On Wed, 14 Nov 2018 21:59:08 -0800 A Z <a...@zegelin.com> wrote: > This is an issue that involves a combination of different software > packages, so my apologies in advance if this is the wrong list to post on. > > I'm experiencing terrible boot times when I assign a large amount of RAM to > a VM when used in combination with VIFO/PCI-passthrough. > > On a VM with a Nvidia GTX 970 + USB controller and 24GiB of RAM assigned, > the time to the TianoCore splash screen is ~5 minutes. It's then ~30 > seconds before Windows 10 begins to boot (spinning dots). During this time, > the QEMU CPU core threads are 100% busy. > > According to `perf`, the QMU CPU core threads are spending most of their > time waiting on a spinlock over kvm->mmu_lock that's created by > kvm_zap_gfn_range. > > I'm fairly certain that ~1 year ago (if not longer) the same configuration > didn't take this long to boot.
If you can confirm that your system uses a NUMA memory configuration and that 24G exceeds the node local memory, then this is a known regression in the Linux kernel mm. AIUI it's being worked upstream, workarounds are to either use hugepages or locked memory for the VM. Thanks, Alex _______________________________________________ vfio-users mailing list vfio-users@redhat.com https://www.redhat.com/mailman/listinfo/vfio-users
