On 2/7/22 20:38, Halil Pasic wrote:
On Mon, 7 Feb 2022 14:39:31 -0500
Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:
Back to the topic of locking: it looks to me that on this path you
do the filtering and thus the accesses to matrix_mdev->shadow_apcb,
matrix_mdev->matrix and matrix_dev->config_info some of which are
of type write whithout the matrix_dev->lock held. More precisely
only with the matrix_dev->guests_lock held in "read" mode.
Did I misread the code? If not, how is that OK?
You make a valid point, a struct rw_semaphore is not adequate for the
purposes
it is used in this patch series. It needs to be a mutex.
Good we agree that v17 is racy.
For v18 which is forthcoming probably this week, I've been reworking the
locking
based on your observation that the struct ap_guest is not necessary given we
already have a list of the mediated devices which contain the KVM
pointer. On the other
[..]
BTW I got delayed on my "locking rules" writeup. Sorry for that!
No worries, I've been writing up a vfio-ap-locking.rst document to
include with the next
version of the patch series.
I'm looking forward to v18 including that document. I prefer not to
discuss what you wrote about the approach taken in v18 now. It is easier
to me when I have both the text stating the intended design, and the
code that is supposed to adhere to this design.
Regards,
Halil
Coming soon to a theater near you:)
