HAMMER2 DESIGN DOCUMENT
Matthew Dillon
dillon@backplane.com
08-Dec-2018 (v6)
24-Jul-2017 (v5)
09-Jul-2016 (v4)
03-Apr-2015 (v3)
14-May-2013 (v2)
08-Feb-2012 (v1)
Current Status as of document date
* Filesystem Core - operational
- bulkfree - operational
- Compression - operational
- Snapshots - operational
- Deduper - live operational, batch specced
- Subhierarchy quotas - scrapped (still possible on a limited basis)
- Logical Encryption - not specced yet
- Copies - not specced yet
- fsync bypass - not specced yet
- FS consistency - operational
* Clustering core
- Network msg core - operational
- Network blk device - operational
- Error handling - under development
- Quorum Protocol - under development
- Synchronization - under development
- Transaction replay - not specced yet
- Cache coherency - not specced yet
Recent Document Changes
* Reorganized the feature list to indicate currently operational features
first, and moving the future features to another section (since they
are taking so long to implement).
Current Features List
* Standard filesystem semantics with full hardlink and softlink support.
64-bit hardlink count field.
* The topology is indexed with a dynamic radix tree rooted in several
places: The super-root, the PFS root inode, and any inode boundary.
Index keys are 64-bits. Each element is referenced with a blockref
structure (described below) that is capable of referencing a power-of-2
sized block. The block size is currently capped at 64KB to play
nice(r) with the buffer cache and SSDs.
The dynamic radix tree pushes elements into new indirect blocks only
when the current level fills up, and will delete empty indirect blocks
when a level is cleaned out.
* Block-copy-on-write filesystem mechanism for both the main topology
and for the freemap. Media-level block frees are deferred and flushes
rotate between (up to) 4 volume headers (capped at 4 if the filesystem
is > ~8GB). Recovery will choose the most recent fully-valid volume
header and can thus work around failures which cause partial volume
header writes.
Modifications issue copy-on-write updates up to the volume root.
* Utilizes a fat blockref structure (128 bytes) which can store up to
64 bytes (512 bits) of check code data for each referenced block.
In the original implementation I had gone with 64 byte blockrefs,
but I eventually decided that I wanted to support up to a 512-bit
hash (which eats 64 bytes), so I bumped it up to 128 bytes. This
turned out to be fortuitous because it made it possible to store
most directory entries directly in the blockref structure without
having to reference a separate data block via the blockref structure.
* 1KB 'fat' inode structure. The inode structure directly embeds four
blockrefs so small files and directories can be represented without
requiring an indirect block to be allocated. The inode structure can
also overload the same space to store up to 512 bytes of direct
file data (for files which are <= 512 bytes long).
The super-root and PFS root inodes are directly represented in the
topology, without the use of directory entries. A combination of
normal directory entries and separtely-indexed inodes are implemented
under each PFS.
Normal filesystem inodes (other than inode 1) are indexed under the PFS
root inode by their inode number. Directory entries are indexed under the
same PFS root by their filename hash. Bit 63 is used to distinguish and
partition the two. Filename hash collisions are handled by incrementing
reserved low bits in the filename hash code.
* Directory entries representing filenames that are less than 64 bytes
long are directly stored AS blockrefs. This means that an inode
representing a small directory can store up to 4 directory entries in
the inode itself before resorting to indirect blocks, and then those
indirect blocks themselves can directly embed up to 512 directory entries.
Directory entries with long filenames reference an indirect data block
to hold the filename instead of directly-embedding the filename.
This results in *very* compact directories in terms of I/O bandwidth.
Not as compact as e.g. UFS's variable-length directory entries, but still
very good with a nominal 128 real bytes per directory entry.
Because directory entries are represented using a dynamic radix tree via
its blockrefs, directory entries can be randomly looked up without having
to scan the whole directory.
* Multiple PFSs. In HAMMER2, all PFSs are implemented the same way, with
the kernel choosing a default PFS name for the mount if none is specified.
For example, "ROOT" is the default PFS name for a root mount. You can
create as many PFSs as you like and you can specify the PFS name in the
mount command using the <device_path>@<pfs_name> notation.
* Snapshots are implemented as PFSs. Due to the copy-on-write nature of
the filesystem, taking a snapshot is a trivial operation requiring only
a normal filesystme sync and copying of the PFS root inode (1KB), and
that's it.
On the minus side, can complicate the bulkfree operation that is responsible
for freeing up disk space. It can take significantly longer when many
snapshots are present.
* SNAPSHOTS ARE READ-WRITE. You can mount any PFS read-write, including
snapshots. For example, you can revert to an earlier 'root' that you
made a snapshot of simply by changing what the system mounts as the root
filesystem.
* Full filesystem coherency at both the radix tree level and the filesystem
semantics level. This is true for all filesystem syncs, recovery after
a crash, and snapshots.
The filesystem syncs fully vfsync the buffer cache for the files
that are part of the sync group, and keeps track of dependencies to
ensure that all inter-dependent inodes are flushed in the same sync
group. Atomic filesystem ops such as write()s are guaranteed to remain
atomic across a sync, snapshot, and crash.
