Gregory Farnum <gfarnum@xxxxxxxxxx> writes: > On Tue, Aug 21, 2018 at 2:40 AM, Yan, Zheng <ukernel@xxxxxxxxx> wrote: >> On Tue, Aug 21, 2018 at 5:21 PM Luis Henriques <lhenriques@xxxxxxxx> wrote: >>> >>> "Yan, Zheng" <ukernel@xxxxxxxxx> writes: >>> ... >>> >> + ret = ceph_get_caps(dst_ci, CEPH_CAP_FILE_WR, >>> >> + CEPH_CAP_FILE_BUFFER, >>> >> + endoff, &got, NULL); >>> >> + if (ret < 0) >>> >> + goto out; >>> >> + >>> > >>> > Getting FILE_RD caps for src inode? I think adding another >>> > ceph_get_caps() call here may cause deadlock (shouldn't hold any cap >>> > while waiting). >>> >>> I thought about that initially, but since I'm using do_splice_direct for >>> doing any manual copies I assumed that this would eventually be handled >>> somewhere else. Because we don't really FILE_RD if we're doing only >>> full (remote) object copies, do we? >>> >> >> We still need to get FILE_RD for full object copies. Because other >> client may have dirty data in page cache. we need to until the dirty >> data get flushed > > Plus it's not like the OSD has any idea about the filesystem's caps; > the client needs to hold them all on the OSD's behalf while it's doing > remote data copies. Yep, caps knowledge is on the MDS side. Makes sense, and I'm currently looking at adding the FILE_RD capability to the source file (still trying to find out about Zheng's deadlock warning). Cheers, -- Luis
