On Thu, Jul 23, 2020 at 09:37:40AM -0400, Colin Walters wrote:
On Tue, Jul 21, 2020, at 11:58 AM, Stefano Garzarella wrote:
quoted
my use case concerns virtualization. The idea, that I described in the
proposal of io-uring restrictions [1], is to share io_uring CQ and SQ queues
with a guest VM for block operations.
Virtualization being a strong security barrier is in eternal conflict
with maximizing performance. All of these "let's add a special
guest/host channel" are high risk areas.
And this effort in particular - is it *really* worth it to expose a
brand new, fast moving Linux kernel interface (that probably hasn't
been fuzzed as much as it needs to be) to virtual machines?
It is an experiment to explore the potential of io_uring. In addition
the restrictions can also be useful for other use case, for example if
a process wants to allow another process to use io_uring, but only allowing
a subset of operations.
People who want maximum performance at the cost of a bit of security
already have the choice to use Linux containers, where they can use
io_uring natively.
Thanks,
Stefano