Say you've got a big slow raid 6, and an X-25E or three. Wouldn't it be
nice if you could use them as cache... Hence bcache.

Wiki and git repositories are at:
  http://bcache.evilpiepirate.org
  http://evilpiepirate.org/git/linux-bcache.git
  http://evilpiepirate.org/git/bcache-tools.git

It's designed around the performance characteristics of SSDs - it only allocates
in erase block sized buckets, and it uses a hybrid btree/log to track cached
extants (which can be anywhere from a single sector to the bucket size). It's
designed to avoid random writes at all costs; it fills up an erase block
sequentially, then issues a discard before reusing it.

Both writethrough and writeback caching are supported. Writeback defaults to
off, but can be switched on and off arbitrarily at runtime. Bcache goes to
great lengths to protect your data - it reliably handles unclean shutdown. (It
doesn't even have a notion of a clean shutdown; bcache simply doesn't return
writes as completed until they're on stable storage).

Writeback caching can use most of the cache for buffering writes - writing
dirty data to the backing device is always done sequentially, scanning from the
start to the end of the index.

Since random IO is what SSDs excel at, there generally won't be much benefit
to caching large sequential IO. Bcache detects sequential IO and skips it;
it also keeps a rolling average of the IO sizes per task, and as long as the
average is above the cutoff it will skip all IO from that task - instead of
caching the first 512k after every seek. Backups and large file copies should
thus entirely bypass the cache.

In the event of a data IO error on the flash it will try to recover by reading
from disk or invalidating cache entries.  For unrecoverable errors (meta data
or dirty data), caching is automatically disabled; if dirty data was present
in the cache it first disables writeback caching and waits for all dirty data
to be flushed.

Getting started:
You'll need make-bcache from the bcache-tools repository. Both the cache device
and backing device must be formatted before use.
  make-bcache -B /dev/sdb
  make-bcache -C /dev/sdc

make-bcache has the ability to format multiple devices at the same time - if
you format your backing devices and cache device at the same time, you won't
have to manually attach:
  make-bcache -B /dev/sda /dev/sdb -C /dev/sdc

To make bcache devices known to the kernel, echo them to /sys/fs/bcache/register:

  echo /dev/sdb > /sys/fs/bcache/register
  echo /dev/sdc > /sys/fs/bcache/register

To register your bcache devices automatically, you could add something like
this to an init script:

  echo /dev/sd* > /sys/fs/bcache/register_quiet

It'll look for bcache superblocks and ignore everything that doesn't have one.

Registering the backing device makes the bcache show up in /dev; you can now
format it and use it as normal. But the first time using a new bcache device,
it'll be running in passthrough mode until you attach it to a cache. See the
section on attaching.

The devices show up at /dev/bcacheN, and can be controlled via sysfs from
/sys/block/bcacheN/bcache:

  mkfs.ext4 /dev/bcache0
  mount /dev/bcache0 /mnt

Cache devices are managed as sets; multiple caches per set isn't supported yet
but will allow for mirroring of metadata and dirty data in the future. Your new
cache set shows up as /sys/fs/bcache/<UUID>

ATTACHING:

After your cache device and backing device are registered, the backing device
must be attached to your cache set to enable caching. Attaching a backing
device to a cache set is done thusly, with the UUID of the cache set in
/sys/fs/bcache:

  echo <UUID> > /sys/block/bcache0/bcache/attach

This only has to be done once. The next time you reboot, just reregister all
your bcache devices. If a backing device has data in a cache somewhere, the
/dev/bcache# device won't be created until the cache shows up - particularly
important if you have writeback caching turned on.

If you're booting up and your cache device is gone and never coming back, you
can force run the backing device:

  echo 1 > /sys/block/sdb/bcache/running

(You need to use /sys/block/sdb (or whatever your backing device is called), not
/sys/block/bcache0, because bcache0 doesn't exist yet. If you're using a
partition, the bcache directory would be at /sys/block/sdb/sdb2/bcache)

The backing device will still use that cache set if it shows up in the future,
but all the cached data will be invalidated. If there was dirty data in the
cache, don't expect the filesystem to be recoverable - you will have massive
filesystem corruption, though ext4's fsck does work miracles.

ERROR HANDLING:

Bcache tries to transparently handle IO errors to/from the cache device without
affecting normal operation; if it sees too many errors (the threshold is
configurable, and defaults to 0) it shuts down the cache device and switches all
the backing devices to passthrough mode.

 - For reads from the cache, if they error we just retry the read from the
   backing device.

 - For writethrough writes, if the write to the cache errors we just switch to
   invalidating the data at that lba in the cache (i.e. the same thing we do for
   a write that bypasses the cache)

 - For writeback writes, we currently pass that error back up to the
   filesystem/userspace. This could be improved - we could retry it as a write
   that skips the cache so we don't have to error the write.

