summaryrefslogtreecommitdiff
path: root/Documentation/admin-guide/device-mapper/dm-zoned.rst
blob: e635041351bc779afe9906780937707969d292be (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
========
dm-zoned
========

The dm-zoned device mapper target exposes a zoned block device (ZBC and
ZAC compliant devices) as a regular block device without any write
pattern constraints. In effect, it implements a drive-managed zoned
block device which hides from the user (a file system or an application
doing raw block device accesses) the sequential write constraints of
host-managed zoned block devices and can mitigate the potential
device-side performance degradation due to excessive random writes on
host-aware zoned block devices.

For a more detailed description of the zoned block device models and
their constraints see (for SCSI devices):

https://www.t10.org/drafts.htm#ZBC_Family

and (for ATA devices):

http://www.t13.org/Documents/UploadedDocuments/docs2015/di537r05-Zoned_Device_ATA_Command_Set_ZAC.pdf

The dm-zoned implementation is simple and minimizes system overhead (CPU
and memory usage as well as storage capacity loss). For a 10TB
host-managed disk with 256 MB zones, dm-zoned memory usage per disk
instance is at most 4.5 MB and as little as 5 zones will be used
internally for storing metadata and performaing reclaim operations.

dm-zoned target devices are formatted and checked using the dmzadm
utility available at:

https://github.com/hgst/dm-zoned-tools

Algorithm
=========

dm-zoned implements an on-disk buffering scheme to handle non-sequential
write accesses to the sequential zones of a zoned block device.
Conventional zones are used for caching as well as for storing internal
metadata. It can also use a regular block device together with the zoned
block device; in that case the regular block device will be split logically
in zones with the same size as the zoned block device. These zones will be
placed in front of the zones from the zoned block device and will be handled
just like conventional zones.

The zones of the device(s) are separated into 2 types:

1) Metadata zones: these are conventional zones used to store metadata.
Metadata zones are not reported as useable capacity to the user.

2) Data zones: all remaining zones, the vast majority of which will be
sequential zones used exclusively to store user data. The conventional
zones of the device may be used also for buffering user random writes.
Data in these zones may be directly mapped to the conventional zone, but
later moved to a sequential zone so that the conventional zone can be
reused for buffering incoming random writes.

dm-zoned exposes a logical device with a sector size of 4096 bytes,
irrespective of the physical sector size of the backend zoned block
device being used. This allows reducing the amount of metadata needed to
manage valid blocks (blocks written).

The on-disk metadata format is as follows:

1) The first block of the first conventional zone found contains the
super block which describes the on disk amount and position of metadata
blocks.

2) Following the super block, a set of blocks is used to describe the
mapping of the logical device blocks. The mapping is done per chunk of
blocks, with the chunk size equal to the zoned block device size. The
mapping table is indexed by chunk number and each mapping entry
indicates the zone number of the device storing the chunk of data. Each
mapping entry may also indicate if the zone number of a conventional
zone used to buffer random modification to the data zone.

3) A set of blocks used to store bitmaps indicating the validity of
blocks in the data zones follows the mapping table. A valid block is
defined as a block that was written and not discarded. For a buffered
data chunk, a block is always valid only in the data zone mapping the
chunk or in the buffer zone of the chunk.

For a logical chunk mapped to a conventional zone, all write operations
are processed by directly writing to the zone. If the mapping zone is a
sequential zone, the write operation is processed directly only if the
write offset within the logical chunk is equal to the write pointer
offset within of the sequential data zone (i.e. the write operation is
aligned on the zone write pointer). Otherwise, write operations are
processed indirectly using a buffer zone. In that case, an unused
conventional zone is allocated and assigned to the chunk being
accessed. Writing a block to the buffer zone of a chunk will
automatically invalidate the same block in the sequential zone mapping
the chunk. If all blocks of the sequential zone become invalid, the zone
is freed and the chunk buffer zone becomes the primary zone mapping the
chunk, resulting in native random write performance similar to a regular
block device.

Read operations are processed according to the block validity
information provided by the bitmaps. Valid blocks are read either from
the sequential zone mapping a chunk, or if the chunk is buffered, from
the buffer zone assigned. If the accessed chunk has no mapping, or the
accessed blocks are invalid, the read buffer is zeroed and the read
operation terminated.

After some time, the limited number of convnetional zones available may
be exhausted (all used to map chunks or buffer sequential zones) and
unaligned writes to unbuffered chunks become impossible. To avoid this
situation, a reclaim process regularly scans used conventional zones and
tries to reclaim the least recently used zones by copying the valid
blocks of the buffer zone to a free sequential zone. Once the copy
completes, the chunk mapping is updated to point to the sequential zone
and the buffer zone freed for reuse.

Metadata Protection
===================

To protect metadata against corruption in case of sudden power loss or
system crash, 2 sets of metadata zones are used. One set, the primary
set, is used as the main metadata region, while the secondary set is
used as a staging area. Modified metadata is first written to the
secondary set and validated by updating the super block in the secondary
set, a generation counter is used to indicate that this set contains the
newest metadata. Once this operation completes, in place of metadata
block updates can be done in the primary metadata set. This ensures that
one of the set is always consistent (all modifications committed or none
at all). Flush operations are used as a commit point. Upon reception of
a flush request, metadata modification activity is temporarily blocked
(for both incoming BIO processing and reclaim process) and all dirty
metadata blocks are staged and updated. Normal operation is then
resumed. Flushing metadata thus only temporarily delays write and
discard requests. Read requests can be processed concurrently while
metadata flush is being executed.

If a regular device is used in conjunction with the zoned block device,
a third set of metadata (without the zone bitmaps) is written to the
start of the zoned block device. This metadata has a generation counter of
'0' and will never be updated during normal operation; it just serves for
identification purposes. The first and second copy of the metadata
are located at the start of the regular block device.

Usage
=====

A zoned block device must first be formatted using the dmzadm tool. This
will analyze the device zone configuration, determine where to place the
metadata sets on the device and initialize the metadata sets.

Ex::

	dmzadm --format /dev/sdxx


If two drives are to be used, both devices must be specified, with the
regular block device as the first device.

Ex::

	dmzadm --format /dev/sdxx /dev/sdyy


Fomatted device(s) can be started with the dmzadm utility, too.:

Ex::

	dmzadm --start /dev/sdxx /dev/sdyy


Information about the internal layout and current usage of the zones can
be obtained with the 'status' callback from dmsetup:

Ex::

	dmsetup status /dev/dm-X

will return a line

	0 <size> zoned <nr_zones> zones <nr_unmap_rnd>/<nr_rnd> random <nr_unmap_seq>/<nr_seq> sequential

where <nr_zones> is the total number of zones, <nr_unmap_rnd> is the number
of unmapped (ie free) random zones, <nr_rnd> the total number of zones,
<nr_unmap_seq> the number of unmapped sequential zones, and <nr_seq> the
total number of sequential zones.

Normally the reclaim process will be started once there are less than 50
percent free random zones. In order to start the reclaim process manually
even before reaching this threshold the 'dmsetup message' function can be
used:

Ex::

	dmsetup message /dev/dm-X 0 reclaim

will start the reclaim process and random zones will be moved to sequential
zones.