On Tue, 2020-04-14 at 17:08 -0700, Andrew Morton wrote:
> On Tue, 7 Apr 2020 17:51:07 +0800 Miles Chen <miles.chen@xxxxxxxxxxxx> wrote:
>
> > In fixup_user_fault(), it is possible that unlocked is NULL,
> > so we should test unlocked before using it.
> >
> > For example, in arch/arc/kernel/process.c, NULL is passed
> > to fixup_user_fault().
> >
> > SYSCALL_DEFINE3(arc_usr_cmpxchg, int *, uaddr, int, expected, int, new)
> > {
> > ...
> > ret = fixup_user_fault(current, current->mm, (unsigned long) uaddr,
> > FAULT_FLAG_WRITE, NULL);
> > ...
> > }
>
> (cc Peter)
>
> > --- a/mm/gup.c
> > +++ b/mm/gup.c
> > @@ -1230,7 +1230,8 @@ int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
> > if (ret & VM_FAULT_RETRY) {
> > down_read(&mm->mmap_sem);
> > if (!(fault_flags & FAULT_FLAG_TRIED)) {
> > - *unlocked = true;
> > + if (unlocked)
> > + *unlocked = true;
> > fault_flags |= FAULT_FLAG_TRIED;
> > goto retry;
> > }
>
> Not sure. If the caller passes FAULT_FLAG_ALLOW_RETRY then they must
> also pass in a valid non-NULL `unlocked'. If the caller passed
> FAULT_FLAG_ALLOW_RETRY and unlocked==NULL then the resulting oops is an
> appropriate way of reporting this mistake. I think?
>
Agree. How about put "unlocked==NULL must not be used with
FAULT_FLAG_ALLOW_RETRY." in the comment? Make it easier to understand
the oops.
e.g.,
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -1176,7 +1176,8 @@ static bool vma_permits_fault(struct
vm_area_struct *vma,
* @address: user address
* @fault_flags:flags to pass down to handle_mm_fault()
* @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if
caller
- * does not allow retry
+ * does not allow retry. If NULL, the caller must guarantee
+ * the fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