 - When we detach, we first try to flush any dirty data (if we were running in
   writeback mode). It currently doesn't do anything intelligent if it fails to
   read some of the dirty data, though.

TROUBLESHOOTING PERFORMANCE:

Bcache has a bunch of config options and tunables. The defaults are intended to
be reasonable for typical desktop and server workloads, but they're not what you
want for getting the best possible numbers when benchmarking.

 - Bad write performance

   If write performance is not what you expected, you probably wanted to be
   running in writeback mode, which isn't the default (not due to a lack of
   maturity, but simply because in writeback mode you'll lose data if something
   happens to your SSD)

   # echo writeback > /sys/block/bcache0/cache_mode

 - Bad performance, or traffic not going to the SSD that you'd expect

   By default, bcache doesn't cache everything. It tries to skip sequential IO -
   because you really want to be caching the random IO, and if you copy a 10
   gigabyte file you probably don't want that pushing 10 gigabytes of randomly
   accessed data out of your cache.

   But if you want to benchmark reads from cache, and you start out with fio
   writing an 8 gigabyte test file - so you want to disable that.

   # echo 0 > /sys/block/bcache0/bcache/sequential_cutoff

   To set it back to the default (4 mb), do

   # echo 4M > /sys/block/bcache0/bcache/sequential_cutoff

 - Traffic's still going to the spindle/still getting cache misses

   In the real world, SSDs don't always keep up with disks - particularly with
   slower SSDs, many disks being cached by one SSD, or mostly sequential IO. So
   you want to avoid being bottlenecked by the SSD and having it slow everything
   down.

   To avoid that bcache tracks latency to the cache device, and gradually
   throttles traffic if the latency exceeds a threshold (it does this by
   cranking down the sequential bypass).

   You can disable this if you need to by setting the thresholds to 0:

   # echo 0 > /sys/fs/bcache/<cache set>/congested_read_threshold_us
   # echo 0 > /sys/fs/bcache/<cache set>/congested_write_threshold_us

   The default is 2000 us (2 milliseconds) for reads, and 20000 for writes.

 - Still getting cache misses, of the same data

   One last issue that sometimes trips people up is actually an old bug, due to
   the way cache coherency is handled for cache misses. If a btree node is full,
   a cache miss won't be able to insert a key for the new data and the data
   won't be written to the cache.

   In practice this isn't an issue because as soon as a write comes along it'll
   cause the btree node to be split, and you need almost no write traffic for
   this to not show up enough to be noticable (especially since bcache's btree
   nodes are huge and index large regions of the device). But when you're
   benchmarking, if you're trying to warm the cache by reading a bunch of data
   and there's no other traffic - that can be a problem.

   Solution: warm the cache by doing writes, or use the testing branch (there's
   a fix for the issue there).

SYSFS - BACKING DEVICE:

attach
  Echo the UUID of a cache set to this file to enable caching.

cache_mode
  Can be one of either writethrough, writeback, writearound or none.

clear_stats
  Writing to this file resets the running total stats (not the day/hour/5 minute
  decaying versions).

detach
  Write to this file to detach from a cache set. If there is dirty data in the
  cache, it will be flushed first.

dirty_data
  Amount of dirty data for this backing device in the cache. Continuously
  updated unlike the cache set's version, but may be slightly off.

label
  Name of underlying device.

readahead
  Size of readahead that should be performed.  Defaults to 0.  If set to e.g.
  1M, it will round cache miss reads up to that size, but without overlapping
  existing cache entries.

running
  1 if bcache is running (i.e. whether the /dev/bcache device exists, whether
  it's in passthrough mode or caching).

sequential_cutoff
  A sequential IO will bypass the cache once it passes this threshhold; the
  most recent 128 IOs are tracked so sequential IO can be detected even when
  it isn't all done at once.

sequential_merge
  If non zero, bcache keeps a list of the last 128 requests submitted to compare
  against all new requests to determine which new requests are sequential
  continuations of previous requests for the purpose of determining sequential
  cutoff. This is necessary if the sequential cutoff value is greater than the
  maximum acceptable sequential size for any single request. 

state
  The backing device can be in one of four different states:

  no cache: Has never been attached to a cache set.

  clean: Part of a cache set, and there is no cached dirty data.

  dirty: Part of a cache set, and there is cached dirty data.

  inconsistent: The backing device was forcibly run by the user when there was
  dirty data cached but the cache set was unavailable; whatever data was on the
  backing device has likely been corrupted.

stop
  Write to this file to shut down the bcache device and close the backing
  device.

writeback_delay
  When dirty data is written to the cache and it previously did not contain
  any, waits some number of seconds before initiating writeback. Defaults to
  30.

writeback_percent
  If nonzero, bcache tries to keep around this percentage of the cache dirty by
  throttling background writeback and using a PD controller to smoothly adjust
  the rate.

writeback_rate
  Rate in sectors per second - if writeback_percent is nonzero, background
  writeback is throttled to this rate. Continuously adjusted by bcache but may
  also be set by the user.

writeback_running
  If off, writeback of dirty data will not take place at all. Dirty data will
  still be added to the cache until it is mostly full; only meant for
  benchmarking. Defaults to on.

