.qcow2``. Here,
858we use ``bitmap0`` again to create a new incremental backup that targets the
859previous one, creating a chain of three images:
860
861.. admonition:: Diagram
862
863 .. code:: text
864
865   +-------------------+   +-------------------+   +-------------------+
866   | drive0.full.qcow2 |<--| drive0.inc0.qcow2 |<--| drive0.inc1.qcow2 |
867   +-------------------+   +-------------------+   +-------------------+
868
869Each new incremental backup re-synchronizes the bitmap to the latest backup
870authored, allowing a user to continue to "consume" it to create new backups on
871top of an existing chain.
872
873In the above diagram, neither drive0.inc1.qcow2 nor drive0.inc0.qcow2 are
874complete images by themselves, but rely on their backing chain to reconstruct
875a full image. The dependency terminates with each full backup.
876
877Each backup in this chain remains independent, and is unchanged by new entries
878made later in the chain. For instance, drive0.inc0.qcow2 remains a perfectly
879valid backup of the disk as it was when that backup was issued.
880
881Example: Incremental Push Backups without Backing Files
882~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
883
884Backup images are best kept off-site, so we often will not have the preceding
885backups in a chain available to link against. This is not a problem at backup
886time; we simply do not set the backing image when creating the destination
887image:
888
889#. Create a new destination image with no backing file set. We will need to
890   specify the size of the base image, because the backing file isn't
891   available for QEMU to use to determine it.
892
893   .. code:: bash
894
895       $ qemu-img create -f qcow2 drive0.inc2.qcow2 64G
896
897   .. note:: Alternatively, you can omit ``mode: "existing"`` from the push
898             backup commands to have QEMU create an image without a backing
899             file for you, but you lose control over format options like
900             compatibility and preallocation presets.
901
902#. Add target block node:
903
904   .. code-block:: QMP
905
906    -> {
907         "execute": "blockdev-add",
908         "arguments": {
909           "node-name": "target0",
910           "driver": "qcow2",
911           "file": {
912             "driver": "file",
913             "filename": "drive0.inc2.qcow2"
914           }
915         }
916       }
917
918    <- { "return": {} }
919
920#. Issue a new incremental backup command. Apart from the new destination
921   image, there is no difference from the last two examples.
922
923   .. code-block:: QMP
924
925    -> {
926         "execute": "blockdev-backup",
927         "arguments": {
928           "device": "drive0",
929           "bitmap": "bitmap0",
930           "target": "target0",
931           "sync": "incremental"
932         }
933       }
934
935    <- { "return": {} }
936
937    ...
938
939    <- {
940         "timestamp": {...},
941         "data": {
942           "device": "drive0",
943           "type": "backup",
944           "speed": 0,
945           "len": 68719476736,
946           "offset": 68719476736
947         },
948         "event": "BLOCK_JOB_COMPLETED"
949       }
950
951    ...
952
953The only difference from the perspective of the user is that you will need to
954set the backing image when attempting to restore the backup:
955
956.. code:: bash
957
958    $ qemu-img rebase drive0.inc2.qcow2 \
959      -u -b drive0.inc1.qcow2
960
961This uses the "unsafe" rebase mode to simply set the backing file to a file
962that isn't present.
963
964It is also possible to use ``--image-opts`` to specify the entire backing
965chain by hand as an ephemeral property at runtime, but that is beyond the
966scope of this document.
967
968Example: Multi-drive Incremental Backup
969~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
970
971Assume we have a VM with two drives, "drive0" and "drive1" and we wish to back
972both of them up such that the two backups represent the same crash-consistent
973point in time.
