[sheepdog] [PATCH v8 00/19] object reclaim based on generational reference counting
Hitoshi Mitake
mitake.hitoshi at gmail.com
Thu May 15 17:22:27 CEST 2014
The object reclaim doesn't support hypervolume yet. But hypervolume cannot be
used as a virtual disk (both of qemu and tgt don't support it) currently. And
the removal of old vdi deletion is acceptable for hypervolume because it doesn't
support snapshot, etc. So I think this patchset can be applied to the master
branch.
The same code is pushed to:
https://github.com/sheepdog/sheepdog/tree/snapshot-object-reclaim
There is a problem which can be caused by discard operation. But the
problem can be solved as an individual topic. I'll post a patchset for
it later.
The leak problem was (also) caused by bugs in QEMU's sheepdog
driver. The fixed version of QEMU driver is here:
https://github.com/sheepdog/qemu/tree/inode-sync
I'll post it to the QEMU list later.
v8:
- let COW and snapshot be excluded mutually
-- This change introduces limitation that "dog vdi snapshot" must be
executed on a same node which execute QEMU. But it is temporal and
removed easily.
- migrate VDI deletion from sheep to dog
- pass Valerio's test
v7:
- reorder invalid tests group
- rebase on master
- correct invalid inheritence of reference count
-- add a new test for checking the case
- refine test output
v6:
- recycle old child_vdi_id and move btree_counter to the place
- update CHANGELOG.md
v5:
- rebase on master
- update store version and break compatibility explicitly
- remove coding style problems
-- don't use pthread_mutex_t directly
-- rename inode.gref -> inode.gref
- resolve regression in http feature (10th patch)
- forbid creating snapshot of hypervolumes temporally
- enhance tests (14th patch)
- remove various problems exposed by the patchset (1st - 3rd patches)
v4:
- remove some bugs
- refine tests
-- the patchset can pass every test in vdi group
v3:
- rebase onto master
v2:
- rebase onto master
- add test to check object reclaim feature
- rename *_obj() to *_object()
- add documentation about the algorithm
Currently, sheepdog doesn't reclaim objects until all the snapshots
are deleted. For example:
$ dog vdi create image
$ dog vdi write image < some_data
$ dog vdi snapshot image -s snap1
$ dog vdi write image < some_data
$ dog vdi delete image <- this doesn't reclaim the objects
of the image
$ dog vdi delete image -s snap1 <- this reclaims all the data objects
of both image and image:snap1
To resolve this problem we need to introduce something like reference
counting mechanism. However, a naive approach doesn't work well on
Sheepdog. It is because, when we create a snapshot or clone vdi, we
have to increment all the reference count of its data objects and send
many requests to all over the nodes.
This series uses generation reference counting [1], which is more
appropriate algorithm for distributed system like Sheepdog. With this
series, we can reclaim objects just when they are no longer necessary.
[1] B. Goldberg, Generation reference counting: A reduced
communication distributed storage reclamation scheme, PLDI '89
==
Here is a brief explanation ob the object reclaim algorithm. If you
want to know the details or correctness proof of the generational
reference counting, please refer the original paper:
Each data object contains the following structure in its inode object:
struct generation_reference {
int32_t generation;
int32_t count;
} gref;
The generation field identifies the generation of the reference and
the count field records the number of references copied from this
particular reference.
We also introduce a new type of object, a ledger object. The ledger
object exists for each data object and contains an array of int32_t
values. The i-th element of the array contains information about the
number of outstanding data object references in the i-th generation.
Generational reference counting is performed as follows:
1. When a vdi A is created, the initial data object references are
initialized as follows:
for (i = 0; i < MAX_DATA_OBJS; i++) {
A.gref[i].generation = 0;
A.gref[i].count = 0;
}
and a ledger object for each data object is initialized as follows:
ledger_object[0] = 1;
for (i = 1; i < MAX_DATA_OBJS; i++) {
ledger_object[i] = 0;
}
In practice, the ledger object is created and initialized when the
object is accessed for the first time, so we don't have to create
all the ledger objects when the vdi is created.
2. When a VDI B is created as a clone (or snapshot) of A, the new
reference B is initialized as follows:
for (i = 0; i < MAX_DATA_OBJS; i++) {
B.gref[i].generation = A.gref[i].generation + 1;
B.gref[i].count = 0;
}
In addition, A.gref.count's are incremented:
for (i = 0; i < MAX_DATA_OBJS; i++) {
A.gref[i].count++;
}
3. When a object o is removed, a decrement message is sent to its
ledger object containing the values of o.generation and o.count.
4. When a decrement message to o is received, in which the generation
field has the value i and the count field has the value c, the
following actions are performed:
o.ledger[i]--;
o.ledger[i + 1] += c;
When every element of the ledger becomes zero (for all i >= 0,
o.ledger[i] = 0), o can be reclaimed.
Note that some elements of the ledger may hold negative values.
Hitoshi Mitake (19):
sheep: introduce generational reference counting for object reclaim
sheep: introduce ledger objects
sheep: decrement generational reference count on vdi deletion
sheep: decrement generational reference count on copy-on-write
sheep: remove max children limit
sheep: introduce sparse objects
sheepdog proto: recycle old child_vdi_id for storing btree_counter
sheep: get_request() for deferred response
sheep, dog: prevent COW during snapshot creation
sheep: reduce needless remove request during object reclamation
sheep, dog: migrate deletion process from sheep to dog
dog: print status of generational reference of inodes in dump-inode
sheep: update store format version to v5
tests: show ledger objects in list
doc: add documentation of object reclaim algorithm
tests/functional: add test for object reclaim
update CHANGELOG.md for new object reclaim scheme
tests/functional: add a new test of the object reclaim for cloned VDIs
tests/functional: new test of object reclaim for a case of
continuously snapshotted VDI
CHANGELOG.md | 3 +
doc/object-reclaim.txt | 73 +++
dog/vdi.c | 99 +++-
include/internal_proto.h | 4 +
include/sheepdog_proto.h | 59 ++-
include/util.h | 3 +
lib/util.c | 62 +++
sheep/config.c | 2 +-
sheep/gateway.c | 99 +++-
sheep/group.c | 3 +
sheep/migrate.c | 16 +
sheep/ops.c | 169 +++++++
sheep/plain_store.c | 83 +++-
sheep/request.c | 42 +-
sheep/sheep_priv.h | 40 ++
sheep/store.c | 26 ++
sheep/vdi.c | 387 ++++------------
tests/functional/016.out | 14 +-
tests/functional/042 | 4 +-
tests/functional/042.out | 44 +-
tests/functional/044.out | 14 +-
tests/functional/057.out | 140 +++---
tests/functional/058.out | 14 +-
tests/functional/063.out | 28 +-
tests/functional/064.out | 38 +-
tests/functional/086 | 36 +-
tests/functional/086.out | 1068 ++------------------------------------------
tests/functional/092 | 47 ++
tests/functional/092.out | 42 ++
tests/functional/093 | 32 ++
tests/functional/093.out | 26 ++
tests/functional/common.rc | 12 +-
tests/functional/group | 4 +-
33 files changed, 1206 insertions(+), 1527 deletions(-)
create mode 100644 doc/object-reclaim.txt
create mode 100755 tests/functional/092
create mode 100644 tests/functional/092.out
create mode 100755 tests/functional/093
create mode 100644 tests/functional/093.out
--
1.9.1
More information about the sheepdog
mailing list