Donation from Maigurs Stalidzans. Thanks!

[ikiwiki.git] / docs / unknown / Filesystem.mdwn

Initial specification of the new DragonFly filesystem.\r
Currently there is no name for this, DFS would have been nice, but is already overloaded.\r
\r
## Feature Summary \r
\r

* On-demand filesystem check and recovery.  No need to scan the entire\r
  filesystem before going live after a reboot.\r
\r

* Infinite snapshots (e.g. on the 30 second filesystem sync), ability\r
  to collapse snapshots for any time interval as a means of\r
  recovering space.\r
\r

* Multi-master operation, including the ability to self-heal a\r
  corrupted filesystem by accessing replicated data.  Multi-master \r
  means that each replicated store can act as a master.  New\r
  filesystem ops can run independantly on any of the replicated stores\r
  and will flow to the others.  This is done by giving the controller\r
  of each store a certain amount of guarenteed free space that can be\r
  filled without having to notify other controllers.\r
\r

* Infinite logless Replication.  No requirement to keep a log of changes\r
  for the purposes of replication, meaning that replication targets can\r
  be offline for 'days' without effecting performance or operation.\r
  ('mobile' computing, but also replication over slow links for backup\r
  purposes and other things).  Certain information is still logged,\r
  but only to streamline performance.\r
\r

* 64 bit file space, 64 bit filesystem space.  No space restrictions\r
  whatsoever.\r
\r

* Reliably handles data storage for huge multi-hundred-terrabyte\r
  filesystems without fear of unrecoverable corruption.\r
\r

* Cluster operation - ability to commit data to locally replicated\r
  store independantly of other replication nodes, with access governed\r
  by cache coherency protocols.\r
\r

* Independant index.  Data is laid out in a highly recoverable fashion,\r
  independant of index generation.  Indexes can be regenerated from\r
  scratch and thus indexes can be updated asynchronously.\r
\r
## Segmentation \r
\r
The physical storage backing a filesystem is broken up into large\r
1MB-4GB segments (64MB is a typical value).  Each segment is\r
self-identifying and contains its own header, data table, and record\r
table.  The operating system glues together filesystems and determines\r
availability based on the segments it finds.\r
\r
All segments on a block device are the same size (power-of-2 restricted)\r
and the OS can probe the disk without any other information simply by\r
locating a segment header, starting with the largest possible segment\r
size.  Segment size is typically chosen such that a scan of all segment\r
headers takes no longer then a few seconds.\r
\r

* The segment header contains all the information required to associate\r
  the segment with a filesystem.   Certain information in the segment\r
  header is updated with an ordered write to 'commit' other work already\r
  flushed out into the segment.  The segment header also contains a \r
  bit indicating whether the segment is 'open' or not, to aid in \r
  on-demand recovery.\r
\r
  The segment ID is stored in the segment header, allowing segments\r
  to be easily relocated.  That is, the segment's location in the\r
  physical backing store is irrelevant.\r
\r

* The data table consists of pure data, laid out linearly in the forward\r
  direction within the segment.   Data blocks are variable-sized entities\r
  containing pure data, with no other identifying information, suitable\r
  for direct DMA.  The segment header has a simple append index for\r
  the data table.\r
\r

* The record table consists of fixed-sized records and a reference to\r
  data in the data table.  The record table is built backwards from\r
  the end of the segment.  The segment header has a simple pre-append\r
  index for the record table.  Each record consists of:\r
\r