974
975#. For each drive, create an empty image:
976
977   .. code:: bash
978
979    $ qemu-img create -f qcow2 drive0.full.qcow2 64G
980    $ qemu-img create -f qcow2 drive1.full.qcow2 64G
981
982#. Add target block nodes:
983
984   .. code-block:: QMP
985
986    -> {
987         "execute": "blockdev-add",
988         "arguments": {
989           "node-name": "target0",
990           "driver": "qcow2",
991           "file": {
992             "driver": "file",
993             "filename": "drive0.full.qcow2"
994           }
995         }
996       }
997
998    <- { "return": {} }
999
1000    -> {
1001         "execute": "blockdev-add",
1002         "arguments": {
1003           "node-name": "target1",
1004           "driver": "qcow2",
1005           "file": {
1006             "driver": "file",
1007             "filename": "drive1.full.qcow2"
1008           }
1009         }
1010       }
1011
1012    <- { "return": {} }
1013
1014#. Create a full (anchor) backup for each drive, with accompanying bitmaps:
1015
1016   .. code-block:: QMP
1017
1018    -> {
1019         "execute": "transaction",
1020         "arguments": {
1021           "actions": [
1022             {
1023               "type": "block-dirty-bitmap-add",
1024               "data": {
1025                 "node": "drive0",
1026                 "name": "bitmap0"
1027               }
1028             },
1029             {
1030               "type": "block-dirty-bitmap-add",
1031               "data": {
1032                 "node": "drive1",
1033                 "name": "bitmap0"
1034               }
1035             },
1036             {
1037               "type": "blockdev-backup",
1038               "data": {
1039                 "device": "drive0",
1040                 "target": "target0",
1041                 "sync": "full"
1042               }
1043             },
1044             {
1045               "type": "blockdev-backup",
1046               "data": {
1047                 "device": "drive1",
1048                 "target": "target1",
1049                 "sync": "full"
1050               }
1051             }
1052           ]
1053         }
1054       }
1055
1056    <- { "return": {} }
1057
1058    ...
1059
1060    <- {
1061         "timestamp": {...},
1062         "data": {
1063           "device": "drive0",
1064           "type": "backup",
1065           "speed": 0,
1066           "len": 68719476736,
1067           "offset": 68719476736
1068         },
1069         "event": "BLOCK_JOB_COMPLETED"
1070       }
1071
1072    ...
1073
1074    <- {
1075         "timestamp": {...},
1076         "data": {
1077           "device": "drive1",
1078           "type": "backup",
1079           "speed": 0,
1080           "len": 68719476736,
1081           "offset": 68719476736
1082         },
1083         "event": "BLOCK_JOB_COMPLETED"
1084       }
1085
1086    ...
1087
1088#. Later, create new destination images for each of the incremental backups
1089   that point to their respective full backups:
1090
1091   .. code:: bash
1092
1093     $ qemu-img create -f qcow2 drive0.inc0.qcow2 \
1094       -b drive0.full.qcow2 -F qcow2
1095     $ qemu-img create -f qcow2 drive1.inc0.qcow2 \
1096       -b drive1.full.qcow2 -F qcow2
1097
1098#. Add target block nodes:
1099
1100   .. code-block:: QMP
1101
1102    -> {
1103         "execute": "blockdev-add",
1104         "arguments": {
1105           "node-name": "target0",
1106           "driver": "qcow2",
1107           "file": {
1108             "driver": "file",
1109             "filename": "drive0.inc0.qcow2"
1110           }
1111         }
1112       }
1113
1114    <- { "return": {} }
1115
1116    -> {
1117         "execute": "blockdev-add",
1118         "arguments": {
1119           "node-name": "target1",
1120           "driver": "qcow2",
1121           "file": {
1122             "driver": "file",
1123             "filename": "drive1.inc0.qcow2"
1124           }
1125         }
1126       }
1127
1128    <- { "return": {} }
1129
1130#. Issue a multi-drive incremental push backup transaction:
1131
1132   .. code-block:: QMP
1133
1134    -> {
1135         "execute": "transaction",
1136         "arguments": {
1137           "actions": [
1138             {
1139               "type": "blockev-backup",
1140               "data": {
1141                 "device": "drive0",
1142                 "bitmap": "bitmap0",
1143                 "sync": "incremental",
1144                 "target": "target0"
1145               }
1146             },
1147             {
1148               "type": "blockdev-backup",
1149               "data": {
1150                 "device": "drive1",
1151                 "bitmap": "bitmap0",
1152                 "sync": "incremental",
1153                 "target": "target1"
1154               }
1155             },
1156           ]
1157         }
1158       }
1159
1160    <- { "return": {} }
1161
1162    ...
1163
1164    <- {
1165         "timestamp": {...},
1166         "data": {
1167           "device": "drive0",
1168           "type": "backup",
1169           "speed": 0,
1170           "len": 68719476736,
1171           "offset": 68719476736
1172         },
1173         "event": "BLOCK_JOB_COMPLETED"
1174       }
1175
1176    ...
1177
1178    <- {
1179         "timestamp": {...},
1180         "data": {
1181           "device": "drive1",
1182           "type": "backup",
1183           "speed": 0,
1184           "len": 68719476736,
1185           "offset": 68719476736
1186         },
1187         "event": "BLOCK_JOB_COMPLETED"
1188       }
1189
1190    ...