* Flushes and syncs are almost entirely asynchronous and will run concurrent
with frontend operations. This feature is implemented by adding inodes
to the sync group currently being flushed on-the-fly as new dependencies
are created, and reordering inodes in the sync queue to prioritize inodes
which the frontend is stalled on.
By reprioritizing inodes in the syncq, frontend stalls are minimized.
The only synchronous disk operations is the final sync of the volume
header which updates the ultimate root of the filesystem. A disk flush
command is issued synchronously, then the write of the volume header is
issued synchronously. All other writes to the disk, regardless of the
complexity of the dependencies, occur asynchronously and can make very
good use of high-speed I/O and SSD bandwidth.
* Low memory footprint. Except for the volume header, the buffer cache
is completely asynchronous and dirty buffers can be retired by the OS
directly to backing store with no further interactions with the filesystem.
* Compression support. Multiple algorithms are supported and can be
configured on a subdirectory hierarchy or individual file basis.
Block compression up to 64KB will be used. Only compression ratios at
powers of 2 that are at least 2:1 (e.g. 2:1, 4:1, 8:1, etc) will work in
this scheme because physical block allocations in HAMMER2 are always
power-of-2.
Modest compression can be achieved with low overhead, is turned on
by default, and is compatible with deduplication.
Compression is extremely useful and often gives you anywhere from 25%
to 400% the logical storage as you have physical blocks, depending.
Of course, .tgz and other pre-compressed files cannot be compressed
further by the filesystem.
The usefulness shnould not be underestimated, our users are constantly
being surprised at things the filesystem is able to compres that just
makes life a lot easier. For example, 30GB core dumps tend to contain
a great deal of highly compressable data. Source trees, web files,
executables, general data... this is why HAMMER2 turns modest compression
on by default. It just works.
* De-duplication support. HAMMER2 uses a relatively simple freemap
scheme that allows the filesystem to discard block references
asynchronously. The same scheme allows essentially unlimited references
to the same data block in the hierarchy. Thus, both live de-duplication
and bulk deduplication are relatively easy to implement.
HAMMER2 currently implements only live de-duplications. This means that
typical situations such as when copying files or whole directory hierarchies
will naturally de-duplicate. Simply reading filesystem data in makes
it available for deduplication later. HAMMER2 will index a potentially
very large number of blocks in memory, even beyond what the buffer cache
can hold, for deduplication purposes.
* Zero-fill detection on write (writing all-zeros), which requires the data
buffer to be scanned, is fully supported. This allows the writing of 0's
to create holes.
Generally speaking pre-writing zerod blocks to reserve space doesn't work
well on copy-on-write filesystems. However, if both compression and
check codes are disabled on a file, H2 will also disable zero-detection,
allowing the file blocks to be pre-reserved (by actually zeroing them and
reusing them later on), and allow data overwrites to write to the same
sector. Please be aware that DISABLING THE CHECK CODE IN THIS MANNER ALSO
MEANS THAT SNAPSHOTS WILL NOT WORK. The snapshot will contain the latest
data for the file and not the data as-of the snapshot. This is NOT turned
on by default in HAMMER2 and is not recommended except in special
well-controlled circumstances.
* Multiple supporting kernel threads, breaking up frontend VOP operation
from backend I/O, compression, and decompression operation. Buffer cache
I/O and VOP ops message the backend. Actual I/O is handled by the backend
and not by the frontend, which will theoretically allow us to survive
stalled devices and nodes when implementing multi-node support.
Pending Features
(not yet implemented)
* Constructing a filesystem across multiple nodes. Each low-level H2 device
would be able to accommodate nodes belonging to multiple cluster components
as well as nodes that are simply local to the device or machine.
CURRENT STATUS: Not yet operational.
* Incremental synchronization via highest-transaction id propagation
within the radix tree. This is a queueless, incremental design.
CURRENT STATUS: Due to the flat inode hierarchy now being employed,
the current synchronization code which silently recurses indirect nodes
will be inefficient due to the fact that all the inodes are at the
same logical level in the topology. To fix this, the code will need
to explicitly iterate indirect nodes and keep track of the related
key ranges to match them up on an indirect-block basis, which would
be incredibly efficient.
* Background synchronization and mirroring occurs at the logical layer
rather than the physical layer. This allows cluster components to
have differing storage arrangements.
In addition, this mechanism will fully correct any out of sync nodes
in the cluster as long as a sufficient number of other nodes agree on
what the proper state should be.
CURRENT STATUS: Not yet operational.
* Encryption. Whole-disk encryption is supported by another layer, but I
intend to give H2 an encryption feature at the logical layer which works
approximately as follows:
- Encryption controlled by the client on an inode/sub-tree basis.
- Server has no visibility to decrypted data.
- Encrypt filenames in directory entries. Since the filename[] array
is 256 bytes wide, client can add random bytes after the normal
terminator to make it virtually impossible for an attacker to figure
out the filename.
- Encrypt file size and most inode contents.
