338 lines · plain
1=====2Cache3=====4 5Introduction6============7 8dm-cache is a device mapper target written by Joe Thornber, Heinz9Mauelshagen, and Mike Snitzer.10 11It aims to improve performance of a block device (eg, a spindle) by12dynamically migrating some of its data to a faster, smaller device13(eg, an SSD).14 15This device-mapper solution allows us to insert this caching at16different levels of the dm stack, for instance above the data device for17a thin-provisioning pool. Caching solutions that are integrated more18closely with the virtual memory system should give better performance.19 20The target reuses the metadata library used in the thin-provisioning21library.22 23The decision as to what data to migrate and when is left to a plug-in24policy module. Several of these have been written as we experiment,25and we hope other people will contribute others for specific io26scenarios (eg. a vm image server).27 28Glossary29========30 31 Migration32 Movement of the primary copy of a logical block from one33 device to the other.34 Promotion35 Migration from slow device to fast device.36 Demotion37 Migration from fast device to slow device.38 39The origin device always contains a copy of the logical block, which40may be out of date or kept in sync with the copy on the cache device41(depending on policy).42 43Design44======45 46Sub-devices47-----------48 49The target is constructed by passing three devices to it (along with50other parameters detailed later):51 521. An origin device - the big, slow one.53 542. A cache device - the small, fast one.55 563. A small metadata device - records which blocks are in the cache,57 which are dirty, and extra hints for use by the policy object.58 This information could be put on the cache device, but having it59 separate allows the volume manager to configure it differently,60 e.g. as a mirror for extra robustness. This metadata device may only61 be used by a single cache device.62 63Fixed block size64----------------65 66The origin is divided up into blocks of a fixed size. This block size67is configurable when you first create the cache. Typically we've been68using block sizes of 256KB - 1024KB. The block size must be between 6469sectors (32KB) and 2097152 sectors (1GB) and a multiple of 64 sectors (32KB).70 71Having a fixed block size simplifies the target a lot. But it is72something of a compromise. For instance, a small part of a block may be73getting hit a lot, yet the whole block will be promoted to the cache.74So large block sizes are bad because they waste cache space. And small75block sizes are bad because they increase the amount of metadata (both76in core and on disk).77 78Cache operating modes79---------------------80 81The cache has three operating modes: writeback, writethrough and82passthrough.83 84If writeback, the default, is selected then a write to a block that is85cached will go only to the cache and the block will be marked dirty in86the metadata.87 88If writethrough is selected then a write to a cached block will not89complete until it has hit both the origin and cache devices. Clean90blocks should remain clean.91 92If passthrough is selected, useful when the cache contents are not known93to be coherent with the origin device, then all reads are served from94the origin device (all reads miss the cache) and all writes are95forwarded to the origin device; additionally, write hits cause cache96block invalidates. To enable passthrough mode the cache must be clean.97Passthrough mode allows a cache device to be activated without having to98worry about coherency. Coherency that exists is maintained, although99the cache will gradually cool as writes take place. If the coherency of100the cache can later be verified, or established through use of the101"invalidate_cblocks" message, the cache device can be transitioned to102writethrough or writeback mode while still warm. Otherwise, the cache103contents can be discarded prior to transitioning to the desired104operating mode.105 106A simple cleaner policy is provided, which will clean (write back) all107dirty blocks in a cache. Useful for decommissioning a cache or when108shrinking a cache. Shrinking the cache's fast device requires all cache109blocks, in the area of the cache being removed, to be clean. If the110area being removed from the cache still contains dirty blocks the resize111will fail. Care must be taken to never reduce the volume used for the112cache's fast device until the cache is clean. This is of particular113importance if writeback mode is used. Writethrough and passthrough114modes already maintain a clean cache. Future support to partially clean115the cache, above a specified threshold, will allow for keeping the cache116warm and in writeback mode during resize.117 118Migration throttling119--------------------120 121Migrating data between the origin and cache device uses bandwidth.122The user can set a throttle to prevent more than a certain amount of123migration occurring at any one time. Currently we're not taking any124account of normal io traffic going to the devices. More work needs125doing here to avoid migrating during those peak io moments.126 127For the time being, a message "migration_threshold <#sectors>"128can be used to set the maximum number of sectors being migrated,129the default being 2048 sectors (1MB).130 131Updating on-disk metadata132-------------------------133 134On-disk metadata is committed every time a FLUSH or FUA bio is written.135If no such requests are made then commits will occur every second. This136means the cache behaves like a physical disk that has a volatile write137cache. If power is lost you may lose some recent writes. The metadata138should always be consistent in spite of any crash.139 140The 'dirty' state for a cache block changes far too frequently for us141to keep updating it on the fly. So we treat it as a hint. In normal142operation it will be written when the dm device is suspended. If the143system crashes all cache blocks will be assumed dirty when restarted.144 145Per-block policy hints146----------------------147 148Policy plug-ins can store a chunk of data per cache block. It's up to149the policy how big this chunk is, but it should be kept small. Like the150dirty flags this data is lost if there's a crash so a safe fallback151value should always be possible.152 153Policy hints affect performance, not correctness.154 155Policy messaging156----------------157 158Policies will have different tunables, specific to each one, so we159need a generic way of getting and setting these. Device-mapper160messages are used. Refer to cache-policies.txt.161 162Discard bitset resolution163-------------------------164 165We can avoid copying data during migration if we know the block has166been