2 HAMMER2 DESIGN DOCUMENT
4 * These features have been speced in the media structures.
6 * Implementation work has begun.
8 * A working filesystem with some features implemented is expected by July 2012.
10 * A fully functional filesystem with most (but not all) features is expected
13 * All elements of the filesystem have been designed except for the freemap
14 (which isn't needed for initial work). 8MB per 2GB of filesystem
15 storage has been reserved for the freemap. The design of the freemap
16 is expected to be completely speced by mid-year.
18 * This is my only project this year. I'm not going to be doing any major
19 kernel bug hunting this year.
23 * Multiple roots (allowing snapshots to be mounted). This is implemented
24 via the super-root concept. When mounting a HAMMER2 filesystem you specify
25 a device path and a directory name in the super-root.
27 * HAMMER1 had PFS's. HAMMER2 does not. Instead, in HAMMER2 any directory
28 in the tree can be configured as a PFS, causing all elements recursively
29 underneath that directory to become a part of that PFS.
31 * Writable snapshots. Any subdirectory tree can be snapshotted. Snapshots
32 show up in the super-root. It is possible to snapshot a subdirectory
33 and then later snapshot a parent of that subdirectory... really there are
36 * Directory sub-hierarchy based quotas and space and inode usage tracking.
37 Any directory sub-tree, whether at a mount point or not, tracks aggregate
38 inode use and data space use. This is stored in the directory inode all
41 * Incremental queueless mirroring / mirroring-streams. Because HAMMER2 is
42 block-oriented and copy-on-write each blockref tracks both direct
43 modifications to the referenced data via (modify_tid) and indirect
44 modifications to the referenced data or any sub-tree via (mirror_tid).
45 This makes it possible to do an incremental scan of meta-data that covers
46 only changes made since the mirror_tid recorded in a prior-run.
48 This feature is also intended to be used to locate recently allocated
49 blocks and thus be able to fixup the freemap after a crash.
51 HAMMER2 mirroring works a bit differently than HAMMER1 mirroring in
52 that HAMMER2 does not keep track of 'deleted' records. Instead any
53 recursion by the mirroring code which finds that (modify_tid) has
54 been updated must also send the direct block table or indirect block
55 table state it winds up recursing through so the target can check
56 similar key ranges and locate elements to be deleted. This can be
57 avoided if the mirroring stream is mostly caught up in that very recent
58 deletions will be cached in memory and can be queried, allowing shorter
59 record deletions to be passed in the stream instead.
61 * Will support multiple compression algorithms configured on subdirectory
62 tree basis and on a file basis. Up to 64K block compression will be used.
63 Only compression ratios near powers of 2 that are at least 2:1 (e.g. 2:1,
64 4:1, 8:1, etc) will work in this scheme because physical block allocations
65 in HAMMER2 are always power-of-2.
67 Compression algorithm #0 will mean no compression and no zero-checking.
68 Compression algorithm #1 will mean zero-checking but no other compression.
69 Real compression will be supported starting with algorithm 2.
71 * Zero detection on write (writing all-zeros), which requires the data
72 buffer to be scanned, will be supported as compression algorithm #1.
73 This allows the writing of 0's to create holes and will be the default
74 compression algorithm for HAMMER2.
76 * Copies support for redundancy. The media blockref structure would
77 have become too bloated but I found a clean way to do copies using the
78 blockset structure (which is a set of 8 fully associative blockref's).
80 The design is such that the filesystem should be able to function at
81 full speed even if disks are pulled or inserted, as long as at least one
82 good copy is present. A background task will be needed to resynchronize
83 missing copies (or remove excessive copies in the case where the copies
84 value is reduced on a live filesystem).
86 * Intended to be clusterable, with a multi-master protocol under design
87 but not expected to be fully operational until mid-2013. The media
88 format for HAMMER1 was less condusive to logical clustering than I had
89 hoped so I was never able to get that aspect of my personal goals
90 working with HAMMER1. HAMMER2 effectively solves the issues that cropped
91 up with HAMMER1 (mainly that HAMMER1's B-Tree did not reflect the logical
92 file/directory hierarchy, making cache coherency very difficult).
94 * Hardlinks will be supported. All other standard features will be supported
95 too of course. Hardlinks in this sort of filesystem require significant
98 * The media blockref structure is now large enough to support up to a 192-bit
99 check value, which would typically be a cryptographic hash of some sort.
100 Multiple check value algorithms will be supported with the default being
101 a simple 32-bit iSCSI CRC.
103 * Fully verified deduplication will be supported and automatic (and
104 necessary in many respects).
