4 dm-cache is a device mapper target written by Joe Thornber, Heinz
5 Mauelshagen, and Mike Snitzer.
7 It aims to improve performance of a block device (eg, a spindle) by
8 dynamically migrating some of its data to a faster, smaller device
11 This device-mapper solution allows us to insert this caching at
12 different levels of the dm stack, for instance above the data device for
13 a thin-provisioning pool. Caching solutions that are integrated more
14 closely with the virtual memory system should give better performance.
16 The target reuses the metadata library used in the thin-provisioning
19 The decision as to what data to migrate and when is left to a plug-in
20 policy module. Several of these have been written as we experiment,
21 and we hope other people will contribute others for specific io
22 scenarios (eg. a vm image server).
27 Migration - Movement of the primary copy of a logical block from one
29 Promotion - Migration from slow device to fast device.
30 Demotion - Migration from fast device to slow device.
32 The origin device always contains a copy of the logical block, which
33 may be out of date or kept in sync with the copy on the cache device
34 (depending on policy).
42 The target is constructed by passing three devices to it (along with
43 other parameters detailed later):
45 1. An origin device - the big, slow one.
47 2. A cache device - the small, fast one.
49 3. A small metadata device - records which blocks are in the cache,
50 which are dirty, and extra hints for use by the policy object.
51 This information could be put on the cache device, but having it
52 separate allows the volume manager to configure it differently,
53 e.g. as a mirror for extra robustness. This metadata device may only
54 be used by a single cache device.
59 The origin is divided up into blocks of a fixed size. This block size
60 is configurable when you first create the cache. Typically we've been
61 using block sizes of 256KB - 1024KB. The block size must be between 64
62 (32KB) and 2097152 (1GB) and a multiple of 64 (32KB).
64 Having a fixed block size simplifies the target a lot. But it is
65 something of a compromise. For instance, a small part of a block may be
66 getting hit a lot, yet the whole block will be promoted to the cache.
67 So large block sizes are bad because they waste cache space. And small
68 block sizes are bad because they increase the amount of metadata (both
74 The cache has three operating modes: writeback, writethrough and
77 If writeback, the default, is selected then a write to a block that is
78 cached will go only to the cache and the block will be marked dirty in
81 If writethrough is selected then a write to a cached block will not
82 complete until it has hit both the origin and cache devices. Clean
83 blocks should remain clean.
85 If passthrough is selected, useful when the cache contents are not known
86 to be coherent with the origin device, then all reads are served from
87 the origin device (all reads miss the cache) and all writes are
88 forwarded to the origin device; additionally, write hits cause cache
89 block invalidates. To enable passthrough mode the cache must be clean.
90 Passthrough mode allows a cache device to be activated without having to
91 worry about coherency. Coherency that exists is maintained, although
92 the cache will gradually cool as writes take place. If the coherency of
93 the cache can later be verified, or established through use of the
94 "invalidate_cblocks" message, the cache device can be transitioned to
95 writethrough or writeback mode while still warm. Otherwise, the cache
96 contents can be discarded prior to transitioning to the desired
99 A simple cleaner policy is provided, which will clean (write back) all
100 dirty blocks in a cache. Useful for decommissioning a cache or when
101 shrinking a cache. Shrinking the cache's fast device requires all cache
102 blocks, in the area of the cache being removed, to be clean. If the
103 area being removed from the cache still contains dirty blocks the resize
104 will fail. Care must be taken to never reduce the volume used for the
105 cache's fast device until the cache is clean. This is of particular
106 importance if writeback mode is used. Writethrough and passthrough
107 modes already maintain a clean cache. Future support to partially clean
108 the cache, above a specified threshold, will allow for keeping the cache
109 warm and in writeback mode during resize.
114 Migrating data between the origin and cache device uses bandwidth.
115 The user can set a throttle to prevent more than a certain amount of
116 migration occurring at any one time. Currently we're not taking any
117 account of normal io traffic going to the devices. More work needs
118 doing here to avoid migrating during those peak io moments.
120 For the time being, a message "migration_threshold <#sectors>"
121 can be used to set the maximum number of sectors being migrated,
122 the default being 204800 sectors (or 100MB).
124 Updating on-disk metadata
125 -------------------------
127 On-disk metadata is committed every time a REQ_SYNC or REQ_FUA bio is
128 written. If no such requests are made then commits will occur every
129 second. This means the cache behaves like a physical disk that has a
130 write cache (the same is true of the thin-provisioning target). If
131 power is lost you may lose some recent writes. The metadata should
132 always be consistent in spite of any crash.
134 The 'dirty' state for a cache block changes far too frequently for us
135 to keep updating it on the fly. So we treat it as a hint. In normal
136 operation it will be written when the dm device is suspended. If the
137 system crashes all cache blocks will be assumed dirty when restarted.
139 Per-block policy hints
140 ----------------------
142 Policy plug-ins can store a chunk of data per cache block. It's up to
143 the policy how big this chunk is, but it should be kept small. Like the
144 dirty flags this data is lost if there's a crash so a safe fallback
145 value should always be possible.
