On Sat, Sep 28, 2019 at 09:40:45AM +0800, Peter Xu wrote: > Previously we've tried to unlimit ioeventfd creation (6ea34c9b78c1, > "kvm: exclude ioeventfd from counting kvm_io_range limit", > 2013-06-04), because that can be easily done by fd limitations and > otherwise it can easily reach the current maximum of 1000 iodevices. > Meanwhile, we still use the counter to limit the maximum allowed kvm > io devices to be created besides ioeventfd. > > 6ea34c9b78c1 achieved that in most cases, however it'll still fali the > ioeventfd creation when non-ioeventfd io devices overflows to 1000. > Then the next ioeventfd creation will fail while logically it should > be the next non-ioeventfd iodevice creation to fail. > > That's not really a big problem at all because when it happens it > probably means something has leaked in userspace (or even malicious > program) so it's a bug to fix there. However the error message like > "ioeventfd creation failed" with an -ENOSPACE is really confusing and > may let people think about the fact that it's the ioeventfd that is > leaked (while in most cases it's not!). > > Let's use this patch to unlimit the creation of ioeventfd for real > this time, assuming this is also a bugfix of 6ea34c9b78c1. To me more > importantly, when with a bug in userspace this patch can probably give > us another more meaningful failure on what has overflowed/leaked > rather than "ioeventfd creation failure: -ENOSPC". > > CC: Dr. David Alan Gilbert <dgilbert@xxxxxxxxxx> > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx> Ping - just in case it fell through the cracks. -- Peter Xu
