On 2012-08-14 17:53, Myklebust, Trond wrote: > On Tue, 2012-08-14 at 22:30 +0800, Peng Tao wrote: >> On Tue, Aug 14, 2012 at 9:45 PM, Myklebust, Trond >> <Trond.Myklebust@xxxxxxxxxx> wrote: >>> On Tue, 2012-08-14 at 10:48 +0300, Benny Halevy wrote: >>>> On 2012-08-09 18:39, Myklebust, Trond wrote: >>>>> On Thu, 2012-08-09 at 23:01 +0800, Peng Tao wrote: >>>>>> On Thu, Aug 9, 2012 at 10:36 PM, Myklebust, Trond >>>>>> <Trond.Myklebust@xxxxxxxxxx> wrote: >>>>>>> On Thu, 2012-08-09 at 22:30 +0800, Peng Tao wrote: >>>>>>>> On Thu, Aug 9, 2012 at 4:21 AM, Trond Myklebust >>>>>>>> <Trond.Myklebust@xxxxxxxxxx> wrote: >>>>>>>>> Ever since commit 0a57cdac3f (NFSv4.1 send layoutreturn to fence >>>>>>>>> disconnected data server) we've been sending layoutreturn calls >>>>>>>>> while there is potentially still outstanding I/O to the data >>>>>>>>> servers. The reason we do this is to avoid races between replayed >>>>>>>>> writes to the MDS and the original writes to the DS. >>>>>>>>> >>>>>>>>> When this happens, the BUG_ON() in nfs4_layoutreturn_done can >>>>>>>>> be triggered because it assumes that we would never call >>>>>>>>> layoutreturn without knowing that all I/O to the DS is >>>>>>>>> finished. The fix is to remove the BUG_ON() now that the >>>>>>>>> assumptions behind the test are obsolete. >>>>>>>>> >>>>>>>> Isn't MDS supposed to recall the layout if races are possible between >>>>>>>> outstanding write-to-DS and write-through-MDS? >>>>>>> >>>>>>> Where do you read that in RFC5661? >>>>>>> >>>>>> That's my (maybe mis-)understanding of how server works... But looking >>>>>> at rfc5661 section 18.44.3. layoutreturn implementation. >>>>>> " >>>>>> After this call, >>>>>> the client MUST NOT use the returned layout(s) and the associated >>>>>> storage protocol to access the file data. >>>>>> " >>>>>> And given commit 0a57cdac3f, client is using the layout even after >>>>>> layoutreturn, which IMHO is a violation of rfc5661. >>>>> >>>>> No. It is using the layoutreturn to tell the MDS to fence off I/O to a >>>>> data server that is not responding. It isn't attempting to use the >>>>> layout after the layoutreturn: the whole point is that we are attempting >>>>> write-through-MDS after the attempt to write through the DS timed out. >>>>> >>>> >>>> I hear you, but this use case is valid after a time out / disconnect >>>> (which will translate to PNFS_OSD_ERR_UNREACHABLE for the objects layout) >>>> In other cases, I/Os to the DS might obviously be in flight and the BUG_ON >>>> indicates that. >>>> >>>> IMO, the right way to implement that is to initially mark the lsegs invalid >>>> and increment plh_block_lgets, as we do today in _pnfs_return_layout >>>> but actually send the layout return only when the last segment is dereferenced. >>> >>> This is what we do for object and block layout types, so your >>> objects-specific objection is unfounded. >>> >> object layout is also doing layout return on IO error (commit >> fe0fe83585f8). And it doesn't take care of draining concurrent >> in-flight IO. I guess that's why Boaz saw the same BUG_ON. > > Yes. I did notice that code when I was looking into this. However that's > Boaz's own patch, and it _only_ applies to the objects layout type. I > assumed that he had tested it when I applied it... > > One way to fix that would be to keep a count of "outstanding > read/writes" in the layout, so that when the error occurs, and we want > to fall back to MDS, we just increment plh_block_lgets, invalidate the > layout, and then let the outstanding read/writes fall to zero before > sending the layoutreturn. Sounds reasonable to me too. > If the objects layout wants to do that, then I have no objection. As > I've said multiple times, though, I'm not convinced we want to do that > for the files layout. > I just fear that removing the BUG_ON will prevent us from detecting cases where a LAYOUTRETURN is sent while there are layout segments in use in the error free or non-timeout case. Benny -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