1191
1192Push Backup Errors & Recovery
1193-----------------------------
1194
1195In the event of an error that occurs after a push backup job is successfully
1196launched, either by an individual QMP command or a QMP transaction, the user
1197will receive a ``BLOCK_JOB_COMPLETE`` event with a failure message,
1198accompanied by a ``BLOCK_JOB_ERROR`` event.
1199
1200In the case of a job being cancelled, the user will receive a
1201``BLOCK_JOB_CANCELLED`` event instead of a pair of COMPLETE and ERROR
1202events.
1203
1204In either failure case, the bitmap used for the failed operation is not
1205cleared. It will contain all of the dirty bits it did at the start of the
1206operation, plus any new bits that got marked during the operation.
1207
1208Effectively, the "point in time" that a bitmap is recording differences
1209against is kept at the issuance of the last successful incremental backup,
1210instead of being moved forward to the start of this now-failed backup.
1211
1212Once the underlying problem is addressed (e.g. more storage space is allocated
1213on the destination), the incremental backup command can be retried with the
1214same bitmap.
1215
1216Example: Individual Failures
1217~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1218
1219Incremental Push Backup jobs that fail individually behave simply as
1220described above. This example demonstrates the single-job failure case:
1221
1222#. Create a target image:
1223
1224   .. code:: bash
1225
1226       $ qemu-img create -f qcow2 drive0.inc0.qcow2 \
1227         -b drive0.full.qcow2 -F qcow2
1228
1229#. Add target block node:
1230
1231   .. code-block:: QMP
1232
1233    -> {
1234         "execute": "blockdev-add",
1235         "arguments": {
1236           "node-name": "target0",
1237           "driver": "qcow2",
1238           "file": {
1239             "driver": "file",
1240             "filename": "drive0.inc0.qcow2"
1241           }
1242         }
1243       }
1244
1245    <- { "return": {} }
1246
1247#. Attempt to create an incremental backup via QMP:
1248
1249   .. code-block:: QMP
1250
1251    -> {
1252         "execute": "blockdev-backup",
1253         "arguments": {
1254           "device": "drive0",
1255           "bitmap": "bitmap0",
1256           "target": "target0",
1257           "sync": "incremental"
1258         }
1259       }
1260
1261    <- { "return": {} }
1262
1263#. Receive a pair of events indicating failure:
1264
1265   .. code-block:: QMP
1266
1267    <- {
1268         "timestamp": {...},
1269         "data": {
1270           "device": "drive0",
1271           "action": "report",
1272           "operation": "write"
1273         },
1274         "event": "BLOCK_JOB_ERROR"
1275       }
1276
1277    <- {
1278         "timestamp": {...},
1279         "data": {
1280           "speed": 0,
1281           "offset": 0,
1282           "len": 67108864,
1283           "error": "No space left on device",
1284           "device": "drive0",
1285           "type": "backup"
1286         },
1287         "event": "BLOCK_JOB_COMPLETED"
1288       }
1289
1290#. Remove target node:
1291
1292   .. code-block:: QMP
1293
1294    -> {
1295         "execute": "blockdev-del",
1296         "arguments": {
1297           "node-name": "target0",
1298         }
1299       }
1300
1301    <- { "return": {} }
1302
1303#. Delete the failed image, and re-create it.
1304
1305   .. code:: bash
1306
1307       $ rm drive0.inc0.qcow2
1308       $ qemu-img create -f qcow2 drive0.inc0.qcow2 \
1309         -b drive0.full.qcow2 -F qcow2
1310
1311#. Add target block node:
1312
1313   .. code-block:: QMP
1314
1315    -> {
1316         "execute": "blockdev-add",
1317         "arguments": {
1318           "node-name": "target0",
1319           "driver": "qcow2",
1320           "file": {
1321             "driver": "file",
1322             "filename": "drive0.inc0.qcow2"
1323           }
1324         }
1325       }
1326
1327    <- { "return": {} }
1328
1329#. Retry the command after fixing the underlying problem, such as
1330   freeing up space on the backup volume:
1331
1332   .. code-block:: QMP
1333
1334    -> {
1335         "execute": "blockdev-backup",
1336         "arguments": {
1337           "device": "drive0",
1338           "bitmap": "bitmap0",
1339           "target": "target0",
1340           "sync": "incremental"
1341         }
1342       }
1343
1344    <- { "return": {} }
1345
1346#. Receive confirmation that the job completed successfully:
1347
1348   .. code-block:: QMP
1349
1350    <- {
1351         "timestamp": {...},
1352         "data": {
1353           "device": "drive0",
1354           "type": "backup",
1355           "speed": 0,
1356           "len": 67108864,
1357           "offset": 67108864
1358         },
1359         "event": "BLOCK_JOB_COMPLETED"
1360       }
1361
1362Example: Partial Transactional Failures
1363~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1364
1365QMP commands like `blockdev-backup
1366<qemu-qmp-ref.html#index-blockdev_002dbackup>`_
1367conceptually only start a job, and so transactions containing these commands
1368may succeed even if the job it created later fails. This might have surprising
1369interactions with notions of how a "transaction" ought to behave.
