|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [PATCH] xen/pvcalls: fix potential endless loop in pvcalls-front.c
On Mon, 13 Nov 2017, Juergen Gross wrote:
> On 11/11/17 00:57, Stefano Stabellini wrote:
> > On Tue, 7 Nov 2017, Juergen Gross wrote:
> >> On 06/11/17 23:17, Stefano Stabellini wrote:
> >>> mutex_trylock() returns 1 if you take the lock and 0 if not. Assume you
> >>> take in_mutex on the first try, but you can't take out_mutex. Next times
> >>> you call mutex_trylock() in_mutex is going to fail. It's an endless
> >>> loop.
> >>>
> >>> Solve the problem by moving the two mutex_trylock calls to two separate
> >>> loops.
> >>>
> >>> Reported-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> >>> Signed-off-by: Stefano Stabellini <sstabellini@xxxxxxxxxx>
> >>> CC: boris.ostrovsky@xxxxxxxxxx
> >>> CC: jgross@xxxxxxxx
> >>> ---
> >>> drivers/xen/pvcalls-front.c | 5 +++--
> >>> 1 file changed, 3 insertions(+), 2 deletions(-)
> >>>
> >>> diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
> >>> index 0c1ec68..047dce7 100644
> >>> --- a/drivers/xen/pvcalls-front.c
> >>> +++ b/drivers/xen/pvcalls-front.c
> >>> @@ -1048,8 +1048,9 @@ int pvcalls_front_release(struct socket *sock)
> >>> * is set to NULL -- we only need to wait for the existing
> >>> * waiters to return.
> >>> */
> >>> - while (!mutex_trylock(&map->active.in_mutex) ||
> >>> - !mutex_trylock(&map->active.out_mutex))
> >>> + while (!mutex_trylock(&map->active.in_mutex))
> >>> + cpu_relax();
> >>> + while (!mutex_trylock(&map->active.out_mutex))
> >>> cpu_relax();
> >>
> >> Any reason you don't just use mutex_lock()?
> >
> > Hi Juergen, sorry for the late reply.
> >
> > Yes, you are right. Given the patch, it would be just the same to use
> > mutex_lock.
> >
> > This is where I realized that actually we have a problem: no matter if
> > we use mutex_lock or mutex_trylock, there are no guarantees that we'll
> > be the last to take the in/out_mutex. Other waiters could be still
> > outstanding.
> >
> > We solved the same problem using a refcount in pvcalls_front_remove. In
> > this case, I was thinking of reusing the mutex internal counter for
> > efficiency, instead of adding one more refcount.
> >
> > For using the mutex as a refcount, there is really no need to call
> > mutex_trylock or mutex_lock. I suggest checking on the mutex counter
> > directly:
> >
> >
> > while (atomic_long_read(&map->active.in_mutex.owner) != 0UL ||
> > atomic_long_read(&map->active.out_mutex.owner) != 0UL)
> > cpu_relax();
> >
> > Cheers,
> >
> > Stefano
> >
> >
> > ---
> >
> > xen/pvcalls: fix potential endless loop in pvcalls-front.c
> >
> > mutex_trylock() returns 1 if you take the lock and 0 if not. Assume you
> > take in_mutex on the first try, but you can't take out_mutex. Next time
> > you call mutex_trylock() in_mutex is going to fail. It's an endless
> > loop.
> >
> > Actually, we don't want to use mutex_trylock at all: we don't need to
> > take the mutex, we only need to wait until the last mutex waiter/holder
> > releases it.
> >
> > Instead of calling mutex_trylock or mutex_lock, just check on the mutex
> > refcount instead.
> >
> > Reported-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> > Signed-off-by: Stefano Stabellini <sstabellini@xxxxxxxxxx>
> > CC: boris.ostrovsky@xxxxxxxxxx
> > CC: jgross@xxxxxxxx
> >
> > diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c
> > index 0c1ec68..9f33cb8 100644
> > --- a/drivers/xen/pvcalls-front.c
> > +++ b/drivers/xen/pvcalls-front.c
> > @@ -1048,8 +1048,8 @@ int pvcalls_front_release(struct socket *sock)
> > * is set to NULL -- we only need to wait for the existing
> > * waiters to return.
> > */
> > - while (!mutex_trylock(&map->active.in_mutex) ||
> > - !mutex_trylock(&map->active.out_mutex))
> > + while (atomic_long_read(&map->active.in_mutex.owner) != 0UL ||
> > + atomic_long_read(&map->active.out_mutex.owner) != 0UL)
>
> I don't like this.
>
> Can't you use a kref here? Even if it looks like more overhead it is
> much cleaner. There will be no questions regarding possible races,
> while with an approach like yours will always smell racy (can't there
> be someone taking the mutex just after above test?).
>
> In no case you should make use of the mutex internals.
Boris' suggestion solves that problem well. Would you be OK with the
proposed
while(mutex_is_locked(&map->active.in_mutex.owner) ||
mutex_is_locked(&map->active.out_mutex.owner))
cpu_relax();
?
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |