> On Jun 15, 2022, at 11:28 AM, Wang Yugui <wangyugui@xxxxxxxxxxxx> wrote: > > Hi, > >>> On Jun 12, 2022, at 3:22 AM, Wang Yugui <wangyugui@xxxxxxxxxxxx> wrote: >>> >>> NFSv4 need to close a file completely (no lingering open) when it does >>> a CLOSE or DELEGRETURN. >>> >>> When multiple NFSv4/OPEN from different clients, we need to check the >>> reference count. The flowing reference-count-check change the behavior >>> of NFSv3 nfsd_rename()/nfsd_unlink() too. >>> >>> Link: https://bugzilla.linux-nfs.org/show_bug.cgi?id=387 >>> Signed-off-by: Wang Yugui <wangyugui@xxxxxxxxxxxx> >>> --- >>> TO-CHECK: >>> 1) NFSv3 nfsd_rename()/nfsd_unlink() feature change is OK? >>> 2) Can we do better performance than nfsd_file_close_inode_sync()? >>> 3) nfsd_file_close_inode_sync()->nfsd_file_close_inode() in nfsd4_delegreturn() >>> => 'Text file busy' about 4s >>> 4) reference-count-check : refcount_read(&nf->nf_ref) <= 1 or ==0? >>> nfsd_file_alloc() refcount_set(&nf->nf_ref, 1); >>> >>> fs/nfsd/filecache.c | 2 +- >>> fs/nfsd/nfs4state.c | 4 ++++ >>> 2 files changed, 5 insertions(+), 1 deletion(-) >> >> I suppose I owe you (and Frank) a progress report on #386. I've fixed >> the LRU algorithm and added some observability features to measure >> how the fix impacts the cache's efficiency for NFSv3 workloads. >> >> These new features show that the hit rate and average age of cache >> items goes down after the fix is applied. I'm trying to understand >> if I've done something wrong or if the fix is supposed to do that. >> >> To handle the case of hundreds of thousands of open files more >> efficiently, I'd like to convert the filecache to use rhashtable. > > A question about the comming rhashtable. > > Now multiple nfsd export share a cache pool. > > In the coming rhashtable, a nfsd export could use a private cache pool > to improve scale out? That seems like a premature optimization. We don't know that the hashtable, under normal (ie, non-generic/531) workloads, is a scaling problem. However, I am considering (in the future) creating separate filecaches for each nfsd_net. -- Chuck Lever
