Re: BKL removal

[Matthew Wilcox]

On Sun, Jul 07, 2002 at 10:58:04PM -0400, Alexander Viro wrote:
> On Mon, 8 Jul 2002, Matthew Wilcox wrote:
> > one struct file per open(), yes.  however, fork() shares a struct file,
> > as does unix domain fd passing.  so we need protection between different
> > processes.  there's some pretty good reasons to want to use a semaphore
> > to protect the struct file (see fasync code.. bleugh).
> 
> ??? What exactly do you want to protect there?

andrea & i discussed this off-list a few days ago... see fs/fcntl.c

                case F_SETFL:
                        lock_kernel();
                        err = setfl(fd, filp, arg);
                        unlock_kernel();

setfl() does:

        if ((arg ^ filp->f_flags) & FASYNC) {
                if (filp->f_op && filp->f_op->fasync) {
                        error = filp->f_op->fasync(fd, filp, (arg & FASYNC) != 0);

and:

        if (arg & O_DIRECT) {
                down(&inode->i_sem);
                if (!filp->f_iobuf)
                        error = alloc_kiovec(1, &filp->f_iobuf);
                up(&inode->i_sem);

and finally:

        filp->f_flags = (arg & SETFL_MASK) | (filp->f_flags & ~SETFL_MASK);

i pointed out that if alloc_kiovec slept, the BKL provides no protection
against someone else doing a setfl at the same time, so we can get the
wrong number of fasync events sent.  Marcus Alanen pointed out that
fasync can also sleep, so we're at risk anyway.  i don't think that
abusing i_sem as andrea did is the Right Thing to do...

-- 
Revolutions do not require corporate support.