librbd: do write_full for whole object write

[wonzhq]

This is useful when ec is the base tier in the cache tiering setup. The
write_full can be proxied from the cache tier to the base tier. So that
we don't need to promote the object.

Signed-off-by: Zhiqiang Wang zhiqiang.wang@intel.com
Suggested-by: Nick Fisknick@fisk.me.uk

[wonzhq]

@fiskn here comes the implementation of what we discussed in the ceph-user mail list yesterday :-)

[fiskn]

Wow, that's quick progress. Couple of questions if its ok?

1. m_object_off==0 in the IF statement, this checks that the write is aligned to the start of the object?
2. m_object_len == m_ictx->get_object_size(). This checks that the write IO is equal to the object size. I assume this means if the IO is bigger than the object, then it will do a normal write instead of a write_full? I wonder if its possible to split an IO into several write_full requests to handle the case where an IO > Obj_size.

[loic-bot]

SUCCESS: http://jenkins.ceph.dachary.org/job/ceph/7408/

• SUCCESS http://jenkins.ceph.dachary.org/job/ceph/LABELS=centos-7&&x86_64/7408/
• SUCCESS http://jenkins.ceph.dachary.org/job/ceph/LABELS=ubuntu-14.04&&x86_64/7408/

[ghost]

@wonzhq could you run part of the rbd suite to show it passes ? You would have to chose at least one job that walks that code path and link the pulpito result as a comment to this pull request.

[ghost]

@wonzhq you will need to wait for http://tracker.ceph.com/issues/12903 to be fixed before running a successfull test.

[wonzhq]

@fiskn

m_object_off==0 in the IF statement, this checks that the write is aligned to the start of the object?

Yes.

m_object_len == m_ictx->get_object_size(). This checks that the write IO is equal to the object size. I assume this means if the IO is bigger than the object, then it will do a normal write instead of a write_full? I wonder if its possible to split an IO into several write_full requests to handle the case where an IO > Obj_size.

Actually the IO has been split when reaching here. This is done by file_to_extents in aio_write.

[wonzhq]

@dachary i proposed #5776 to fix 12903.

[ghost]

@wonzhq wonzhq-rbd-write-full is building in http://ceph.com/gitbuilder.cgi

[fiskn]

@wonzhq Thanks for the response, all makes sense, really looking forward to giving this a test.

[idryomov]

I'll do the same for the kernel client once this is merged.

[dillaman]

LGTM once tested

[fiskn]

Should this also be considered for the CephFS clients?

[wonzhq]

Yes, we should be able to do the same thing for cephfs and rgw. Please let me get familiar with the cephfs and rgw code first.

[jdurgin]

lgtm too. rgw already uses write_full when possible I think

[tchaikov]

tested at

• http://pulpito.ceph.com/kchai-2015-09-11_21:01:10-rados-wip-kefu-testing---basic-multi/
• http://pulpito.ceph.com/kchai-2015-09-11_07:49:14-rados-wip-kefu-testing---basic-multi/
• http://pulpito.ceph.com/kchai-2015-09-09_11:54:39-rados-wip-kefu-testing---basic-multi/

[tchaikov]

@jdurgin shall we run some rbd tests for it also?

[jdurgin]

@tchaikov yes, a rados run doesn't execute this code. I'll schedule an rbd run with this and maybe a few other branches