- Encrypt file data (holes are not encrypted).
- Encryption occurs after compression, with random filler.
- Check codes calculated after encryption & compression (not before).
- Blockrefs are not encrypted.
- Directory and File Topology is not encrypted.
- Encryption is not sub-topology validation. Client would have to keep
track of that itself. Server or other clients can still e.g. remove
files, rename, etc.
In particular, note that even though the file size field can be encrypted,
the server does have visibility on the block topology and thus has a pretty
good idea how big the file is. However, a client could add junk blocks
at the end of a file to make this less apparent, at the cost of space.
If a client really wants a fully validated H2-encrypted space the easiest
solution is to format a filesystem within an encrypted file by treating it
as a block device, but I digress.
CURRENT STATUS: Not yet operational.
* Device ganging, copies for redundancy, and file splitting.
Device ganging - The idea here is not to gang devices into a single
physical volume but to instead format each device independently
and allow crossover-references in the blockref to other devices in
the set.
One of the things we want to accomplish is to ensure that a failed
device does not prevent access to radix tree elements in other devices
in the gang, and that the failed device can be reconstructed. To do
this, each device implements complete reachability from the node root
to all elements underneath it. When a device fails, the sychronization
code can theoretically reconstruct the missing material in other
devices making up the gang. New devices can be added to the gang and
existing devices can be removed from the gang.
Redundant copies - This is actually a fairly tough problem. The
solution I would like to implement is to use the device ganging feature
to also implement redundancy, that way if a device fails within the
gang there's a good chance that it can still remain completely functional
without having to resynchronize. But making this work is difficult to say
the least.
CURRENT STATUS: Not yet operational.
* MESI Cache coherency for multi-master/multi-client clustering operations.
The servers hosting the MASTERs are also responsible for keeping track of
the cache state.
This is a feature that we would need to implement coherent cross-machine
multi-threading and migration.
CURRENT STATUS: Not yet operational.
* Implement unverified de-duplication (where only the check code is tested,
avoiding having to actually read data blocks to calculate a de-duplication.
This would make use of the blockref structure's widest check field
(512 bits).
Out of necessity this type of feature would be settable on a file or
recursive directory tree basis, but should only be used when the data
is throw-away or can be reconstructed since data corruption (mismatched
duplicates with the same hash) is still possible even with a 512-bit
check code.
The Unverified dedup feature is intended only for those files where
occasional corruption is ok, such as in a web-crawler data store or
other situations where the data content is not critically important
or can be externally recovered if it becomes corrupt.
CURRENT STATUS: Not yet operational.
GENERAL DESIGN
HAMMER2 generally implements a copy-on-write block design for the filesystem,
which is very different from HAMMER1's B-Tree design. Because the design
is copy-on-write it can be trivially snapshotted simply by making a copy
of the block table we desire to snapshot. Snapshotting the root inode
effectively snapshots the entire filesystem, whereas snapshotting a file
inode only snapshots that one file. Snapshotting a directory inode is
generally unhelpful since it only contains directory entries and the
underlying files are not arranged under it in the radix tree.
The copy-on-write design implements a block table as a radix-tree,
with a small fan-out in the volume header and inode (typically 4x) and
a large fan-out for indirect blocks (typically 128x and 512x depending).
The table is built bottom-up. Intermediate radixes are only created when
necessary so small files and directories will have a much shallower radix
tree.
HAMMER2 implements several space optimizations:
1. Directory entries with filenames <= 64 bytes will fit entirely
in the 128-byte blockref structure and do not require additional data
block references. Since blockrefs are the core elements making up
block tables, most directories should have good locality of reference
for directory scans.
Filenames > 64 bytes require a 1KB data-block reference, which
is clearly less optimal, but very few files in a filesystem tend
to be larger than 64 bytes so it works out. This also simplifies
the handling for large filenames as we can allow filenames up to
1023 bytes long with this mechanism with no major changes to the
code.
2. Inodes embed 4 blockrefs, so files up to 256KB and directories with
up to four directory entries (not including "." or "..") can be
accommodated without requiring any indirecct blocks.
3. Indirect blocks can be sized to any power of two up to 65536 bytes,
and H2 typically uses 16384 and 65536 bytes. The smaller size is
used for initial indirect blocks to reduce storage overhead for
medium-sized files and directories.
4. The File inode itself can directly hold the data for small
files <= 512 bytes in size, overloading the space also used
by its four 128 bytes blockrefs (which are not needed if the
file is <= 512 bytes in size). This works out great for small
files and directories.
5. The last block in a file will have a storage allocation in powers
of 2 from 1KB to 64KB as needed. Thus a small file in excess of
512 bytes but less than 64KB will not waste a full 64KB block.
6. When compression is enabled, small physical blocks will be allocated
when possible. However, only reductions in powers of 2 are supported.
So if a 64KB data block can be compressed to (16KB+1) to 32KB, then
a 32KB block will be used. This gives H2 modest compression at very
low cost without too much added complexity.