147 For instance, the 'mq' policy, which is currently the default policy,
148 uses this facility to store the hit count of the cache blocks. If
149 there's a crash this information will be lost, which means the cache
150 may be less efficient until those hit counts are regenerated.
152 Policy hints affect performance, not correctness.
157 Policies will have different tunables, specific to each one, so we
158 need a generic way of getting and setting these. Device-mapper
159 messages are used. Refer to cache-policies.txt.
161 Discard bitset resolution
162 -------------------------
164 We can avoid copying data during migration if we know the block has
165 been discarded. A prime example of this is when mkfs discards the
166 whole block device. We store a bitset tracking the discard state of
167 blocks. However, we allow this bitset to have a different block size
168 from the cache blocks. This is because we need to track the discard
169 state for all of the origin device (compare with the dirty bitset
170 which is just for the smaller cache device).
178 cache <metadata dev> <cache dev> <origin dev> <block size>
179 <#feature args> [<feature arg>]*
180 <policy> <#policy args> [policy args]*
182 metadata dev : fast device holding the persistent metadata
183 cache dev : fast device holding cached data blocks
184 origin dev : slow device holding original data blocks
185 block size : cache unit size in sectors
187 #feature args : number of feature arguments passed
188 feature args : writethrough or passthrough (The default is writeback.)
190 policy : the replacement policy to use
191 #policy args : an even number of arguments corresponding to
192 key/value pairs passed to the policy
193 policy args : key/value pairs passed to the policy
194 E.g. 'sequential_threshold 1024'
195 See cache-policies.txt for details.
197 Optional feature arguments are:
198 writethrough : write through caching that prohibits cache block
199 content from being different from origin block content.
200 Without this argument, the default behaviour is to write
201 back cache block contents later for performance reasons,
202 so they may differ from the corresponding origin blocks.
204 passthrough : a degraded mode useful for various cache coherency
205 situations (e.g., rolling back snapshots of
206 underlying storage). Reads and writes always go to
207 the origin. If a write goes to a cached origin
208 block, then the cache block is invalidated.
209 To enable passthrough mode the cache must be clean.
211 A policy called 'default' is always registered. This is an alias for
212 the policy we currently think is giving best all round performance.
214 As the default policy could vary between kernels, if you are relying on
215 the characteristics of a specific policy, always request it by name.
220 <metadata block size> <#used metadata blocks>/<#total metadata blocks>
221 <cache block size> <#used cache blocks>/<#total cache blocks>
222 <#read hits> <#read misses> <#write hits> <#write misses>
223 <#demotions> <#promotions> <#dirty> <#features> <features>*
224 <#core args> <core args>* <policy name> <#policy args> <policy args>*
226 metadata block size : Fixed block size for each metadata block in
228 #used metadata blocks : Number of metadata blocks used
229 #total metadata blocks : Total number of metadata blocks
230 cache block size : Configurable block size for the cache device
232 #used cache blocks : Number of blocks resident in the cache
233 #total cache blocks : Total number of cache blocks
234 #read hits : Number of times a READ bio has been mapped
236 #read misses : Number of times a READ bio has been mapped
238 #write hits : Number of times a WRITE bio has been mapped
240 #write misses : Number of times a WRITE bio has been
242 #demotions : Number of times a block has been removed
244 #promotions : Number of times a block has been moved to
246 #dirty : Number of blocks in the cache that differ
248 #feature args : Number of feature args to follow
249 feature args : 'writethrough' (optional)
250 #core args : Number of core arguments (must be even)
251 core args : Key/value pairs for tuning the core
252 e.g. migration_threshold
253 policy name : Name of the policy
254 #policy args : Number of policy arguments to follow (must be even)
255 policy args : Key/value pairs
256 e.g. sequential_threshold
261 Policies will have different tunables, specific to each one, so we
262 need a generic way of getting and setting these. Device-mapper
263 messages are used. (A sysfs interface would also be possible.)
265 The message format is:
270 dmsetup message my_cache 0 sequential_threshold 1024
273 Invalidation is removing an entry from the cache without writing it
274 back. Cache blocks can be invalidated via the invalidate_cblocks
275 message, which takes an arbitrary number of cblock ranges. Each cblock
276 range's end value is "one past the end", meaning 5-10 expresses a range
277 of values from 5 to 9. Each cblock must be expressed as a decimal
278 value, in the future a variant message that takes cblock ranges
279 expressed in hexidecimal may be needed to better support efficient
280 invalidation of larger caches. The cache must be in passthrough mode
281 when invalidate_cblocks is used.
283 invalidate_cblocks [<cblock>|<cblock begin>-<cblock end>]*
286 dmsetup message my_cache 0 invalidate_cblocks 2345 3456-4567 5678-6789
291 The test suite can be found here:
293 https://github.com/jthornber/device-mapper-test-suite
295 dmsetup create my_cache --table '0 41943040 cache /dev/mapper/metadata \
296 /dev/mapper/ssd /dev/mapper/origin 512 1 writeback default 0'
297 dmsetup create my_cache --table '0 41943040 cache /dev/mapper/metadata \
298 /dev/mapper/ssd /dev/mapper/origin 1024 1 writeback \
299 mq 4 sequential_threshold 1024 random_threshold 8'