1370
1371This distinction means that on occasion, a transaction containing such job
1372launching commands may appear to succeed and return success, but later
1373individual jobs associated with the transaction may fail. It is possible that
1374a management application may have to deal with a partial backup failure after
1375a "successful" transaction.
1376
1377If multiple backup jobs are specified in a single transaction, if one of those
1378jobs fails, it will not interact with the other backup jobs in any way by
1379default. The job(s) that succeeded will clear the dirty bitmap associated with
1380the operation, but the job(s) that failed will not. It is therefore not safe
1381to delete any incremental backups that were created successfully in this
1382scenario, even though others failed.
1383
1384This example illustrates a transaction with two backup jobs, where one fails
1385and one succeeds:
1386
1387#. Issue the transaction to start a backup of both drives.
1388
1389   .. code-block:: QMP
1390
1391    -> {
1392         "execute": "transaction",
1393         "arguments": {
1394           "actions": [
1395           {
1396             "type": "blockdev-backup",
1397             "data": {
1398               "device": "drive0",
1399               "bitmap": "bitmap0",
1400               "sync": "incremental",
1401               "target": "target0"
1402             }
1403           },
1404           {
1405             "type": "blockdev-backup",
1406             "data": {
1407               "device": "drive1",
1408               "bitmap": "bitmap0",
1409               "sync": "incremental",
1410               "target": "target1"
1411             }
1412           }]
1413         }
1414       }
1415
1416#. Receive notice that the Transaction was accepted, and jobs were
1417   launched:
1418
1419   .. code-block:: QMP
1420
1421    <- { "return": {} }
1422
1423#. Receive notice that the first job has completed:
1424
1425   .. code-block:: QMP
1426
1427    <- {
1428         "timestamp": {...},
1429         "data": {
1430           "device": "drive0",
1431           "type": "backup",
1432           "speed": 0,
1433           "len": 67108864,
1434           "offset": 67108864
1435         },
1436         "event": "BLOCK_JOB_COMPLETED"
1437       }
1438
1439#. Receive notice that the second job has failed:
1440
1441   .. code-block:: QMP
1442
1443    <- {
1444         "timestamp": {...},
1445         "data": {
1446           "device": "drive1",
1447           "action": "report",
1448           "operation": "read"
1449         },
1450         "event": "BLOCK_JOB_ERROR"
1451       }
1452
1453    ...
1454
1455    <- {
1456         "timestamp": {...},
1457         "data": {
1458           "speed": 0,
1459           "offset": 0,
1460           "len": 67108864,
1461           "error": "Input/output error",
1462           "device": "drive1",
1463           "type": "backup"
1464         },
1465         "event": "BLOCK_JOB_COMPLETED"
1466       }
1467
1468At the conclusion of the above example, ``drive0.inc0.qcow2`` is valid and
1469must be kept, but ``drive1.inc0.qcow2`` is incomplete and should be
1470deleted. If a VM-wide incremental backup of all drives at a point-in-time is
1471to be made, new backups for both drives will need to be made, taking into
1472account that a new incremental backup for drive0 needs to be based on top of
1473``drive0.inc0.qcow2``.
1474
1475For this example, an incremental backup for ``drive0`` was created, but not
1476for ``drive1``. The last VM-wide crash-consistent backup that is available in
1477this case is the full backup:
1478
1479.. code:: text
1480
1481          [drive0.full.qcow2] <-- [drive0.inc0.qcow2]
1482          [drive1.full.qcow2]
1483
1484To repair this, issue a new incremental backup across both drives. The result
1485will be backup chains that resemble the following:
1486
1487.. code:: text
1488
1489          [drive0.full.qcow2] <-- [drive0.inc0.qcow2] <-- [drive0.inc1.qcow2]
1490          [drive1.full.qcow2] <-------------------------- [drive1.inc1.qcow2]
1491
1492Example: Grouped Completion Mode
1493~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1494
1495While jobs launched by transactions normally complete or fail individually,
1496it's possible to instruct them to complete or fail together as a group. QMP
1497transactions take an optional properties structure that can affect the
1498behavior of the transaction.