7. Live de-dup will attempt to share data blocks when file copying is
detected, significantly reducing actual physical writes to storage
and the storage used. Bulk de-dup (when implemented), will catch
other cases of de-duplication.
Directories contain directory entries which are indexed using a hash of
their filename. The hash is carefully designed to maintain some natural
sort ordering. The directory entries are implemented AS blockrefs. So
an inode can contain up to 4 before requiring an indirect block, and
each indirect block can contain up to 512 entries, with further data block
references required for any directory entry whos filename is > 64 bytes.
Because the directory entries are blockrefs, random access lookups are
maximally efficient. The directory hash is designed to very loosely try
to retain some alphanumeric sorting to bundle similarly-named files together
and reduce random lookups.
The copy-on-write nature of the filesystem means that any modification
whatsoever will have to eventually synchronize new disk blocks all the way
to the super-root of the filesystem and then to the volume header itself.
This forms the basis for crash recovery and also ensures that recovery
occurs on a completed high-level transaction boundary. All disk writes are
to new blocks except for the volume header (which cycles through 4 copies),
thus allowing all writes to run asynchronously and concurrently prior to
and during a flush, and then just doing a final synchronization and volume
header update at the end. Many of HAMMER2s features are enabled by this
core design feature.
The Freemap is also implemented using a radix tree via a set of pre-reserved
blocks (approximately 4MB for every 2GB of storage), and also cycles through
multiple copies to ensure that crash recovery can restore the state of the
filesystem quickly at mount time.
HAMMER2 tries to maintain a small footprint and one way it does this is
by using the normal buffer cache for data and meta-data, and allowing the
kernel to asynchronously flush device buffers at any time (even during
synchronization). The volume root is flushed separately, separated from
the asynchronous flushes by a synchronizing BUF_CMD_FLUSH op. This means
that HAMMER2 has very low resource overhead from the point of view of the
operating system and is very much unlike HAMMER1 which had to lock dirty
buffers into memory for long periods of time. HAMMER2 has no such
requirement.
Buffer cache overhead is very well bounded and can handle filesystem
operations of any complexity, even on boxes with very small amounts
of physical memory. Buffer cache overhead is significantly lower with H2
than with H1 (and orders of magnitude lower than ZFS).
At some point I intend to implement a shortcut to make fsync()'s run fast,
and that is to allow deep updates to blockrefs to shortcut to auxillary
space in the volume header to satisfy the fsync requirement. The related
blockref is then recorded when the filesystem is mounted after a crash and
the update chain is reconstituted when a matching blockref is encountered
again during normal operation of the filesystem.
FILESYSTEM SYNC SEQUENCING
HAMMER2 implements a filesystem sync mechanism that allows the frontend
to continue doing modifying operations concurrent with the sync. The
general sync mechanism operates in four phases:
1. Individual file and directory inodes are fsync()d to disk,
updated the blockrefs in the parent block above the inode, and
removed from the syncq.
Once removed from the syncq, the frontend can do a modifying
operation on these file and directory inodes without further
effecting the filesystem sync. These modifications will be
flushed to disk on the next filesystem sync.
To reduce frontend stall times, an inode blocked on by the frontend
which is on the syncq will be reordered to the front of the syncq
to give the syncer a shot at it more quickly, in order to unstall
the frontend ASAP.
If a frontend operations creates an unavoidable dependency between
an inode on the syncq and an inode not on the syncq, both inodes
are placed on (or back onto) the syncq as needed to ensure filesystem
consistency for the filesystem sync. This can extend the filesystem
sync time, but even under heavy loads syncs are still able to be
retired.
2. The PFS ROOT is fsync()d to storage along with the subhierarchy
representing the inode index (whos inodes were flushed in (1)).
This brings the block copy-on-write up to the root inode.
3. The SUPER-ROOT inode is fsync()d to storage along with the
subhierarchy representing the PFS ROOTs for the volume.
4. Finally, a physical disk flush command is issued to the storage
device, and then the volume header is written to disk. All
I/O prior to this step occurred asynchronously. This is the only
step which must occur synchronously.
MIRROR_TID, MODIFY_TID, UPDATE_TID
In HAMMER2, the core block reference is a 128-byte structure called a blockref.
The blockref contains various bits of information including the 64-bit radix
key (typically a directory hash if a directory entry, inode number if a
hidden hardlink target, or file offset if a file block), number of significant
key bits for ranged recursion of indirect blocks, a 64-bit device seek that
encodes the radix of the physical block size in the low bits (physical block
size can be different from logical block size due to compression),
three 64-bit transaction ids, type information, and up to 512 bits worth
of check data for the block being reference which can be anything from
a simple CRC to a strong cryptographic hash.
mirror_tid - This is a media-centric (as in physical disk partition)
transaction id which tracks media-level updates. The mirror_tid
can be different at the same point on different nodes in a
cluster.
