On Sun, Jan 16, 2022 at 03:44:21PM +0300, Vasily Averin wrote: > On 03.01.2022 22:53, J. Bruce Fields wrote: > > On Wed, Dec 29, 2021 at 11:24:43AM +0300, Vasily Averin wrote: > >> nfsd and lockd use F_SETLK cmd with the FL_SLEEP flag set to request > >> asynchronous processing of blocking locks. > >> > >> Currently nfs4 use locks_lock_inode_wait() function which is blocked > >> for such requests. To handle them correctly FL_SLEEP flag should be > >> temporarily reset before executing the locks_lock_inode_wait() function. > >> > >> Additionally block flag is forced to set, to translate blocking lock to > >> remote nfs server, expecting it supports async processing of the blocking > >> locks too. > > > > But this on its own isn't enough for the client to support asynchronous > > blocking locks, right? Don't we also need the logic that calls knfsd's > > lm_notify when it gets a CB_NOTIFY_LOCK from the server? > > No, I think this should be enough. > We are here a nfs client, > we can get F_SETLK with FL_SLEEP from nfsd only (i.e. in re-export case) > we need to avoid blocking if lock is already taken, > so we need to call locks_lock_inode_wait without FL_SLEEP, > then we submit _sleeping_ request to NFS server (i.e. set )data->arg.block = 1) > and waiting for reply from server. > > Here we rely that server will NOT block on such request too, so our reply wel not be blocked too. Just on that one point: if there's a lock conflict, an NFSv4 server will return NFS4ERR_DENIED immediately and leave it to the client to poll. Or if you're using NFS version >= 4.1, the server has the option of calling back to the client with a CB_NOTIFY_LOCK to let the client know when the lock might be available. (See https://datatracker.ietf.org/doc/html/rfc8881#section-20.11 for details.) But if a server that blocked and didn't reply to the original LOCK request until the lock became available, that would be a bug. (Apologies for responding just to that one point, I'm also trying to get caught back up again here....). --b. > Under "block" I mean that handler can sleep or process request for a very long time > but it will NOT BE BLOCKED if lock is taken already, it WILL NOT WAIT when lock will be released, > it just return some error in this case. > > I think it is correct. > Do you think I am wrong or maybe I missed something? > > Thank you, > Vasily Averin > > However I noticed now that past is incorrect, > temporally dropped FL_SLEEP should be restored back in _nfs4_proc_setlk before _nfs4_do_setlk() call. > I'll fix it in next version of this patch-set. > > >> https://bugzilla.kernel.org/show_bug.cgi?id=215383 > >> Signed-off-by: Vasily Averin <vvs@xxxxxxxxxxxxx> > >> --- > >> fs/nfs/nfs4proc.c | 5 ++++- > >> 1 file changed, 4 insertions(+), 1 deletion(-) > >> > >> diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c > >> index ee3bc79f6ca3..9b1380c4223c 100644 > >> --- a/fs/nfs/nfs4proc.c > >> +++ b/fs/nfs/nfs4proc.c > >> @@ -7094,7 +7094,7 @@ static int _nfs4_do_setlk(struct nfs4_state *state, int cmd, struct file_lock *f > >> recovery_type == NFS_LOCK_NEW ? GFP_KERNEL : GFP_NOFS); > >> if (data == NULL) > >> return -ENOMEM; > >> - if (IS_SETLKW(cmd)) > >> + if (IS_SETLKW(cmd) || (fl->fl_flags & FL_SLEEP)) > >> data->arg.block = 1; > >> nfs4_init_sequence(&data->arg.seq_args, &data->res.seq_res, 1, > >> recovery_type > NFS_LOCK_NEW); > >> @@ -7200,6 +7200,9 @@ static int _nfs4_proc_setlk(struct nfs4_state *state, int cmd, struct file_lock > >> int status; > >> > >> request->fl_flags |= FL_ACCESS; > >> + if (((fl_flags & FL_SLEEP_POSIX) == FL_SLEEP_POSIX) && IS_SETLK(cmd)) > >> + request->fl_flags &= ~FL_SLEEP; > >> + > >> status = locks_lock_inode_wait(state->inode, request); > >> if (status < 0) > >> goto out; > >> -- > >> 2.25.1
