Re: [PATCH v5 1/3] FUSE: Avoid lookups in fuse create

[Vivek Goyal <vgoyal@redhat.com>]

On Wed, May 18, 2022 at 01:41:02PM -0400, Vivek Goyal wrote:
> On Tue, May 17, 2022 at 03:37:42PM +0530, Dharmendra Singh wrote:
> 
> [..]
> > diff --git a/include/uapi/linux/fuse.h b/include/uapi/linux/fuse.h
> > index d6ccee961891..bebe4be3f1cb 100644
> > --- a/include/uapi/linux/fuse.h
> > +++ b/include/uapi/linux/fuse.h
> > @@ -301,6 +301,7 @@ struct fuse_file_lock {
> >   * FOPEN_CACHE_DIR: allow caching this directory
> >   * FOPEN_STREAM: the file is stream-like (no file position at all)
> >   * FOPEN_NOFLUSH: don't flush data cache on close (unless FUSE_WRITEBACK_CACHE)
> > + * FOPEN_FILE_CREATED: the file was actually created
> >   */
> >  #define FOPEN_DIRECT_IO		(1 << 0)
> >  #define FOPEN_KEEP_CACHE	(1 << 1)
> > @@ -308,6 +309,7 @@ struct fuse_file_lock {
> >  #define FOPEN_CACHE_DIR		(1 << 3)
> >  #define FOPEN_STREAM		(1 << 4)
> >  #define FOPEN_NOFLUSH		(1 << 5)
> > +#define FOPEN_FILE_CREATED	(1 << 6)
> >  
> >  /**
> >   * INIT request/reply flags
> > @@ -537,6 +539,7 @@ enum fuse_opcode {
> >  	FUSE_SETUPMAPPING	= 48,
> >  	FUSE_REMOVEMAPPING	= 49,
> >  	FUSE_SYNCFS		= 50,
> > +	FUSE_CREATE_EXT		= 51,
> 
> I am wondering if we really have to introduce a new opcode for this. Both
> FUSE_CREATE and FUSE_CREATE_EXT prepare and send fuse_create_in{} and
> expect fuse_entry_out and fuse_open_out in response. So no new structures
> are being added. Only thing FUSE_CREATE_EXT does extra is that it also
> reports back whether file was actually created or not.
> 
> May be instead of adding an new fuse_opcode, we could simply add a
> new flag which we send in fuse_create_in and that reqeusts to report
> if file was created or not. This is along the lines of
> FUSE_OPEN_KILL_SUIDGID.
> 
> So say, a new flag FUSE_OPEN_REPORT_CREATE flag. Which we will set in
> fuse_create_in->open_flags. If file server sees this flag is set, it
> knows that it needs to set FOPEN_FILE_CREATED flag in response.
> 
> To me creating a new flag FUSE_OPEN_REPORT_CREATE seems better instead
> of adding a new opcode.

Actually I take that back. If we were to use a flag, then we will have to
do feature negotiation in advance at init time and only then we can set
FUSE_OPEN_REPORT_CREATE. But we are relying on no new feature bit instead
-ENOSYS will be returned if server does not support FUSE_CREATE_EXT.
So adding a new opcode is better.

Thanks
Vivek