[Gluster-devel] directory filehandles

[Emmanuel Dreyfus]

Hello

I have trouble figuring the whole story about how to cope with FUSE
directory filehandles in the NetBSD implementation.

libfuse makes a special use of filehandles exposed to filesystem for
OPENDIR, READDIR, FSYNCDIR, and RELEASEDIR. For that four operations,
the fh is a pointer to a struct fuse_dh, in which the fh field is
exposed to the filesystem. All other filesystem operations pass the fh
as is from kernel to filesystem back and forth.

That means that a fh obtained by OPENDIR should never be passed to
operations others than (READDIR, FSYNCDIR and RELEASEDIR). For instance,
when porting ltfs to NetBSD, I experienced that passing a fh obtained
from OPENDIR to SETATTR would crash.

glusterfs implementation differs from libfuse because it seems the
filesystem is always passed as is: there is nothing like libfuse struct
fuse_dh. It will therefore happily accept fh obtained by OPENDIR for any
operation, something that I do not expect to happen in libfuse based
filesystems.

My real concern is SETLK on directory. Here glusterfs really wants a fh
or it will report an error. The NetBSD implementation passes the fh it
got from OPENDIR, but I expect a libfuse based filesystem to crash in
such a situation. For now I did not find any libfuse-based filesystem
that implements locking, so I could not test that. 

Could someone clarify this? What are the FUSE operations that should be
sent to filesystem on that kind of program?

        int fd;

        /* NetBSD  calls FUSE LOOKUP and OPENDIR */
        if ((fd = open("/gfs/tmp", O_RDONLY, 0)) == -1)
                err(1, "open failed");

        /* NetBSD calls FUSE SETLKW */
        if (flock(fd, LOCK_EX) == -1)
                err(1, "flock failed");



-- 
Emmanuel Dreyfus
http://hcpnet.free.fr/pubz
manu at netbsd.org