106 * Non-verified de-duplication will be supported as a configurable option on
107 a file or subdirectory tree. Non-verified deduplication would use the
108 largest available check code (192 bits) and not bother to verify data
109 matches during the dedup pass, which is necessary on extremely large
110 filesystems with a great deal of deduplicable data (as otherwise a large
111 chunk of the media would have to be read to implement the dedup).
113 This feature is intended only for those files where occassional corruption
114 is ok, such as in a large data store of farmed web content.
118 HAMMER2 generally implements a copy-on-write block design for the filesystem,
119 which is very different from HAMMER1's B-Tree design. Because the design
120 is copy-on-write it can be trivially snapshotted simply by referencing an
121 existing block, and because the media structures logically match a standard
122 filesystem directory/file hierarchy snapshots and other similar operations
123 can be trivially performed on an entire subdirectory tree at any level in
126 The copy-on-write nature of the filesystem implies that any modification
127 whatsoever will have to eventually synchronize new disk blocks all the way
128 to the super-root of the filesystem and the volume header itself. This forms
129 the basis for crash recovery. All disk writes are to new blocks except for
130 the volume header, thus allowing all writes to run concurrently except for
131 the volume header update at the end.
133 Clearly this method requires intermediate modifications to the chain to be
134 cached so multiple modifications can be aggregated prior to being
135 synchronized. One advantage, however, is that the cache can be flushed at
136 any time WITHOUT having to allocate yet another new block when further
137 modifications are made as long as the volume header has not yet been flushed.
138 This means that buffer cache overhead is very well bounded and can handle
139 filesystem operations of any complexity even on boxes with very small amounts
142 I intend to implement a shortcut to make fsync()'s run fast, and that is to
143 allow deep updates to blockrefs to shortcut to auxillary space in the
144 volume header to satisfy the fsync requirement. The related blockref is
145 then recorded when the filesystem is mounted after a crash and the update
146 chain is reconstituted when a matching blockref is encountered again during
147 normal operation of the filesystem.
149 Basically this means that no real work needs to be done at mount-time
152 Directories are hashed, and another major design element is that directory
153 entries ARE INODES. They are one and the same. In addition to directory
154 entries being inodes the data for very small files (512 bytes or smaller)
155 can be directly embedded in the inode (overloaded onto the same space that
156 the direct blockref array uses). This should result in very high
159 Inode numbers are not spatially referenced, which complicates NFS servers
160 but doesn't complicate anything else. The inode number is stored in the
161 inode itself, an absolutely necessary feature in order to support the
162 hugely flexible snapshots that we want to have in HAMMER2.
166 Hardlinks are a particularly sticky problem for HAMMER2 due to the lack of
167 a spatial reference to the inode number. We do not want to have to have
168 an index of inode numbers for any basic HAMMER2 feature if we can help it.
170 Hardlinks are handled by placing the inode for a multiply-hardlinked file
171 in the closest common parent directory. If "a/x" and "a/y" are hardlinked
172 the inode for the hardlinked file will be placed in directory "a", e.g.
173 "a/3239944", but it will be invisible and will be in an out-of-band namespace.
174 The directory entries "a/x" and "a/y" will be given the same inode number
175 but in fact just be placemarks that cause HAMMER2 to recurse upwards through
176 the directory tree to find the invisible inode number.
178 Because directories are hashed and a different namespace (hash key range)
179 is used for hardlinked inodes, standard directory scans are able to trivially
180 skip this invisible namespace and inode-specific lookups can restrict their
181 lookup to within this space.
183 The nature of snapshotting makes handling link-count 2->1 and 1->2 cases
184 trivial. Basically the inode media structure is copied as needed to break-up
185 or re-form the standard directory entry/inode. There are no backpointers in
186 HAMMER2 and no reference counts on the blocks (see FREEMAP NOTES below), so
187 it is an utterly trivial operation.
191 In order to implement fast snapshots (and writable snapshots for that
192 matter), HAMMER2 does NOT ref-count allocations. The freemap which
193 is still under design just won't do that. All the freemap does is
194 keep track of 100% free blocks.
196 This not only trivializes all the snapshot features it also trivializes
197 hardlink handling and solves the problem of keeping the freemap sychronized
198 in the event of a crash. Now all we have to do after a crash is make
199 sure blocks allocated before the freemap was flushed are properly
200 marked as allocated in the allocmap. This is a trivial exercise using the
201 same algorithm the mirror streaming code uses (which is very similar to
202 HAMMER1)... an incremental meta-data scan that covers only the blocks that
203 might have been allocated between the last allocation map sync and now.
205 Thus the freemap does not have to be synchronized during a fsync().
207 The complexity is in figuring out what can be freed... that is, when one
208 can mark blocks in the freemap as being free. HAMMER2 implements this as
209 a background task which essentially must scan available meta-data to
210 determine which blocks are not being referenced.