* an object identifier (e.g. inode number)\r

* creation transid\r

* deletion transid (updated in-place)\r

* indexing key (offset or filename key)\r

* data table reference (offset, bytes)  (up to a 1GB contiguous ref)\r

* integrity check\r
\r
All data elements are completely identified by the record table.  All\r
modifications are historical in nature... that is, no actual data is\r
destroyed and one can access a 'snapshot' of the segment as of any\r
transaction id (i.e. timestamp) simply by ignoring those records\r
marked as deleted prior to the specified time or created after\r
the specified time.  \r
\r
Certain updates to the object table occur in-place.  In particular,\r
the deletion transaction id is updated in-place.  However, such\r
updates are not considered 'committed' until the segment header itself\r
is updated with the latest committed transaction id and a recovery\r
operation will undo any deletion transaction id greater then the\r
committed transaction id in the segment header, as well as free\r
any uncommitted objects and their related data.\r
\r
The entire filesystem can be recovered from just the record and data\r
tables.  Indexing and cross-segment spanning is described in later\r
sections.\r
\r
## Physical space recovery \r
\r
Physical space recovery requires actually destroying records marked\r
as deleted.  Snapshots can be collapsed by destroying records whos\r
creation AND deletion id's fall within the collapsed space.  The oldest\r
data is freed up by destroying records whos deletion id is less then\r
the terminal timestamp. \r
\r
Record destruction creates holes in both the data table and the record\r
table.  Any holes adjacent to the data table append point or the record\r
table prepend point are immediately recovered by adjusting the \r
appropriate indices in the segment header.  The operating system may\r
cache a record of non-adjacent holes (in memory) and reuse the space,\r
and can also generate an in-memory index of available holes on the\r
fly when space is very tight (which requires scanning the record table),\r
but otherwise the recovery of any space not adjacent to the data table\r
append point requires a performance reorganization of the segment.\r
\r
## Locating Data objects, and the Master Record \r
\r
Data objects are basically the amalgamation of all records with\r
the same 64 bit object identifier.  The top N bits of this identifier\r
indicate the segment the master record for the data object resides in.\r
All 64 bits are made available to userland.\r
\r
The filesystem needs to be free to relocate the master record for\r
a data object on the fly.  Relocation is a dynamic process whereby \r
the master record in a segment becomes a forwarding record to another\r
segment.  Any references to a forwarding record is adjusted on the fly\r
in order to collapse the chain, and any intermediate forwarding records\r
can be physically destroyed once all references to them have been\r
adjusted.  However, the ORIGINAL forwarding record must be retained\r
for all time in order to maintain the uniqueness of the originally\r
assigned user-visible inode number and to give us a way to locate\r
the master record given the inode number.  We cannot change the inode\r
number.  Overhead is minimal.\r
\r
A forwarding record is created in two situations:\r
1. To move the  master record to improve performance\r
2. If the current segment does not have sufficient space to increase the\r
size the master record if it becomes necessary to do so.\r
\r
A forwarding record is a degenerate case of a master record and the\r
forwarding information is embedded in the record itself, with no\r
data table reference at all.  The space used is only the space required\r
to store a record table entry.\r
\r
The master record for a data object is a special record in the record\r
table.  It is special because it is NOT part of the historical data\r
store.  The creation and deletion transaction ids represent the creation\r
or deletion of the entire data object, and the data block contains\r
information on where to find the other bits and pieces of the data\r
object (in particular, if the data object spans more then one segment).\r
***'A corrupted Master Record can be completely reconstructed by scanninG\r
the filesystem***'.\r
\r
The master record can be thought of as a persistent cache.  It gives\r
the filesystem a way to quickly locate all the segments that might\r
contain records for the data object and reserves a limited amount of\r
space for the filesystem to store direct references to data residing\r
in the same segment as the master record.\r
\r
Master record updates are fairly rare.  For the most part a master\r
record must only be updated if a file or directory grows into a\r
new segment.\r
\r
## Performance reorganization \r
\r
Segments can be repacked in order to clean up fragmentation and \r
reorganize data and records for more optimal access.  Repacking is\r
accomplished by copying the segment's data into a wholely new\r
segment on the physical disk, then destroying the old segment.\r
\r
Since a segment is identified by its header the actual location of\r
the segment on the physical media is irrelevant.\r
\r
For example, master records can wind up strewn about the record\r
table for a segment.  Repacking would pack all of those master\r
records next to each other.\r
\r
Similarly, a file or directory might have bits and pieces of data\r
strewn about a segment.  Repacking would de-fragment those entities,\r
as well as pack together the data related to small files and \r
separate out the larger chunks related to larger files.\r
\r
## Segment Allocation \r
\r
Remember that since the actual physical location of a segment is\r
independant of the segment identifier (typically an 8 or 16 bit\r
number), the allocation of segment numbers does not have to be\r
linear.  The filesystem will typically be generous in its allocation\r
of segment numbers in order to allow for spill over growth of files\r
into logically adjacent segments, thus simplifying the segment\r
range stored in the master record that describes which segments\r
might contain data records for a data object.  For example,\r
the first segment allocated by the filesystem when using a 16 bit\r
segment id would not be 0x0000 or 0xffff, but probably 0x8000.  The\r
next segment allocated (when not doing a spill-over) might be 0x4000\r
or 0xc000, and so forth.  The entire segment space is used even\r
if there are fewer actual (physical) segments.\r
\r
## Large cross-segment objects \r
\r
A Data object can wind up being far larger then a segment.  For that\r
matter, even a small data object with a lot of history can wind up being\r
far larger then a segment.\r
\r
The filesystem does its best to organize the data for such objects\r
to reduce the size of the master records required to describe them\r
and to separate historical data from current data, to maintain the\r
performance of the live filesystem.\r
\r
## Indexing \r
\r
An object's master record generally describes the segments containing\r
the data for the object, and may contain direct references to data\r
in the same segment as the master record (an optimization or performance \r
reorganization that is desireable for files smaller then a single\r
segment).\r
\r
However, data objects can grow to be many records due to fragmentation,\r
simply being large files, or due to the retention of a great deal of\r
history.  The records pertaining to these objects may have to be indexed\r
beyond the space the filesystem is willing to reserve in the master\r
record in order to improve access.\r
\r
To make it clear, indexing is an optimization.  The index is not\r
required to recover a segment or to recover a filesystem.  The intent\r
is for the master record to always contain sufficient information\r
to reduce the number I/O's required to resolve a file access request\r
to a reasonable number, even if no index is available.\r
\r
Indexing can occur in three places:\r
\r

