os/bluestore: get rid off blob's ref_map for non-shared objects #9988
Conversation
| @@ -196,6 +196,7 @@ struct bluestore_blob_t { | |||
| } | |||
|
|
|||
| vector<bluestore_pextent_t> extents;///< raw data position on device | |||
| uint32_t compressed_length_orig = 0;///< original length of compressed blob if any | |||
There was a problem hiding this comment.
We could just call this 'length'...
There was a problem hiding this comment.
I'd prefer to highlight that this member is applied to compressed blobs only. And we ensure its' persistence in this case only. 'length' hides such semantic IMHO.
|
This looks good to me! It needs a rebase, though. |
|
I'm not wild about having to calculate a ref map before we can figure out which refs we are discarding, but I don't see a better way. Perhaps in the future we can keep the (up to date) ref_map in the Blob so that we only generate it on the first put after the Onode is loaded off disk. (Can do that later though..) |
ifed01
force-pushed
the
wip-bluestore-reduce-blob4
branch
from
09c03e4
to
5d9480b
Compare
|
@liewegas: Fixed and rebased |
|
Hi Igor, does this pass "ceph_test_objectstore --gtest_filter=*/2"? |
Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
… track original length for compressed blobs Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
…ernal ref_map for recerence counting Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
…e_extent_ref_map_t Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
…e_onode_t interface to handle reference counting for both ref_map/no ref_map cases in a single place Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
…e_t::punch_hole and others Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
ifed01
force-pushed
the
wip-bluestore-reduce-blob4
branch
from
5d9480b
to
53af175
Compare
|
@markhpc, just rerun the object store test suite - everything is OK |
|
Adding comment to see if I can get jenkins to rerun |
|
Retest this please |
|
Hi Igor, it looks like this is failing one of the unit tests: The following tests FAILED: Can you please fix and update? |
…nt ref_map locations Signed-off-by: Igor Fedotov <ifedotov@mirantis.com>
|
@markhpc - fixed |
|
While testing this I hit a segfault during ceph_test_objectstore, but on subsequent runs I'm not able to reproduce. This PR may not be responsible for the segfault, but I'm going to try to see if I can track down what's going on before merging. |
|
Igor mentioned that he has hit this segfault in the past as well and it seems to be related to running out of memory during the test. Given that it has been observed before this PR and that I have not been able to reproduce the segfault in question, I'm going to merge this as it is passing all tests now. |
Ref_map is built on the fly for onodes that are not shared. Hence provides saving for both serialized and in-memory onode representation.
Signed-off-by: Igor Fedotov ifedotov@mirantis.com