212 Part of the ongoing design work is finding ways to reduce the scope of this
213 meta-data scan so the entire filesystem's meta-data does not need to be
214 scanned (though in tests with HAMMER1, even full meta-data scans have
215 turned out to be fairly low cost). In other words, its an area that we
216 can continue to improve on as the filesystem matures. Not only that, but
217 we can completely change the freemap algorithms without creating
218 incompatibilities (at worse simply having to require that a R+W mount do
219 a full meta-data scan when upgrading or downgrading the freemap algorithm).
223 Clustering, as always, is the most difficult bit but we have some advantages
224 with HAMMER2 that we did not have with HAMMER1. First, HAMMER2's media
225 structures generally follow the kernel's filesystem hiearchy. Second,
226 HAMMER2's writable snapshots make it possible to implement several forms
227 of multi-master clustering.
229 The general mechanics for most of the multi-master clustering implementations
232 (a) Use the copies mechanism to specify all elements of the cluster,
233 both local and remote (networked).
235 (b) The core synchronization state operates just as it does for copies,
236 simply requiring a fully-flushed ack from the remote in order to
237 mark the blocks as having been fully synchronized.
239 The mirror_tid may be used to locate these blocks, allowing the
240 synchronization state to be updated on the fly at a much later
241 time without requiring the state to be maintained in-memory.
242 (also for crash recovery resynchronization purposes).
244 (c) Data/meta-data can be retrieved from those copies which are marked
245 as being synchronized, with priority given to the local storage
246 relative to any given physical machine.
248 This means that e.g. even in a master-slave orientation the slave
249 may be able to satisfy a request from a program when the slave
250 happens to be the local storage.
252 (d) Transaction id synchronization between all elements of the cluster,
253 typically through masking (assigning a cluster number using the low
254 3 bits of the transaction id).
256 (e) General access (synchronized or otherwise) may require cache
257 coherency mechanisms to run over the network.
259 Implementing cache coherency is a major complexity issue.
261 (f) General access (synchronized or otherwise) may require quorum
262 agreement, using the synchronization flags in the blockrefs
263 to determine whether agreement has been reached.
265 Implementing quorum voting is a major complexity issue.
267 There are lots of ways to implement multi-master environments using the
268 above core features but the implementation is going to be fairly complex
269 even with HAMMER2's feature set.
271 Keep in mind that modifications propagate all the way to the super-root
272 and volume header, so in any clustered arrangement the use of (modify_tid)
273 and (mirror_tid) is critical in determining the synchronization state of
274 portion(s) of the filesystem.
276 Specifically, since any modification propagates to the root the (mirror_tid)
277 in higher level directories is going to be in a constant state of flux. This
278 state of flux DOES NOT invalidate the cache state for these higher levels
279 of directories. Instead, the (modify_tid) is used on a node-by-node basis
280 to determine cache state at any given level, and (mirror_tid) is used to
281 determine whether any recursively underlying state is desynchronized.
283 * Simple semi-synchronized multi-master environment.
285 In this environment all nodes are considered masters and modifications
286 can be made on any of them, and then propagate to the others
287 asynchronously via HAMMER2 mirror streams. One difference here is
288 that kernel can activate these userland-managed streams automatically
289 when the copies configuration is used to specify the cluster.
291 The only type of conflict which isn't readily resolvable by comparing
292 the (modify_tid) is when file data is updated. In this case user
293 intervention might be required but, theoretically, it should be
294 possible to automate most merges using a multi-way patch and, if not,
295 choosing one and creating backup copies if the others to allow the
296 user or sysop to resolve the conflict later.
298 * Simple fully synchronized fail-over environment.
300 In this environment there is one designated master and the remaining
301 nodes are slaves. If the master fails all remaining nodes agree on a
302 new master, possibly with the requirement that a quorum be achieved
303 (if you don't want to allow the cluster to split).
305 If network splits are allowed the each sub-cluster operates in this
306 mode but recombining the clusters reverts to the first algorithm.
307 If not allowed whomever no longer has a quorum will be forced to stall.
309 In this environment the current designated master is responsible for
310 managing locks for modifying operations. The designated master will
311 proactively tell the other nodes to mark the blocks related to the
312 modifying operation as no longer being synchronized while any local
313 data at the node that acquired the lock (master or slave) remains
314 marked as being synchronized.
316 The node that succesfully gets the lock then issues the modifying
317 operation to both its local copy and to the master, marking the
318 master as being desynchronized until the master acknowledges receipt.