* First, it can occur in the segment itself to organize the records\r
  in that segment.  Such indexes are usually persistently cached in the\r
  dead space between the data table and the record table, and the \r
  filesystem heuristically determines how much space is needed for\r
  that.  If the data table or the record table grows into the index\r
  the filesystem can choose to relocate, regenerate, or shift the index\r
  data, or to disallow the growth (extend the data object into a new\r
  segment).  These heuristics are fairly simple.\r
\r

* Second, it can occur in the master record to roughly organize the\r
  data space... for example so the filesystem does not have to check\r
  all segments if a particular file offset (or offset range) and all\r
  of its history is known to reside in just one.  The filesystem\r
  typically is only willing to allow so much space in the master record\r
  for a data object to store such an index.  If this level of index\r
  becomes too large it is basically simplified in-place and starts\r
  requiring the use of the per-segment index to further resolve the\r
  location of data records for the object.\r
\r

* Third, it can be generated and cached in memory.  When dealing with\r
  very large multi-segment files it may be beneficial to scan\r
  the relatively few records in the record table for the related segments\r
  and simply index them in memory, rather then store an index on-disk.\r
\r
  For example if you are using 64MB segments and have a 20GB file,\r
  literally only 320 disk accesses (with a few data records read in\r
  each access) are required to construct an index of the entire 20GB\r
  file and the index itself would require very little memory.\r
\r
Indexes are asynchronous entities.  The 'official' filesystem store is\r
the data table and the record table.  The host can cache updates in\r
memory, and asynchronously update the performance-improving index.  \r
\r
Indexes residing in segments are regenerated if the segment is marked\r
open on initial access (due to an unclean shutdown).  This allows\r
persistent per-segment indexes to be updated without requiring any\r
particular transactional disk synchronization or block ordering.\r
\r
Indexes are generally implemented using a B+Tree structure.\r
\r
## Replication \r
\r
Data and record elements are only directly referenced by their offset\r
within a segment when referenced from within that same segment.  This\r
means that replication is possible on a segment-by-segment basis.\r
\r
Segment headers contain asynchronously updated record update log areas\r
for deletion events (because these modify the deletion timestamp in\r
an existing record rather then append a new record).  These log areas\r
are of limited size and can overflow under normal operation.  They are\r
ONLY used to streamline replication.  If a replication target is not\r
fast enough, it will have to resynchronize by scanning the records\r
on the source (which itself doesn't usually take very long to do so it\r
isn't considered a big deal).  But otherwise it can check the log area\r
and simply pick up where it left off.  The intention is to completely\r
support both very slow replication slaves and mostly off-line slaves.\r
\r
The only thing that is actually replicated are the record table\r
entries and (with the exception of a master record) the data table\r
entries.  The data table entry for a master record is never replicated,\r
but (re)generated by the replication target.  The replication target\r
is responsible for maintaining its own indexes and managing its own\r
physical segment space.  This gives any replication target a great\r
deal of robustness and flexibility.\r
\r
Also note that the physical location of the logical segments on the\r
replication target is entirely up to the replication target.  Replication\r
is done logically, not physically.\r
\r
## Segment Expansion - Virtual segments \r
\r
A replication target may wish to retain historical data that the \r
replication source has not, or destroy historical data that the\r
replication source intends to retain.  This can result in the replication\r
