I pull a new request: https://github.com/ceph/ceph/pull/3413 In this PR, I use an additional buffer to avoid modifying other.op_bl when append, so msg can works fine with the new transaction now. But an additional memcpy is needed which may hurt the performance. I also revert the commit d427ca3 to fix http://tracker.ceph.com/issues/10534. Is there any better solution? On 20 January 2015 at 10:19, Dong Yuan <yuandong1222@xxxxxxxxx> wrote: > On 20 January 2015 at 09:43, Sage Weil <sweil@xxxxxxxxxx> wrote: >> When I first looked at the append() function I started working on a branch >> that would avoid the need for it entirely. ObjectStore already >> operates on a list of Transactions, and IIRC the appending is only needed >> for local operations, not for what is sent over the wire. It was >> nontrivial because of various annoying reasons (pointer ownership, >> lifecycle, etc.) so I didn't get very far. > > I totally agree with you. > > Actually I tried to avoid all append at first, but it seems too > complex to refact transaction and remove append in a single BP, so I > give up and focus on the transaction. > >> >> If this is going to be a significant amount of work to fix with the >> current approach it might make sense to bite the bullet and go down that >> path..? > > I want to spend sometime (maybe one or two days) on the current > approach first, if it doesn't make sense. I will go down that path. > Is that OK? > >> >> sage >> >> On Tue, 20 Jan 2015, Dong Yuan wrote: >> >>> If we must keep the order of local_t and op_t, I think the current >>> impl of append is not appropriate. >>> >>> Maybe I must provide something like clone_append to avoid the modify >>> of existing data in the op_buffer, so the msg should be happy with the >>> new transaction format. >>> >>> I will try this solution and give a patch as soon as possible. >>> >>> On 20 January 2015 at 02:44, Samuel Just <sam.just@xxxxxxxxxxx> wrote: >>> > We actually found another problem with this series as well. >>> > >>> > http://tracker.ceph.com/issues/10534 >>> > >>> > Looks like d427ca35404a30e1f428859c3274e030f2f83ef6 reversed the order >>> > of localt (which contains the create_collection) and op_t (which >>> > contains all of the operations on the object in the temp collection). >>> > We need the operations in localt to precede the operations in op_t. >>> > How do we do this? >>> > -Sam >>> > >>> > On Mon, Jan 19, 2015 at 8:07 AM, Dong Yuan <yuandong1222@xxxxxxxxx> wrote: >>> >> Yes, I am working on it since last week. >>> >> >>> >> This seems the same problem which I fix in the ReplicatedPG. So I try >>> >> to fix it following the same way which I did in the ReplicatedPG, but >>> >> it break the make_check test_erasure_code.sh. >>> >> >>> >> Now I am working on the test case, but unfortunately I am not familiar >>> >> with the impl of EC, so it will take me more time. >>> >> >>> >> On 19 January 2015 at 23:41, Sage Weil <sweil@xxxxxxxxxx> wrote: >>> >>> Hi, >>> >>> >>> >>> This has come up a couple times in QA since we merged the fast transaction >>> >>> code. Can you take a look? >>> >>> >>> >>> http://tracker.ceph.com/issues/10517 >>> >>> >>> >>> Ping me if it will help to get access to the test cluster. Haomai has >>> >>> access, and you can get the logs via http (linked from pulpito.ceph.com), >>> >>> but it is less convenient. >>> >>> >>> >>> Thanks! >>> >>> sage >>> >> >>> >> >>> >> >>> >> -- >>> >> Dong Yuan >>> >> Email:yuandong1222@xxxxxxxxx >>> >> -- >>> >> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in >>> >> the body of a message to majordomo@xxxxxxxxxxxxxxx >>> >> More majordomo info at http://vger.kernel.org/majordomo-info.html >>> >>> >>> >>> -- >>> Dong Yuan >>> Email:yuandong1222@xxxxxxxxx >>> >>> > > > > -- > Dong Yuan > Email:yuandong1222@xxxxxxxxx -- Dong Yuan Email:yuandong1222@xxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