discarded. A prime example of this is when mkfs discards the167whole block device. We store a bitset tracking the discard state of168blocks. However, we allow this bitset to have a different block size169from the cache blocks. This is because we need to track the discard170state for all of the origin device (compare with the dirty bitset171which is just for the smaller cache device).172 173Target interface174================175 176Constructor177-----------178 179 ::180 181 cache <metadata dev> <cache dev> <origin dev> <block size>182 <#feature args> [<feature arg>]*183 <policy> <#policy args> [policy args]*184 185 ================ =======================================================186 metadata dev fast device holding the persistent metadata187 cache dev fast device holding cached data blocks188 origin dev slow device holding original data blocks189 block size cache unit size in sectors190 191 #feature args number of feature arguments passed192 feature args writethrough or passthrough (The default is writeback.)193 194 policy the replacement policy to use195 #policy args an even number of arguments corresponding to196 key/value pairs passed to the policy197 policy args key/value pairs passed to the policy198 E.g. 'sequential_threshold 1024'199 See cache-policies.txt for details.200 ================ =======================================================201 202Optional feature arguments are:203 204 205 ==================== ========================================================206 writethrough write through caching that prohibits cache block207 content from being different from origin block content.208 Without this argument, the default behaviour is to write209 back cache block contents later for performance reasons,210 so they may differ from the corresponding origin blocks.211 212 passthrough a degraded mode useful for various cache coherency213 situations (e.g., rolling back snapshots of214 underlying storage). Reads and writes always go to215 the origin. If a write goes to a cached origin216 block, then the cache block is invalidated.217 To enable passthrough mode the cache must be clean.218 219 metadata2 use version 2 of the metadata. This stores the dirty220 bits in a separate btree, which improves speed of221 shutting down the cache.222 223 no_discard_passdown disable passing down discards from the cache224 to the origin's data device.225 ==================== ========================================================226 227A policy called 'default' is always registered. This is an alias for228the policy we currently think is giving best all round performance.229 230As the default policy could vary between kernels, if you are relying on231the characteristics of a specific policy, always request it by name.232 233Status234------235 236::237 238 <metadata block size> <#used metadata blocks>/<#total metadata blocks>239 <cache block size> <#used cache blocks>/<#total cache blocks>240 <#read hits> <#read misses> <#write hits> <#write misses>241 <#demotions> <#promotions> <#dirty> <#features> <features>*242 <#core args> <core args>* <policy name> <#policy args> <policy args>*243 <cache metadata mode>244 245 246========================= =====================================================247metadata block size Fixed block size for each metadata block in248 sectors249#used metadata blocks Number of metadata blocks used250#total metadata blocks Total number of metadata blocks251cache block size Configurable block size for the cache device252 in sectors253#used cache blocks Number of blocks resident in the cache254#total cache blocks Total number of cache blocks255#read hits Number of times a READ bio has been mapped256 to the cache257#read misses Number of times a READ bio has been mapped258 to the origin259#write hits Number of times a WRITE bio has been mapped260 to the cache261#write misses Number of times a WRITE bio has been262 mapped to the origin263#demotions Number of times a block has been removed264 from the cache265#promotions Number of times a block has been moved to266 the cache267#dirty Number of blocks in the cache that differ268 from the origin269#feature args Number of feature args to follow270feature args 'writethrough' (optional)271#core args Number of core arguments (must be even)272core args Key/value pairs for tuning the core273 e.g. migration_threshold274policy name Name of the policy275#policy args Number of policy arguments to follow (must be even)276policy args Key/value pairs e.g. sequential_threshold277cache metadata mode ro if read-only, rw if read-write278 279 In serious cases where even a read-only mode is280 deemed unsafe no further I/O will be permitted and281 the status will just contain the string 'Fail'.282 The userspace recovery tools should then be used.283needs_check 'needs_check' if set, '-' if not set284 A metadata operation has failed, resulting in the285 needs_check flag being set in the metadata's286 superblock. The metadata device must be287 deactivated and checked/repaired before the288 cache can be made fully operational again.289 '-' indicates needs_check is not set.290========================= =====================================================291 292Messages293--------294 295Policies will have different tunables, specific to each one, so we296need a generic way of getting and setting these. Device-mapper297messages are used. (A sysfs interface would also be possible.)298 299The message format is::300 301 <key> <value>302 303E.g.::304 305 dmsetup message my_cache 0 sequential_threshold 1024306 307 308Invalidation is removing an entry from the cache without writing it309back. Cache blocks can be invalidated via the invalidate_cblocks310message, which takes an arbitrary number of cblock ranges. Each cblock311range's end value is "one past the end", meaning 5-10 expresses a range312of values from 5 to 9. Each cblock must be expressed as a decimal313value, in the future a variant message that takes cblock ranges314expressed in hexadecimal may be needed to better support efficient315invalidation of larger caches. The cache must be in passthrough mode316when invalidate_cblocks is used::317 318 invalidate_cblocks [<cblock>|<cblock begin>-<cblock end>]*319 320E.g.::321 322 dmsetup message my_cache 0 invalidate_cblocks 2345 3456-4567 5678-6789323 324Examples325========326 327The test suite can be found here:328 329https://github.com/jthornber/device-mapper-test-suite330 331::332 333 dmsetup create my_cache --table '0 41943040 cache /dev/mapper/metadata \334 /dev/mapper/ssd /dev/mapper/origin 512 1 writeback default 0'335 dmsetup create my_cache --table '0 41943040 cache /dev/mapper/metadata \336 /dev/mapper/ssd /dev/mapper/origin 1024 1 writeback \337 mq 4 sequential_threshold 1024 random_threshold 8'338