target running out of room in a logical segment, and can also complicate\r
recovery operations if the replication source needs to self-heal a\r
corrupted segment.  This presents a problem because all filesystem\r
accesses and all replication and recovery activities are segment-centric.\r
\r
To deal with this situation a logical segment can be turned into a \r
virtual segment.  A virtual segment is made up of multiple logical\r
segments, all with the same identifier plus additional information\r
in the segment distinguishing the pieces from each other.  Each logical\r
segment is maintained independantly but API accesses check both\r
(or all N such segments) when doing lookups, and write to whichever\r
one has free space.  \r
\r
Virtual segments are pretty standard fare on replication slaves which\r
are intended to archive a great deal more history then replication\r
masters, but are intended to be very rare features of replication\r
masters.  If a virtual segment must be created on a replication master\r
the filesystem does all it can to shift data off the virtual segment\r
in order to be able to collapse it back into a logical segment.\r
\r
Virtual segmentation is not space efficient.\r
\r
## Files and Directories \r
\r
As has been described, a data object (which can be a file or directory\r
or other filesystem entity) is identified by a data object id, which\r
is effectively its inode, and a 64 bit key field.  The key field is\r
typically a base offset for a file or a sortable key for a directory\r
entry.  Negative numbers are used for meta-data.  For example, -1 will\r
be used for the inode data & stat information.  Other negative numbers\r
will be used to support other meta-data such as ACLs.\r
\r
Since records support variable-length data, the data storage for a file\r
is effectively extent-based.  Minimum granularity will be something like\r
32 bytes.\r
\r
Files are thus fairly straight forward.\r
\r
From the point of view of the filesystem each directory entry will\r
be a data record.  The data record will basically just contain the\r
inode number, type, and filename.  It will be variable-length insofar\r
as the filename is variable length.\r
\r
Most files will be representable in a single extent (or at least\r
can be optimized into a single extent), and so will not depend very\r
heavily on indexing.\r
\r
Directory lookups WILL depend on the indexing of the directory records\r
for efficiency, otherwise every record would have to be scanned to\r
lookup a filename.  In this regard a B+Tree is probably the best \r
solution.\r
\r
## Inode Number Uniqueness \r
\r
Inode numbers are 64 bits where the top N bits represent the segment\r
in which the inode was originally allocated, identifying the master\r
record for the data object.  Inode numbers do not change even if\r
the master record is relocated and the original master record replaced\r
with a forwarding record.  The forwarding record must be retained\r
in order to guarentee the uniqueness of the inode number.\r
\r
This also allows inode numbers to be allocated on a per-segment basis,\r
with 48 bits of uniqueness (one trillion file creates & deletes) within\r
each segment.\r
\r
The filesystem will always allocate new inode numbers using a sequence\r
number maintained in the segment header.  When all 48 bits are used up,\r
however, the filesystem must check new inode numbers against existing\r
numbers (including any deleted historical objects).  \r
\r
Inode number uniqueness over all time can no longer be guarenteed, but\r
inode number uniqueness over a period of time  **can**  still be guarenteed\r
by retaining the master record for deleted files for a period of time\r
 **even after they have been destroyed** .  The system operator can specify\r
the retention time with the caveat that certain benchmarks might cause\r
the disk to fill up or become somewhat less efficient even when \r
historical data is purged.\r
\r
It is recommended that any exported or clustered filesystems set the\r
inode number retention time to at least a week.  Inode number uniqueness\r
is crucial to cluster operation.\r