On 04/08/16 16:16, David Vrabel wrote: > Currently libxencall using mlocked buffers for hypercall buffers. > This pages are subject to compaction and page migration. A userspace > process may see a hypercall fail with -EFAULT if a page backing a > hypercall buffer is in the process of being migrated. > > Page migration can be prevented by taking an additional reference to > the page. > > There are three possible ways to do this: > > 1. Add a new device which when mmap()'d populated the VMA with > suitable memory that has an additional reference. This would give > VMAs that have appropriate properties set (e.g., VM_DONTCOPY) > without having to do additional madvise() calls. > > 2. Add a new ioctl to privcmd to populate a VMA with suitably > ref-counted pages. However, mmap() on privcmd is already used for > foreign mappings and the VMA properties for hypercall buffers and > foreign mappings might need to be different and the handling of > unmap will need to be different. This might be tricky. > > 3. Add a pair of ioctls to privcmd to lock/unlock a buffer by > getting/putting an additional reference to the page. This is > what's implemented in this series.
This method appears to be insufficient -- I'm still seeing very occasional, spurious -EFAULT errors. The easiest workaround is to set vm.compact_unevictable_allowed = 0. David _______________________________________________ Xen-devel mailing list Xen-devel@lists.xen.org https://lists.xen.org/xen-devel
