On Fri, Nov 30, 2018 at 3:03 PM Olga Kornievskaia <olga.kornievskaia@xxxxxxxxx> wrote: > > This patch series adds support for NFSv4.2 copy offload feature > allowing copy between two different NFS servers. > > This functionality depends on the VFS ability to support generic > copy_file_range() where a copy is done between an NFS file and > a local file system. > > This feature is enabled by the kernel module parameter -- > inter_copy_offload_enable -- and by default is disabled. There is > also a kernel compile configuration of NFSD_V4_2_INTER_SSC that > adds dependency on the NFS client side functions called from the > server. > > These patches work on top of existing async intra copy offload > patches. For the "inter" SSC, the implementation only supports > asynchronous inter copy. > > On the source server, upon receiving a COPY_NOTIFY, it generate a > unique stateid that's kept in the global list. Upon receiving a READ > with a stateid, the code checks the normal list of open stateid and > now additionally, it'll check the copy state list as well before > deciding to either fail with BAD_STATEID or find one that matches. > The stored stateid is only valid to be used for the first time > with a choosen lease period (90s currently). When the source server > received an OFFLOAD_CANCEL, it will remove the stateid from the > global list. Otherwise, the copy stateid is removed upon the removal > of its "parent" stateid (open/lock/delegation stateid). > > On the destination server, upon receiving a COPY request, the server > establishes the necessary clientid/session with the source server. > It calls into the NFS client code to establish the necessary > open stateid, filehandle, file description (without doing an NFS open). > Then the server calls into the copy_file_range() to preform the copy > where the source file will issue NFS READs and then do local file > system writes (this depends on the VFS ability to do cross device > copy_file_range(). > > v2: > -- in on top of 4.20-rc4 + client side inter patch series > -- VFS changes to do enable generic copy_file_range() and then NFS > falls back on generic_copy_file_range() for previous EXDEV/OPNOTSUPP > errors > -- hopefully addressed Bruce's review comments (highlights are): > --- copy_notify patch: addressed naming, sc_cp_list access is > now protected by s2s_cp_lock > --- fillin netloc4 patch: address the size and added WARN_ON > --- add ca_source to COPY: decode only 1 address, dont allocate > memory (the rest into dummy) > --- check stateid against stored: moved the refcount under lock > --- allow stale filehandle: adding a loop to go thru the ops in > the compound, store/manage puttfh if copy is present in the compound > mark the source putfh as "no verify". > > All the patches (client inter) and this patch series is available > from git://linux-nfs.org/projects/aglo/linux.git under the "linux-ssc" > branch > Bruce, Do you have comments on this v2? Once VFS has the patches for the generic copy_file_range() functionality, NFS should be all set to just used it. > Olga Kornievskaia (10): > VFS generic copy_file_range() support > NFS fallback to generic_copy_file_range > NFSD fill-in netloc4 structure > NFSD add ca_source_server<> to COPY > NFSD return nfs4_stid in nfs4_preprocess_stateid_op > NFSD add COPY_NOTIFY operation > NFSD check stateids against copy stateids > NFSD generalize nfsd4_compound_state flag names > NFSD: allow inter server COPY to have a STALE source server fh > NFSD add nfs4 inter ssc to nfsd4_copy > > fs/nfs/nfs4file.c | 9 +- > fs/nfsd/Kconfig | 10 ++ > fs/nfsd/nfs4proc.c | 406 ++++++++++++++++++++++++++++++++++++++++++++++----- > fs/nfsd/nfs4state.c | 124 ++++++++++++++-- > fs/nfsd/nfs4xdr.c | 166 ++++++++++++++++++++- > fs/nfsd/nfsd.h | 32 ++++ > fs/nfsd/nfsfh.h | 5 +- > fs/nfsd/nfssvc.c | 6 + > fs/nfsd/state.h | 21 ++- > fs/nfsd/xdr4.h | 37 ++++- > fs/read_write.c | 66 +++++++-- > include/linux/fs.h | 7 + > include/linux/nfs4.h | 1 + > mm/filemap.c | 6 +- > 14 files changed, 810 insertions(+), 86 deletions(-) > > -- > 1.8.3.1 >
