Re: [Qemu-devel] [PATCH v2 2/5] vfio-ccw: concurrent I/O handling

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, 22 Jan 2019 13:46:12 +0100
Halil Pasic <pasic@xxxxxxxxxxxxx> wrote:

> On Tue, 22 Jan 2019 12:53:22 +0100
> Cornelia Huck <cohuck@xxxxxxxxxx> wrote:
> 
> > On Tue, 22 Jan 2019 12:17:37 +0100
> > Halil Pasic <pasic@xxxxxxxxxxxxx> wrote:
> >   
> > > On Tue, 22 Jan 2019 11:29:26 +0100
> > > Cornelia Huck <cohuck@xxxxxxxxxx> wrote:
> > >   
> > > > On Mon, 21 Jan 2019 21:20:18 +0100
> > > > Halil Pasic <pasic@xxxxxxxxxxxxx> wrote:
> > > >     
> > > > > On Mon, 21 Jan 2019 12:03:51 +0100
> > > > > Cornelia Huck <cohuck@xxxxxxxxxx> wrote:
> > > > >     
> > > > > > Rework handling of multiple I/O requests to return -EAGAIN if
> > > > > > we are already processing an I/O request. Introduce a mutex
> > > > > > to disallow concurrent writes to the I/O region.
> > > > > > 
> > > > > > The expectation is that userspace simply retries the operation
> > > > > > if it gets -EAGAIN.
> > > > > > 
> > > > > > We currently don't allow multiple ssch requests at the same
> > > > > > time, as we don't have support for keeping channel programs
> > > > > > around for more than one request.
> > > > > > 
> > > > > > Signed-off-by: Cornelia Huck <cohuck@xxxxxxxxxx>
> > > > > > ---      
> > > > > 
> > > > > [..]
> > > > >     
> > > > > >  static ssize_t vfio_ccw_mdev_write(struct mdev_device *mdev,
> > > > > > @@ -188,25 +192,30 @@ static ssize_t vfio_ccw_mdev_write(struct mdev_device *mdev,
> > > > > >  {
> > > > > >  	struct vfio_ccw_private *private;
> > > > > >  	struct ccw_io_region *region;
> > > > > > +	int ret;
> > > > > >  
> > > > > >  	if (*ppos + count > sizeof(*region))
> > > > > >  		return -EINVAL;
> > > > > >  
> > > > > >  	private = dev_get_drvdata(mdev_parent_dev(mdev));
> > > > > > -	if (private->state != VFIO_CCW_STATE_IDLE)
> > > > > > +	if (private->state == VFIO_CCW_STATE_NOT_OPER ||
> > > > > > +	    private->state == VFIO_CCW_STATE_STANDBY)
> > > > > >  		return -EACCES;
> > > > > > +	if (!mutex_trylock(&private->io_mutex))
> > > > > > +		return -EAGAIN;
> > > > > >  
> > > > > >  	region = private->io_region;
> > > > > > -	if (copy_from_user((void *)region + *ppos, buf, count))
> > > > > > -		return -EFAULT;
> > > > > > +	if (copy_from_user((void *)region + *ppos, buf, count)) {      
> > > > > 
> > > > > This might race with vfio_ccw_sch_io_todo() on
> > > > > private->io_region->irb_area, or?    
> > > > 
> > > > Ah yes, this should also take the mutex (should work because we're on a
> > > > workqueue).
> > > >     
> > > 
> > > I'm not sure that will do the trick (assumed I understood the
> > > intention correctly). Let's say the things happen in this order:
> > > 1) vfio_ccw_sch_io_todo() goes first, I guess updates
> > > private->io_region->irb_area and releases the mutex.
> > > 2) Then vfio_ccw_mdev_write() destroys the irb_area by zeriong it out,
> > > and finally,
> > > 3) userspace reads the destroyed irb_area using vfio_ccw_mdev_read().
> > > 
> > > Or am I misunderstanding something?   
> > 
> > You're not, but dealing with that race is outside the scope of this
> > patch. If userspace submits a request and then tries to get the old
> > data for a prior request, I suggest that userspace needs to fix their
> > sequencing.
> >   
> 
> I tend to disagree, because I think this would be a degradation compared
> to what we have right now.
> 
> Let me explain. I guess the current idea is that the private->state !=
> VFIO_CCW_STATE_IDLE check safeguards against this. Yes we lack proper
> synchronization (atomic/interlocked access or locks) that would guarantee
> that different thread observe state transitions as required -- no
> splint brain. But if state were atomic the scenario I have in mind can
> not happen, because we get the solicited interrupt in BUSY state (and
> set IDLE in vfio_ccw_sch_io_todo()). 

This BUSY handling is broken for another case: If the guest requests
intermediate interrupts, there may be more than one interrupt by the
hardware -- and we still go out of BUSY state. (Freeing the cp is also
broken in that case.) However, the Linux dasd driver does not seem to
do that.

> Unsolicited interrupts are another
> piece of cake -- I have no idea how may of those do we get.

They at least don't have the "free the cp before we got final state"
bug. But I think both are reasons to get away from "use the BUSY state
to ensure the right sequence".

> And because
> of this the broken sequencing in userspace could actually be the kernels
> fault.

Here, I can't follow you at all :(



[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Index of Archives]     [Kernel Development]     [Kernel Newbies]     [IDE]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite Info]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux ATA RAID]     [Samba]     [Linux Media]     [Device Mapper]

  Powered by Linux