| Message ID | 166689625728.90991.15067635142973595248.stgit@klimt.1015granger.net (mailing list archive) |
|---|---|
| Headers | show |
| Series | Series short description | expand |
On Fri, 28 Oct 2022, Chuck Lever wrote: > I'm proposing this series for v6.2 (for-next). > > For quite some time, we've been encouraged to disable filecache > garbage collection for NFSv4 files, and I think I found a surgical > way to do just that. That is presented in "NFSD: Add an NFSD_FILE_GC > flag to enable nfsd_file garbage collection". > > The other major change in this set is reworking the file_hashtbl to > resize itself dynamically. This reduces the average size of its > bucket chains, greatly speeding up hash insertion, which holds the > state_lock. > > This version seems to pass thread-intensive testing so far. > > Comments and opinions are welcome. All looks good to me - with the understanding that refcount fixes will follow. I cannot comment on the tracepoint changes as I'm not particularly familiar with that code, but for everything else Reviewed-by: NeilBrown <neilb@suse.de> Thanks, NeilBrown > > Changes since v5: > - Wrap hash insertion with inode->i_lock > - Replace hashfn and friends with in-built rhashtable functions > - Add a tracepoint to report delegation return > > Changes since v4: > - Addressed some comments in the GC patch; more to come > - Split clean-ups out of the rhashtable patch, reordered the series > - Removed state_lock from the rhashtable helpers > > Changes since v3: > - the new filehandle alias check was still not right > > Changes since v2: > - Converted nfs4_file_rhashtbl to nfs4_file_rhltable > - Addressed most or all other review comments > > Changes since RFC: > - checking nfs4_files for inode aliases is now done only on hash > insertion > - the nfs4_file reference count is now bumped only while the RCU > read lock is held > - comments and function names have been revised and clarified > > --- > > Chuck Lever (14): > NFSD: Pass the target nfsd_file to nfsd_commit() > NFSD: Revert "NFSD: NFSv4 CLOSE should release an nfsd_file immediately" > NFSD: Add an NFSD_FILE_GC flag to enable nfsd_file garbage collection > NFSD: Clean up nfs4_preprocess_stateid_op() call sites > NFSD: Trace stateids returned via DELEGRETURN > NFSD: Trace delegation revocations > NFSD: Use const pointers as parameters to fh_ helpers > NFSD: Update file_hashtbl() helpers > NFSD: Clean up nfsd4_init_file() > NFSD: Add a remove_nfs4_file() helper > NFSD: Clean up find_or_add_file() > NFSD: Refactor find_file() > NFSD: Allocate an rhashtable for nfs4_file objects > NFSD: Use rhashtable for managing nfs4_file objects > > > fs/nfsd/filecache.c | 81 +++++++++++++++------- > fs/nfsd/filecache.h | 4 +- > fs/nfsd/nfs3proc.c | 10 ++- > fs/nfsd/nfs4proc.c | 42 +++++------ > fs/nfsd/nfs4state.c | 165 ++++++++++++++++++++++++-------------------- > fs/nfsd/nfsfh.h | 10 +-- > fs/nfsd/state.h | 5 +- > fs/nfsd/trace.h | 59 +++++++++++++++- > fs/nfsd/vfs.c | 19 ++--- > fs/nfsd/vfs.h | 3 +- > 10 files changed, 249 insertions(+), 149 deletions(-) > > -- > Chuck Lever > >
I'm proposing this series for v6.2 (for-next). For quite some time, we've been encouraged to disable filecache garbage collection for NFSv4 files, and I think I found a surgical way to do just that. That is presented in "NFSD: Add an NFSD_FILE_GC flag to enable nfsd_file garbage collection". The other major change in this set is reworking the file_hashtbl to resize itself dynamically. This reduces the average size of its bucket chains, greatly speeding up hash insertion, which holds the state_lock. This version seems to pass thread-intensive testing so far. Comments and opinions are welcome. Changes since v5: - Wrap hash insertion with inode->i_lock - Replace hashfn and friends with in-built rhashtable functions - Add a tracepoint to report delegation return Changes since v4: - Addressed some comments in the GC patch; more to come - Split clean-ups out of the rhashtable patch, reordered the series - Removed state_lock from the rhashtable helpers Changes since v3: - the new filehandle alias check was still not right Changes since v2: - Converted nfs4_file_rhashtbl to nfs4_file_rhltable - Addressed most or all other review comments Changes since RFC: - checking nfs4_files for inode aliases is now done only on hash insertion - the nfs4_file reference count is now bumped only while the RCU read lock is held - comments and function names have been revised and clarified --- Chuck Lever (14): NFSD: Pass the target nfsd_file to nfsd_commit() NFSD: Revert "NFSD: NFSv4 CLOSE should release an nfsd_file immediately" NFSD: Add an NFSD_FILE_GC flag to enable nfsd_file garbage collection NFSD: Clean up nfs4_preprocess_stateid_op() call sites NFSD: Trace stateids returned via DELEGRETURN NFSD: Trace delegation revocations NFSD: Use const pointers as parameters to fh_ helpers NFSD: Update file_hashtbl() helpers NFSD: Clean up nfsd4_init_file() NFSD: Add a remove_nfs4_file() helper NFSD: Clean up find_or_add_file() NFSD: Refactor find_file() NFSD: Allocate an rhashtable for nfs4_file objects NFSD: Use rhashtable for managing nfs4_file objects fs/nfsd/filecache.c | 81 +++++++++++++++------- fs/nfsd/filecache.h | 4 +- fs/nfsd/nfs3proc.c | 10 ++- fs/nfsd/nfs4proc.c | 42 +++++------ fs/nfsd/nfs4state.c | 165 ++++++++++++++++++++++++-------------------- fs/nfsd/nfsfh.h | 10 +-- fs/nfsd/state.h | 5 +- fs/nfsd/trace.h | 59 +++++++++++++++- fs/nfsd/vfs.c | 19 ++--- fs/nfsd/vfs.h | 3 +- 10 files changed, 249 insertions(+), 149 deletions(-) -- Chuck Lever