Whenever any block in the media topology is modified, its
mirror_tid is updated with the flush id and will propagate
upward during the flush all the way to the volume header.
mirror_tid is monotonic. It is primarily used for on-mount
recovery and volume root validation. The name is historical
from H1, it is not used for nominal mirroring.
modify_tid - This is a cluster-centric (as in across all the nodes used
to build a cluster) transaction id which tracks filesystem-level
updates.
modify_tid is updated when the front-end of the filesystem makes
a change to an inode or data block. It does NOT propagate upward
during a flush.
update_tid - This is a cluster synchronization transaction id. Modifications
made to the topology will clear this field to 0 as they propagate
up to the root. This gives the synchronizer an easy way to
determine what needs revalidation.
The synchronizer revalidates the cluster bottom-up by validating
a sub-topology and propagating the highest modify_tid in the
validated sub-topology up via the update_tid field.
Update to this field may be optimized by the HAMMER2 VFS to
avoid the double-transition.
The synchronization code updates an out-of-sync node bottom-up and will
dynamically set update_tid as it goes, but media flushes can occur at any
time and these flushes will use mirror_tid for flush and freemap management.
The mirror_tid for each flush propagates upward to the volume header on each
flush. modify_tid is set for any chains modified by a cluster op but does
not propagate up, instead serving as a seed for update_tid.
* The synchronization code is able to determine that a sub-tree is
synchronized simply by observing the update_tid at the root of the sub-tree,
on an inode-by-inode basis and also on a data-block-by-data-block basis.
* The synchronization code is able to do an incremental update of an
out-of-sync node simply by skipping elements with a matching update_tid
(when not 0).
* The synchronization code can be interrupted and restarted at any time,
and is able to pick up where it left off with very low overhead.
* The synchronization code does not inhibit media flushes. Media flushes
can occur (and must occur) while synchronization is ongoing.
There are several other stored transaction ids in HAMMER2. There is a
separate freemap_tid in the volume header that is used to allow freemap
flushes to be deferred, and inodes have a pfs_psnap_tid which is used in
conjunction with CHECK_NONE to allow blocks without a check code which do
not violate the most recent snapshot to be overwritten in-place.
Remember that since this is a copy-on-write filesystem, we can propagate
a considerable amount of information up the tree to the volume header
without adding to the I/O we already have to do.
DIRECTORIES AND INODES
Directories are hashed. In HAMMER2, the PFS ROOT directory (aka inode 1 for
a PFS) can contain a mix of directory entries AND embedded inodes. This was
actually a design mistake, so the code to deal with the index of inodes
vs the directory entries is slightly convoluted (but not too bad).
In the first iteration of HAMMER2 I tried really hard to embed actual
inodes AS the directory entries, but it created a mass of problems for
implementing NFS export support and dealing with hardlinks, so in a later
iteration I implemented small independent directory entries (that wound up
mostly fitting in the blockref structure, so WIN WIN!). However, 'embedded'
inodes AS the directory entries still survive for the SUPER-ROOT and the
PFS-ROOTs under the SUPER-ROOT. They just aren't used in the individual
filesystem that each PFS represents.
Hardlinks are now implemented normally, with multiple directory entries
referencing the same inode and that inode containing a nlinks count.
RECOVERY
H2 allows freemap flushes to lag behind topology flushes. This improves
filesystem sync performance. The freemap flush tracks a separate
transaction id (via mirror_tid) in the volume header.
On mount, HAMMER2 will first locate the highest-sequenced check-code-validated
volume header from the 4 copies available (if the filesystem is big enough,
e.g. > ~8GB or so, there will be 4 copies of the volume header).
HAMMER2 will then run an incremental scan of the topology for mirror_tid
transaction ids between the last freemap flush tid and the last topology
flush tid in order to synchronize the freemap. Because this scan is
incremental the time it takes to run will be relatively short and well-bounded
at mount-time. This is NOT an fsck. Freemap flushes can be avoided for any
number of normal topology flushes but should still occur frequently enough
to avoid long recovery times in case of a crash.
The filesystem is then ready for use.
DISK I/O OPTIMIZATIONS
The freemap implements a 1KB allocation resolution. Each 2MB segment managed
by the freemap is zoned and has a tendency to collect inodes, small data,
indirect blocks, and larger data blocks into separate segments. The idea is
to greatly improve I/O performance (particularly by laying inodes down next
to each other which has a huge effect on directory scans).
The current implementation of HAMMER2 implements a fixed 64KB physical block
size in order to allow the mapping of hammer2_dio's in its IO subsystem
to consumers that might desire different sizes. This way we don't have to
worry about matching the buffer cache / DIO cache to the variable block
size of underlying elements. In addition, 64KB I/Os allow compatibility
with physical sector sizes up to 64KB in the underlying physical storage
with no change in the byte-by-byte format of the filesystem. The DIO
layer also prevents ordering deadlocks between unrelated portions of the
filesystem hierarchy whos logical blocks wind up in the same physical block.
The biggest issue we are avoiding by having a fixed 64KB I/O size is not
actually to help nominal front-end access issue but instead to reduce the
complexity of having to deal with mixed block sizes in the buffer cache,
particularly when blocks are freed and then later reused with a different
block size. HAMMER1 had to have specialized code to check for and
invalidate buffer cache buffers in the free/reuse case. HAMMER2 does not
need such code.
