Re: [PATCH] counter: drop chrdev_lock

[William Breathitt Gray <vilhelm.gray@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 5865 bytes --]

On Wed, Oct 20, 2021 at 05:59:32AM +0900, William Breathitt Gray wrote:
> On Tue, Oct 19, 2021 at 09:44:04AM -0500, David Lechner wrote:
> > On 10/19/21 2:18 AM, William Breathitt Gray wrote:
> > > On Tue, Oct 19, 2021 at 09:07:48AM +0200, Greg KH wrote:
> > >> On Tue, Oct 19, 2021 at 03:53:08PM +0900, William Breathitt Gray wrote:
> > >>> On Mon, Oct 18, 2021 at 11:03:49AM -0500, David Lechner wrote:
> > >>>> On 10/18/21 4:14 AM, William Breathitt Gray wrote:
> > >>>>> On Sun, Oct 17, 2021 at 01:55:21PM -0500, David Lechner wrote:
> > >>>>>> diff --git a/drivers/counter/counter-sysfs.c b/drivers/counter/counter-sysfs.c
> > >>>>>> index 1ccd771da25f..7bf8882ff54d 100644
> > >>>>>> --- a/drivers/counter/counter-sysfs.c
> > >>>>>> +++ b/drivers/counter/counter-sysfs.c
> > >>>>>> @@ -796,25 +796,18 @@ static int counter_events_queue_size_write(struct counter_device *counter,
> > >>>>>>    					   u64 val)
> > >>>>>>    {
> > >>>>>>    	DECLARE_KFIFO_PTR(events, struct counter_event);
> > >>>>>> -	int err = 0;
> > >>>>>> -
> > >>>>>> -	/* Ensure chrdev is not opened more than 1 at a time */
> > >>>>>> -	if (!atomic_add_unless(&counter->chrdev_lock, 1, 1))
> > >>>>>> -		return -EBUSY;
> > >>>>>> +	int err;
> > >>>>>>    
> > >>>>>>    	/* Allocate new events queue */
> > >>>>>>    	err = kfifo_alloc(&events, val, GFP_KERNEL);
> > >>>>>>    	if (err)
> > >>>>>> -		goto exit_early;
> > >>>>>> +		return err;
> > >>>>>>    
> > >>>>>>    	/* Swap in new events queue */
> > >>>>>>    	kfifo_free(&counter->events);
> > >>>>>>    	counter->events.kfifo = events.kfifo;
> > >>>>>
> > >>>>> Do we need to hold the events_lock mutex here for this swap in case
> > >>>>> counter_chrdev_read() is in the middle of reading the kfifo to
> > >>>>> userspace, or do the kfifo macros already protect us from a race
> > >>>>> condition here?
> > >>>>>
> > >>>> Another possibility might be to disallow changing the size while
> > >>>> events are enabled. Otherwise, we also need to protect against
> > >>>> write after free.
> > >>>>
> > >>>> I considered this:
> > >>>>
> > >>>> 	swap(counter->events.kfifo, events.kfifo);
> > >>>> 	kfifo_free(&events);
> > >>>>
> > >>>> But I'm not sure that would be safe enough.
> > >>>
> > >>> I think it depends on whether it's safe to call kfifo_free() while other
> > >>> kfifo_*() calls are executing. I suspect it is not safe because I don't
> > >>> think kfifo_free() waits until all kfifo read/write operations are
> > >>> finished before freeing -- but if I'm wrong here please let me know.
> > >>>
> > >>> Because of that, will need to hold the counter->events_lock afterall so
> > >>> that we don't modify the events fifo while a kfifo read/write is going
> > >>> on, lest we suffer an address fault. This can happen regardless of
> > >>> whether you swap before or after the kfifo_free() because the old fifo
> > >>> address could still be in use within those uncompleted kfifo_*() calls
> > >>> if they were called before the swap but don't complete before the
> > >>> kfifo_free().
> > >>>
> > >>> So we have a problem now that I think you have already noticed: the
> > >>> kfifo_in() call in counter_push_events() also needs protection, but it's
> > >>> executing within an interrupt context so we can't try to lock a mutex
> > >>> lest we end up sleeping.
> > >>>
> > >>> One option we have is as you suggested: we disallow changing size while
> > >>> events are enabled. However, that will require us to keep track of when
> > >>> events are disabled and implement a spinlock to ensure that we don't
> > >>> disable events in the middle of a kfifo_in().
> > >>>
> > >>> Alternatively, we could change events_lock to a spinlock and use it to
> > >>> protect all these operations on the counter->events fifo. Would this
> > >>> alternative be a better option so that we avoid creating another
> > >>> separate lock?
> > >>
> > >> I would recommend just having a single lock here if at all possible,
> > >> until you determine that there a performance problem that can be
> > >> measured that would require it to be split up.
> > >>
> > >> thanks,
> > >>
> > >> greg k-h
> > > 
> > > All right let's go with a single events_lock spinlock then. David, if
> > > you make those changes and submit a v2, I'll be okay with this patch and
> > > can provide my ack for it.
> > > 
> > 
> > We can't use a spin lock for everything since there are operations
> > that can sleep that need to be in the critical sections. Likewise,
> > we can't use a mutex for everything since some critical sections
> > are in interrupt handlers. But, I suppose we can try combining
> > the existing mutexes. Since the kfifo is accessed from both
> > contexts, it seems like it still needs more consideration than
> > just a mutex or a spin lock, e.g. if events are enabled, don't
> > allow swapping out the kfifo buffer.
> 
> I think there is a deadlock case if we combine the ops_exists_lock with
> the n_events_list_lock, so this will need further thought. However, at
> the very least the swap occuring in counter_events_queue_size_write()
> and the kfifo_in() in counter_push_events() require some sort of
> locking; it is trivial to cause a page fault with the code in its
> current state.
> 
> I think this can be fixed if just events_lock is changed to spinlock for
> now without modifying the other locks. We can try to combine the
> remaining locks in a subsequent patch, if they are capable of being
> combined.
> 
> William Breathitt Gray

After considering this further, kfifo_to_user() could possibly sleep so
we can't use a spinlock here afterall. As such, events_lock should
remain as a mutex and instead we'll only allow swapping out the kfifo
buffer when events are disabled.

William Breathitt Gray

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]