[Top][All Lists]

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

Re: [qemu-s390x] [PATCH v2 2/5] vfio-ccw: concurrent I/O handling

From: Eric Farman
Subject: Re: [qemu-s390x] [PATCH v2 2/5] vfio-ccw: concurrent I/O handling
Date: Fri, 25 Jan 2019 15:21:55 -0500
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.9.1

On 01/25/2019 07:58 AM, Halil Pasic wrote:
On Thu, 24 Jan 2019 21:25:10 -0500
Eric Farman <address@hidden> wrote:

        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;

Ah, I see Halil's difficulty here.

It is true there is a race condition today, and that this doesn't
address it.  That's fine, add it to the todo list.  But even with that,
I don't see what the mutex is enforcing?

It is protecting the io regions. AFAIU the idea was that only one
thread is accessing the io region(s) at a time to prevent corruption and
reading half-morphed data.

Two simultaneous SSCHs will be
serialized (one will get kicked out with a failed trylock() call), while
still leaving the window open between cc=0 on the SSCH and the
subsequent interrupt.  In the latter case, a second SSCH will come
through here, do the copy_from_user below, and then jump to fsm_io_busy
to return EAGAIN.  Do we really want to stomp on io_region in that case?

I'm not sure I understood you correctly. The interrupt handler does not
take the lock before writing to the io_region. That is one race but it is
easy to fix.

The bigger problem is that between the interrupt handler has written IRB
area and userspace has read it we may end up destroying it by stomping on
it (to use your words). The userspace reading a wrong (given todays qemu
zeroed out) IRB could lead to follow on problems.

I wasn't thinking about a race between the start and interrupt handler, but rather between two near-simultaneous starts. Looking at it more closely, the orb and scsw structs as well as the ret_code field in ccw_io_region are only referenced under the protection of the new mutex (within fsm_io_request, for example), which I guess is the point.

So that leaves us with just the irb fields, which you'd mentioned a couple days ago (and which I was trying to ignore since it'd seems to have been discussed enough at the time). So I withdraw my concerns on this point. For now. ;-)

   Why can't we simply return EAGAIN if state==BUSY?

Sure we can. That would essentially go back to the old way of things:
if not idle return with error.

I think this happens both before and after this series. With this series, we just update the io_region with things that are never used because we're busy.

Just the error code returned would change
form EACCESS to EAGAIN. Which Isn't necessarily a win, because
conceptually here should be never two interleaved io_requests/start
commands hitting the module.

region = private->io_region;
-       if (copy_from_user((void *)region + *ppos, buf, count))
-               return -EFAULT;
+       if (copy_from_user((void *)region + *ppos, buf, count)) {
+               ret = -EFAULT;
+               goto out_unlock;
+       }

reply via email to

[Prev in Thread] Current Thread [Next in Thread]