1499
1500The ``completion-mode`` transaction property can be either ``individual``
1501which is the default legacy behavior described above, or ``grouped``, detailed
1502below.
1503
1504In ``grouped`` completion mode, no jobs will report success until all jobs are
1505ready to report success. If any job fails, all other jobs will be cancelled.
1506
1507Regardless of if a participating incremental backup job failed or was
1508cancelled, their associated bitmaps will all be held at their existing
1509points-in-time, as in individual failure cases.
1510
1511Here's the same multi-drive backup scenario from `Example: Partial
1512Transactional Failures`_, but with the ``grouped`` completion-mode property
1513applied:
1514
1515#. Issue the multi-drive incremental backup transaction:
1516
1517   .. code-block:: QMP
1518
1519    -> {
1520         "execute": "transaction",
1521         "arguments": {
1522           "properties": {
1523             "completion-mode": "grouped"
1524           },
1525           "actions": [
1526           {
1527             "type": "blockdev-backup",
1528             "data": {
1529               "device": "drive0",
1530               "bitmap": "bitmap0",
1531               "sync": "incremental",
1532               "target": "target0"
1533             }
1534           },
1535           {
1536             "type": "blockdev-backup",
1537             "data": {
1538               "device": "drive1",
1539               "bitmap": "bitmap0",
1540               "sync": "incremental",
1541               "target": "target1"
1542             }
1543           }]
1544         }
1545       }
1546
1547#. Receive notice that the Transaction was accepted, and jobs were launched:
1548
1549   .. code-block:: QMP
1550
1551    <- { "return": {} }
1552
1553#. Receive notification that the backup job for ``drive1`` has failed:
1554
1555   .. code-block:: QMP
1556
1557    <- {
1558         "timestamp": {...},
1559         "data": {
1560           "device": "drive1",
1561           "action": "report",
1562           "operation": "read"
1563         },
1564         "event": "BLOCK_JOB_ERROR"
1565       }
1566
1567    <- {
1568         "timestamp": {...},
1569         "data": {
1570           "speed": 0,
1571           "offset": 0,
1572           "len": 67108864,
1573           "error": "Input/output error",
1574           "device": "drive1",
1575           "type": "backup"
1576         },
1577         "event": "BLOCK_JOB_COMPLETED"
1578       }
1579
1580#. Receive notification that the job for ``drive0`` has been cancelled:
1581
1582   .. code-block:: QMP
1583
1584    <- {
1585         "timestamp": {...},
1586         "data": {
1587           "device": "drive0",
1588           "type": "backup",
1589           "speed": 0,
1590           "len": 67108864,
1591           "offset": 16777216
1592         },
1593         "event": "BLOCK_JOB_CANCELLED"
1594       }
1595
1596At the conclusion of *this* example, both jobs have been aborted due to a
1597failure. Both destination images should be deleted and are no longer of use.
1598
1599The transaction as a whole can simply be re-issued at a later time.
1600
1601.. raw:: html
1602
1603   <!--
1604   The FreeBSD Documentation License
1605
1606   Redistribution and use in source (ReST) and 'compiled' forms (SGML, HTML,
1607   PDF, PostScript, RTF and so forth) with or without modification, are
1608   permitted provided that the following conditions are met:
1609
1610   Redistributions of source code (ReST) must retain the above copyright notice,
1611   this list of conditions and the following disclaimer of this file unmodified.
1612
1613   Redistributions in compiled form (transformed to other DTDs, converted to
1614   PDF, PostScript, RTF and other formats) must reproduce the above copyright
1615   notice, this list of conditions and the following disclaimer in the
1616   documentation and/or other materials provided with the distribution.
1617
1618   THIS DOCUMENTATION IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
1619   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
1620   IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
1621   ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
1622   LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
1623   CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
1624   SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
1625   INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
1626   CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
1627   ARISING IN ANY WAY OUT OF THE USE OF THIS DOCUMENTATION, EVEN IF ADVISED OF
1628   THE POSSIBILITY OF SUCH DAMAGE.
1629   -->
1630