That said, HAMMER2 places no major restrictions on mixing logical block
sizes within a 64KB block. The only restriction is that a logical HAMMER2
block cannot cross a 64KB boundary. The soft restrictions the block
allocator puts in place exist primarily for performance reasons (i.e. to
try to collect 1K inodes together). The 2MB freemap zone granularity
should work very well in this regard.
HAMMER2 also utilizes OS support for ganging 64KB buffers together into even
larger blocks for I/O (OS buffer cache 'clustering'), OS-supported read-ahead,
OS-driven asynchronous retirement, and other performance features typically
provided by the OS at the block-level to ensure smooth system operation.
By avoiding wiring buffers/memory and allowing the OS's buffer cache to
run normally, HAMMER2 winds up with very low OS overhead.
FREEMAP NOTES
The freemap is stored in the reserved blocks situated in the ~4MB reserved
area at the base of every ~1GB level-1 zone of physical storage. The current
implementation reserves 8 copies of every freemap block and cycles through
them in order to make the freemap operate in a copy-on-write fashion.
- Freemap is copy-on-write.
- Freemap operations are transactional, same as everything else.
- All backup volume headers are consistent on-mount.
The Freemap is organized using the same radix blockmap algorithm used for
files and directories, but with fixed radix values. For a maximally-sized
filesystem the Freemap will wind up being a 5-level-deep radix blockmap,
but the top-level is embedded in the volume header so insofar as performance
goes it is really just a 4-level blockmap.
The freemap radix allocation mechanism is also the same, meaning that it is
bottom-up and will not allocate unnecessary intermediate levels for smaller
filesystems. The number of blockmap levels not including the volume header
for various filesystem sizes is as follows:
up-to #of freemap levels
1GB 1-level
256GB 2-level
64TB 3-level
16PB 4-level
4EB 5-level
16EB 6-level
The Freemap has bitmap granularity down to 16KB and a linear iterator that
can linearly allocate space down to 1KB. Due to fragmentation it is possible
for the linear allocator to become marginalized, but it is relatively easy
to reallocate small blocks every once in a while (like once a year if you
care at all) and once the old data cycles out of the snapshots, or you also
rewrite the snapshots (which you can do), the freemap should wind up
relatively optimal again. Generally speaking I believe that algorithms can
be developed to make this a non-problem without requiring any media structure
changes. However, touching all the freemaps will replicate meta-data whereas
the meta-data was mostly shared in the original snapshot. So this is a
problem that needs solving in HAMMER2.
In order to implement fast snapshots (and writable snapshots for that
matter), HAMMER2 does NOT ref-count allocations. All the freemap does is
keep track of 100% free blocks plus some extra bits for staging the bulkfree
scan. The lack of ref-counting makes it possible to:
- Completely trivialize HAMMER2s snapshot operations.
- Completely trivialize HAMMER2s de-dup operations.
- Allows any volume header backup to be used trivially.
- Allows whole sub-trees to be destroyed without having to scan them.
Deleting PFSs and snapshots is instant (though space recovery still
requires two bulkfree scans).
- Simplifies normal crash recovery operations by not having to reconcile
a ref-count.
- Simplifies catastrophic recovery operations for the same reason.
Normal crash recovery is simply a matter of doing an incremental scan
of the topology between the last flushed freemap TID and the last flushed
topology TID. This usually takes only a few seconds and allows:
- Freemap flushes to be be deferred for any number of topology flush
cycles (with some care to ensure that all four volume headers
remain valid).
- Does not have to be flushed for fsync, reducing fsync overhead.
FREEMAP - BULKFREE
Blocks are freed via a bulkfree scan, which is a two-stage meta-data scan.
Blocks are first marked as being possibly free and then finalized in the
second scan. Live filesystem operations are allowed to run during these
scans and any freemap block that is allocated or adjusted after the first
scan will simply be re-marked as allocated and the second scan will not
transition it to being free.
The cost of not doing ref-count tracking is that HAMMER2 must perform two
bulkfree scans of the meta-data to determine which blocks can actually be
freed. This can be complicated by the volume header backups and snapshots
which cause the same meta-data topology to be scanned over and over again,
but mitigated somewhat by keeping a cache of higher-level nodes to detect
when we would scan a sub-topology that we have already scanned. Due to the
copy-on-write nature of the filesystem, such detection is easy to implement.
Part of the ongoing design work is finding ways to reduce the scope of this
meta-data scan so the entire filesystem's meta-data does not need to be
scanned (though in tests with HAMMER1, even full meta-data scans have
turned out to be fairly low cost). In other words, its an area where
improvements can be made without any media format changes.
Another advantage of operating the freemap like this is that some future
version of HAMMER2 might decide to completely change how the freemap works
and would be able to make the change with relatively low downtime.
CLUSTERING
Clustering, as always, is the most difficult bit but we have some advantages
with HAMMER2 that we did not have with HAMMER1. First, HAMMER2's media
structures generally follow the kernel's filesystem hierarchy which allows
cluster operations to use topology cache and lock state. Second,
HAMMER2's writable snapshots make it possible to implement several forms
of multi-master clustering.
