Until ESXi 6.5 VMware used the vmfsSparse format for snapshots (VMDK3 in
QEMU).
This format was lacking in the following:
* Grain directory (L1) and grain table (L2) entries were 32-bit,
allowing access to only 2TB (slightly less) of data.
* The grain size (default) was 512 bytes - leading to data
fragmentation and many grain tables.
* For space reclamation purposes, it was necessary to find all the
grains which are not pointed to by any grain table - so a reverse
mapping of "offset of grain in vmdk" to "grain table" must be
constructed - which takes large amounts of CPU/RAM.
The format specification can be found in VMware's documentation:
https://www.vmware.com/support/developer/vddk/vmdk_50_technote.pdf
In ESXi 6.5, to support snapshot files larger than 2TB, a new format was
introduced: SESparse (Space Efficient).
This format fixes the above issues:
* All entries are now 64-bit.
* The grain size (default) is 4KB.
* Grain directory and grain tables are now located at the beginning
of the file.
+ seSparse format reserves space for all grain tables.
+ Grain tables can be addressed using an index.
+ Grains are located in the end of the file and can also be
addressed with an index.
- seSparse vmdks of large disks (64TB) have huge preallocated
headers - mainly due to L2 tables, even for empty snapshots.
* The header contains a reverse mapping ("backmap") of "offset of
grain in vmdk" to "grain table" and a bitmap ("free bitmap") which
specifies for each grain - whether it is allocated or not.
Using these data structures we can implement space reclamation
efficiently.
* Due to the fact that the header now maintains two mappings:
* The regular one (grain directory & grain tables)
* A reverse one (backmap and free bitmap)
These data structures can lose consistency upon crash and result
in a corrupted VMDK.
Therefore, a journal is also added to the VMDK and is replayed
when the VMware reopens the file after a crash.
Since ESXi 6.7 - SESparse is the only snapshot format available.
Unfortunately, VMware does not provide documentation regarding the new
seSparse format.
This commit is based on black-box research of the seSparse format.
Various in-guest block operations and their effect on the snapshot file
were tested.
The only VMware provided source of information (regarding the underlying
implementation) was a log file on the ESXi:
/var/log/hostd.log
Whenever an seSparse snapshot is created - the log is being populated
with seSparse records.
Relevant log records are of the form:
[...] Const Header:
[...] constMagic = 0xcafebabe
[...] version = 2.1
[...] capacity = 204800
[...] grainSize = 8
[...] grainTableSize = 64
[...] flags = 0
[...] Extents:
[...] Header : <1 : 1>
[...] JournalHdr : <2 : 2>
[...] Journal : <2048 : 2048>
[...] GrainDirectory : <4096 : 2048>
[...] GrainTables : <6144 : 2048>
[...] FreeBitmap : <8192 : 2048>
[...] BackMap : <10240 : 2048>
[...] Grain : <12288 : 204800>
[...] Volatile Header:
[...] volatileMagic = 0xcafecafe
[...] FreeGTNumber = 0
[...] nextTxnSeqNumber = 0
[...] replayJournal = 0
The sizes that are seen in the log file are in sectors.
Extents are of the following format: <offset : size>
This commit is a strict implementation which enforces:
* magics
* version number 2.1
* grain size of 8 sectors (4KB)
* grain table size of 64 sectors
* zero flags
* extent locations
Additionally, this commit proivdes only a subset of the functionality
offered by seSparse's format:
* Read-only
* No journal replay
* No space reclamation
* No unmap support
Hence, journal header, journal, free bitmap and backmap extents are
unused, only the "classic" (L1 -> L2 -> data) grain access is
implemented.
However there are several differences in the grain access itself.
Grain directory (L1):
* Grain directory entries are indexes (not offsets) to grain
tables.
* Valid grain directory entries have their highest nibble set to
0x1.
* Since grain tables are always located in the beginning of the
file - the index can fit into 32 bits - so we can use its low
part if it's valid.
Grain table (L2):
* Grain table entries are indexes (not offsets) to grains.
* If the highest nibble of the entry is:
0x0:
The grain in not allocated.
The rest of the bytes are 0.
0x1:
The grain is unmapped - guest sees a zero grain.
The rest of the bits point to the previously mapped grain,
see 0x3 case.
0x2:
The grain is zero.
0x3:
The grain is allocated - to get the index calculate:
((entry & 0x0fff000000000000) >> 48) |
((entry & 0x0000ffffffffffff) << 12)
* The difference between 0x1 and 0x2 is that 0x1 is an unallocated
grain which results from the guest using sg_unmap to unmap the
grain - but the grain itself still exists in the grain extent - a
space reclamation procedure should delete it.
Unmapping a zero grain has no effect (0x2 will not change to 0x1)
but unmapping an unallocated grain will (0x0 to 0x1) - naturally.
In order to implement seSparse some fields had to be changed to support
both 32-bit and 64-bit entry sizes.
Reviewed-by: Karl Heubaum <karl.heubaum@oracle.com>
Reviewed-by: Eyal Moscovici <eyal.moscovici@oracle.com>
Reviewed-by: Arbel Moshe <arbel.moshe@oracle.com>
Signed-off-by: Sam Eiderman <shmuel.eiderman@oracle.com>
Message-id: 20190620091057.47441-4-shmuel.eiderman@oracle.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
98eb9733f4