SYSFS - BACKING DEVICE STATS:

There are directories with these numbers for a running total, as well as
versions that decay over the past day, hour and 5 minutes; they're also
aggregated in the cache set directory as well.

bypassed
  Amount of IO (both reads and writes) that has bypassed the cache

cache_hits
cache_misses
cache_hit_ratio
  Hits and misses are counted per individual IO as bcache sees them; a
  partial hit is counted as a miss.

cache_bypass_hits
cache_bypass_misses
  Hits and misses for IO that is intended to skip the cache are still counted,
  but broken out here.

cache_miss_collisions
  Counts instances where data was going to be inserted into the cache from a
  cache miss, but raced with a write and data was already present (usually 0
  since the synchronization for cache misses was rewritten)

cache_readaheads
  Count of times readahead occured.

SYSFS - CACHE SET:

average_key_size
  Average data per key in the btree.

bdev<0..n>
  Symlink to each of the attached backing devices.

block_size
  Block size of the cache devices.

btree_cache_size
  Amount of memory currently used by the btree cache

bucket_size
  Size of buckets

cache<0..n>
  Symlink to each of the cache devices comprising this cache set. 

cache_available_percent
  Percentage of cache device which doesn't contain dirty data, and could
  potentially be used for writeback.  This doesn't mean this space isn't used
  for clean cached data; the unused statistic (in priority_stats) is typically
  much lower.

clear_stats
  Clears the statistics associated with this cache

dirty_data
  Amount of dirty data is in the cache (updated when garbage collection runs).

flash_vol_create
  Echoing a size to this file (in human readable units, k/M/G) creates a thinly
  provisioned volume backed by the cache set.

io_error_halflife
io_error_limit
  These determines how many errors we accept before disabling the cache.
  Each error is decayed by the half life (in # ios).  If the decaying count
  reaches io_error_limit dirty data is written out and the cache is disabled.

journal_delay_ms
  Journal writes will delay for up to this many milliseconds, unless a cache
  flush happens sooner. Defaults to 100.

root_usage_percent
  Percentage of the root btree node in use.  If this gets too high the node
  will split, increasing the tree depth.

stop
  Write to this file to shut down the cache set - waits until all attached
  backing devices have been shut down.

tree_depth
  Depth of the btree (A single node btree has depth 0).

unregister
  Detaches all backing devices and closes the cache devices; if dirty data is
  present it will disable writeback caching and wait for it to be flushed.

SYSFS - CACHE SET INTERNAL:

This directory also exposes timings for a number of internal operations, with
separate files for average duration, average frequency, last occurence and max
duration: garbage collection, btree read, btree node sorts and btree splits.

active_journal_entries
  Number of journal entries that are newer than the index.

btree_nodes
  Total nodes in the btree.

btree_used_percent
  Average fraction of btree in use.

bset_tree_stats
  Statistics about the auxiliary search trees

btree_cache_max_chain
  Longest chain in the btree node cache's hash table

cache_read_races
  Counts instances where while data was being read from the cache, the bucket
  was reused and invalidated - i.e. where the pointer was stale after the read
  completed. When this occurs the data is reread from the backing device.

trigger_gc
  Writing to this file forces garbage collection to run.

SYSFS - CACHE DEVICE:

block_size
  Minimum granularity of writes - should match hardware sector size.

btree_written
  Sum of all btree writes, in (kilo/mega/giga) bytes

bucket_size
  Size of buckets

cache_replacement_policy
  One of either lru, fifo or random.

discard
  Boolean; if on a discard/TRIM will be issued to each bucket before it is
  reused. Defaults to off, since SATA TRIM is an unqueued command (and thus
  slow).

freelist_percent
  Size of the freelist as a percentage of nbuckets. Can be written to to
  increase the number of buckets kept on the freelist, which lets you
  artificially reduce the size of the cache at runtime. Mostly for testing
  purposes (i.e. testing how different size caches affect your hit rate), but
  since buckets are discarded when they move on to the freelist will also make
  the SSD's garbage collection easier by effectively giving it more reserved
  space.

io_errors
  Number of errors that have occured, decayed by io_error_halflife.

metadata_written
  Sum of all non data writes (btree writes and all other metadata).

nbuckets
  Total buckets in this cache

priority_stats
  Statistics about how recently data in the cache has been accessed.
  This can reveal your working set size.  Unused is the percentage of
  the cache that doesn't contain any data.  Metadata is bcache's
  metadata overhead.  Average is the average priority of cache buckets.
  Next is a list of quantiles with the priority threshold of each.

written
  Sum of all data that has been written to the cache; comparison with
  btree_written gives the amount of write inflation in bcache.