The mount device path you specify serves to bootstrap your entry into
the cluster. This is typically local media. It can even be a ram-disk
that only contains placemarkers that help HAMMER2 connect to a fully
networked cluster.
With HAMMER2 you mount a directory entry under the super-root. This entry
will contain a cluster identifier that helps HAMMER2 identify and integrate
with the nodes making up the cluster. HAMMER2 will automatically integrate
*all* entries under the super-root when you mount one of them. You have to
mount at least one for HAMMER2 to integrate the block device in the larger
cluster.
For cluster servers every HAMMER2-formatted partition has a "LOCAL" MASTER
which can be mounted in order to make the rest of the elements under the
super-root available to the network. (In a prior specification I emplaced
the cluster connections in the volume header's configuration space but I no
longer do that).
Connecting to the wider networked cluster involves setting up the /etc/hammer2
directory with appropriate IP addresses and keys. The user-mode hammer2
service daemon maintains the connections and performs graph operations
via libdmsg.
Node types within the cluster:
DUMMY - Used as a local placeholder (typically in ramdisk)
CACHE - Used as a local placeholder and cache (typically on a SSD)
SLAVE - A SLAVE in the cluster, can source data on quorum agreement.
MASTER - A MASTER in the cluster, can source and sink data on quorum
agreement.
SOFT_SLAVE - A SLAVE in the cluster, can source data locally without
quorum agreement (must be directly mounted).
SOFT_MASTER - A local MASTER but *not* a MASTER in the cluster. Can source
and sink data locally without quorum agreement, intended to
be synchronized with the real MASTERs when connectivity
allows. Operations are not coherent with the real MASTERS
even when they are available.
NOTE: SNAPSHOT, AUTOSNAP, etc represent sub-types, typically under a
SLAVE. A SNAPSHOT or AUTOSNAP is a SLAVE sub-type that is no longer
synchronized against current masters.
NOTE: Any SLAVE or other copy can be turned into its own writable MASTER
by giving it a unique cluster id, taking it out of the cluster that
originally spawned it.
There are four major protocols:
Quorum protocol
This protocol is used between MASTER nodes to vote on operations
and resolve deadlocks.
This protocol is used between SOFT_MASTER nodes in a sub-cluster
to vote on operations, resolve deadlocks, determine what the latest
transaction id for an element is, and to perform commits.
Cache sub-protocol
This is the MESI sub-protocol which runs under the Quorum
protocol. This protocol is used to maintain cache state for
sub-trees to ensure that operations remain cache coherent.
Depending on administrative rights this protocol may or may
not allow a leaf node in the cluster to hold a cache element
indefinitely. The administrative controller may preemptively
downgrade a leaf with insufficient administrative rights
without giving it a chance to synchronize any modified state
back to the cluster.
Proxy protocol
The Quorum and Cache protocols only operate between MASTER
and SOFT_MASTER nodes. All other node types must use the
Proxy protocol to perform similar actions. This protocol
differs in that proxy requests are typically sent to just
one adjacent node and that node then maintains state and
forwards the request or performs the required operation.
When the link is lost to the proxy, the proxy automatically
forwards a deletion of the state to the other nodes based on
what it has recorded.
If a leaf has insufficient administrative rights it may not
be allowed to actually initiate a quorum operation and may only
be allowed to maintain partial MESI cache state or perhaps none
at all (since cache state can block other machines in the
cluster). Instead a leaf with insufficient rights will have to
make due with a preemptive loss of cache state and any allowed
modifying operations will have to be forwarded to the proxy which
continues forwarding it until a node with sufficient administrative
rights is encountered.
To reduce issues and give the cluster more breath, sub-clusters
made up of SOFT_MASTERs can be formed in order to provide full
cache coherent within a subset of machines and yet still tie them
into a greater cluster that they normally would not have such
access to. This effectively makes it possible to create a two
or three-tier fan-out of groups of machines which are cache-coherent
within the group, but perhaps not between groups, and use other
means to synchronize between the groups.
Media protocol
This is basically the physical media protocol.
MASTER & SLAVE SYNCHRONIZATION
With HAMMER2 I really want to be hard-nosed about the consistency of the
filesystem, including the consistency of SLAVEs (snapshots, etc). In order
to guarantee consistency we take advantage of the copy-on-write nature of
the filesystem by forking consistent nodes and using the forked copy as the
source for synchronization.
Similarly, the target for synchronization is not updated on the fly but instead
is also forked and the forked copy is updated. When synchronization is
complete, forked sources can be thrown away and forked copies can replace
the original synchronization target.
This may seem complex, but 'forking a copy' is actually a virtually free
operation. The top-level inode (under the super-root), on-media, is simply
copied to a new inode and poof, we have an unchanging snapshot to work with.
- Making a snapshot is fast... almost instantanious.
- Snapshots are used for various purposes, including synchronization
of out-of-date nodes.
- A snapshot can be converted into a MASTER or some other PFS type.
