On 08 Apr 2015, at 15:03, Ted Unangst <[email protected]> wrote:
> Also, this only helps if you're sure that the code reading the flag will do so
> in an smp safe way. In many cases, the reading code will also need to acquire
> a lock in order to correctly do something after reading the flag. From the
> diff context, it looks like most of this code will definitely already have
> some other lock.
What do you think about f_offset protection? Lock file struct object within
of_read or fo_write routine?
For example for vn_read()
int
vn_read(struct file *fp, off_t *poff, struct uio *uio, struct ucred *cred)
{
struct vnode *vp = (struct vnode *)fp->f_data;
int error = 0;
size_t count = uio->uio_resid;
struct proc *p = uio->uio_procp;
FILE_LOCK(fp);
/* no wrap around of offsets except on character devices */
if (vp->v_type != VCHR && count > LLONG_MAX - *poff) {
FILE_UNLOCK(fp);
return (EINVAL);
}
vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p);
uio->uio_offset = *poff;
if (vp->v_type != VDIR)
error = VOP_READ(vp, uio,
(fp->f_flag & FNONBLOCK) ? IO_NDELAY : 0, cred);
*poff += count - uio->uio_resid;
VOP_UNLOCK(vp, 0, p);
FILE_UNLOCK(fp);
return (error);
}
Is it a good idea?
