OK, it's an incompatibility between Posix world and Plan 9 world. It's
because mkdir doesn't open the directory, and
create on Plan 9 does.
Here is what I think is the relevant code from applylog.
if(rd.mode&DMDIR){
fd = create(local, OREAD, DMDIR); <=========
creates d--------- --
but it's open.
if(fd < 0 && isdir(local))
fd = open(local, OREAD);
if(fd < 0){
error("mkdir %q: %r", name);
skip = 1;
continue;
}
This sequence works fine on Plan 9.
In devfs-posix.c we have this:
if(mkdir(path, perm & 0777) < 0)
oserror();
if((fd = open(path, 0)) < 0)
oserror();
Which fails very badly on linux if perm is 0, as it is from replica.
This works on Plan 9 because create is one operation, even for
directories. Chalk it up to another Posix botch. I can't recall how
many times this one has bitten me.
I can special-case the code in devfs-posix.c:
if (perm&0400 == 0) then do a mkdir, an open, and an fchmod. Tomorrow,
that is. I think the work-around is pretty easy.
I figure the 0400 might work because mode has to be OREAD at this point anyway.
or, simpler: take this code:
// Be like Plan 9 file servers: inherit mode bits
// and group from parent.
fchmod(fd, perm & st.st_mode & 0777);
which happens AFTER the mkdir, and instead compute the perms first and
do the mkdir with those perms.
Russ, any reason to make the d--------- perm directory and then change
it with the fchmod later? Can we just do the mkdir with the final
permissions? Or is this d--------- guarding against some race
condition in the linux vfs?
ron