320 In this environment any node can access data from local storage if
321 the designated master copy is marked synchronized AND its (modify_tid)
322 matches the slave copy's (modify_tid).
324 However, if a slave disconnects from the master then reconnects the
325 slave will have lost the master's desynchronization stream and must
326 mark its root blockref for the master copy HAMMER2_BREF_DESYNCHLD as
327 well as clear the SYNC1/SYNC2 bits. Setting DESYNCCHLD forces on-demand
328 recursive reverification that the master and slave are (or are not) in
329 sync in order to reestablish on the slave the synchronization state of
332 That might be a bit confusing but the whole point here is to allow
333 read accesses to the filesystem to be satisfied by any node in a
334 multi-master cluster, not just by the current designated master.
336 * Fully cache coherent and synchronized multi-master environment.
338 In this environment a quorum is required to perform any modifying
339 action. All nodes are masters (there is no 'designated' master)
340 and all nodes connect to all other nodes in a cross-bar.
342 The quorum is specified by copies setup in the root volume configuration.
343 A quorum of nodes in the cluster must agree on the copies configuration.
344 If they do not the cluster cannot proceed to mount. Any other nodes
345 not in the quorum which are in the cluster which disagree with the
346 configuration will inherit the copies configuration from the quorum.
348 Any modifying action will initiate a lock request locally to all nodes
349 in the cluster. The modifying action is allowed to proceed the instant
350 a quorum of nodes respond in the affirmative (even if some have not
351 yet responded or are down). The modifying action is considered complete
352 once the two-phase commit protocol succeeds. The modifying action
353 typically creates and commits a temporary snapshot on at least a quorum
354 of masters as phase-1 and then ties the snapshot back into the main
357 These locks are cache-coherency locks and may be passively maintained
358 in order to aggregate multiple operations under the same lock and thus
359 under the same transaction from the point of view of the rest of the
362 A lock request which interferes with a passively maintained lock will
363 force the two-phase commit protocol to complete and then transfer
364 ownership to the requesting entity, thus avoiding having to deal with
365 deadlock protocols at this point in the state machine.
367 Since any node can initiate concurrent lock requests to many other nodes
368 it is possible to deadlock. When two nodes initiate conflicting lock
369 requests to the cluster the one achieving the quorum basically wins and
370 the other is forced to retry (go back one paragraph). In this situation
371 no deadlock will occur.
373 If three are more nodes initiate conflicting lock requests to the
374 cluster a deadlock can occur whereby none of the nodes achieve a quorum.
375 In this case every node will know which of the other nodes was granted
376 the lock(s). Deadlock resolution then proceeds simultaniously on the
377 three nodes (since they have the same information), whereby the lock
378 holders on the losing end of the algorithm transfer their locks to one
379 of the other nodes. The lock state and knowledge of the lock state is
380 updated in real time on all nodes until a quorum is achieved.
382 * Fully cache coherent and synchronized multi-master environment with
383 passive read locking.
385 This is a more complex form of clustering than the previous form.
386 Take the previous form and add the ability to passively hold SHARED
387 locks in addition to the EXCLUSIVE locks the previous form is able
390 The advantage of being able to passively hold a shared lock on a sub-tree
391 (locks can be held on single nodes or entire sub-trees) is that it is
392 then possible for all nodes to validate a node (modify_tid) or entire
393 sub-tree (mirror_tid) with a very short network transaction and then
394 satisfy a large number of requests from local storage.
396 * Fully cache coherent and synchronized multi-master environment with
397 passive read locking and slave-only nodes.
399 This is the MOST complex form of clustering we intend to support.
400 In a multi-master environment requiring a quorum of masters to operate
401 we implement all of the above plus ALSO allow additional nodes to be
402 added to the cluster as slave-only nodes.
404 The difference between a slave-only node and setting up a manual
405 mirror-stream from the cluster to a read-only snapshot on another
406 HAMMER2 filesystem is that the slave-only node will be fully
407 cache coherent with either the cluster proper (if connected to a quorum
408 of masters), or to one or more other nodes in the cluster (if not
409 connected to a quorum of masters), EVEN if the slave itself is not
410 completely caught up.
412 So if the slave-only cluster node is connected to the rest of the cluster
413 over a slow connection you basically get a combination of local disk
414 speeds for any data that is locally in sync and network-limited speeds
415 for any data that is not locally in sync.
417 slave-only cluster nodes run a standard mirror-stream in the background
418 to pull in the data as quickly as possible.
420 This is in constrast to a manual mirror-stream to a read-only
421 snapshot (basically a simple slave), which has no ability to bypass
422 the local storage to handle out-of-date requests (in fact has no ability
423 to detect that the local storage is out-of-date anyway).