- A snapshot can be forked off from its parent cluster entirely and
turned into its own writable filesystem, either as a single MASTER
or this can be done across the cluster by forking a quorum+ of
existing MASTERs and transferring them all to a new cluster id.
More complex is reintegrating the target once the synchronization is complete.
For SLAVEs we just delete the old SLAVE and rename the copy to the same name.
However, if the SLAVE is mounted and not optioned as a static mount (that is
the mounter wants to see updates as they are synchronized), a reconciliation
must occur on the live mount to clean up the vnode, inode, and chain caches
and shift any remaining vnodes over to the updated copy.
- A mounted SLAVE can track updates made to the SLAVE but the
actual mechanism is that the SLAVE PFS is replaced with an
updated copy, typically every 30-60 seconds.
Reintegrating a MASTER which has fallen out of the quorum due to being out
of date is also somewhat more complex. The same updating mechanic is used,
we actually have to throw the 'old' MASTER away once the new one has been
updated. However if the cluster is undergoing heavy modifications the
updated MASTER will be out of date almost the instant its source is
snapshotted. Reintegrating a MASTER thus requires a somewhat more complex
interaction.
- If a MASTER is really out of date we can run one or more
synchronization passes concurrent with modifying operations.
The quorum can remain live.
- A final synchronization pass is required with quorum operations
blocked to reintegrate the now up-to-date MASTER into the cluster.
QUORUM OPERATIONS
Quorum operations can be broken down into HARD BLOCK operations and NETWORK
operations. If your MASTERs are all local mounts, then failures and
sequencing is easy to deal with.
Quorum operations on a networked cluster are more complex. The problems:
- Masters cannot rely on clients to moderate quorum transactions.
Apart from the reliance being unsafe, the client could also
lose contact with one or more masters during the transaction and
leave one or more masters out-of-sync without the master(s) knowing
they are out of sync.
- When many clients are present, we do not want a flakey network
link from one to cause one or more masters to go out of
synchronization and potentially stall the whole works.
- Normal hammer2 mounts allow a virtually unlimited number of modifying
transactions between actual flushes. The media flush rolls everything
up into a single transaction id per flush. Detection of 'missing'
transactions in a concurrent multi-client setup when one or more client
temporarily loses connectivity is thus difficult.
- Clients have a limited amount of time to reconnect to a cluster after
a network disconnect before their MESI cache states are lost.
- Clients may proceed with several transactions before knowing for sure
that earlier transactions were completely successful. Performance is
important, we won't be waiting for a full quorum-verified synchronous
flush to media before allowing a system call to return.
- Masters can decide that a client's MESI cache states were lost (i.e.
that the transaction was too slow) as well.
The solutions (for modifying transactions):
- Masters handle quorum confirmation amongst themselves and do not rely
on the client for that purpose.
- A client can connect to one or more masters regardless of the size of
the quorum and can submit modifying operations to a single master if
desired. The master will take care of the rest.
A client must still validate the quorum (and obtain MESI cache states)
when doing read-only operations in order to present the correct data
to the user process for the VOP.
- Masters will run a 2-phase commit amongst themselves, often concurrent
with other non-conflicting transactions, and will serialize operations
and/or enforce synchronization points for 2-phase completion on
serialized transactions from the same client or when cache state
ownership is shifted from one client to another.
- Clients will usually allow operations to run asynchronously and return
from system calls more or less ASAP once they own the necessary cache
coherency locks. The client can select the validation mode to wait for
with mount options:
(1) Fully async (mount -o async)
(2) Wait for phase-1 ack (mount)
(3) Wait for phase-2 ack (mount -o sync) (fsync - wait p2ack)
(4) Wait for flush (mount -o sync) (fsync - wait flush)
Modifying system calls cannot be told to wait for a full media
flush, as full media flushes are prohibitively expensive. You
still have to fsync().
The fsync wait mode for network links can be selected, either to
return after the phase-2 ack or to return after the media flush.
The default is to wait for the phase-2 ack, which at least guarantees
that a network failure after that point will not disrupt operations
issued before the fsync.
- Clients must adjust the chain state for modifying operations prior to
releasing chain locks / returning from the system call, even if the
masters have not finished the transaction. A late failure by the
cluster will result in desynchronized state which requires erroring
out the whole filesystem or resynchronizing somehow.
- Clients can opt to keep a record of transactions through the phase-2
ack or the actual media flush on the masters.
However, replaying/revalidating the log cannot necessarily guarantee
success. If the masters lose synchronization due to network issues
between masters (or if the client was mounted fully-async), or if enough
masters crash simultaniously such that a quorum fails to flush even
after the phase-2 ack, then it is possible that by the time a client
is able to replay/revalidate, some other client has squeeded in and
committed something that would conflict.
If the client crashes it works similarly to a crash with a local storage
mount... many dirty buffers might be lost. And the same happens in
the cluster case.
TRANSACTION LOG
Keeping a short-term transaction log, much less being able to properly replay
it, is fraught with difficulty and I've made it a separate development task.
For now HAMMER2 does not have one.