Re: Detecting NUMA per pmem

Hi All We encountered a strange issue using pmem emulation. We configure emulated pmem devices in our system by adding these params to the boot command: memmap=32G!224G memmap=32G!480G . PMEM devices look OK. This issue is specific to 4.14 kernel AND centos 7.5, other combinations follow exactly the same flow but don't crash. Happens when ndctl converts an <emulated> /dev/pmem0 to /dev/dax0.0 command: ndctl create-namespace -f -e namespace0.0 --type=pmem --mode=dax This crashes the kernel in udev when trying to free page in do_munmap() syscall. And hangs. echo namespace0.0 > /sys/devices/platform/e820_pmem/ndbus0/region0/namespace0.0/driver/unbind Calltrace(): ndctl_unbind() ndctl_namespace_disable() ndctl_namespace_disable_invalidate() ndctl_namespace_disable_safe() namespace_destroy() namespace_reconfig() do_xaction_namespace() When we use for example ubuntu 16.04 with this kernel version it does not happen. When we use centos7.5 and kernel version 4.9 is also does not happen. When working with actual NVDIMMs this does not happen Did you encounter such an issue? Why does the kernel think that this area is mapped and who might be mapping it? If no one maps it why does the kernel has indication that these pages are mapped? Any help would be highly appreciated. Thanks Oren Berman On 10 January 2018 at 09:41, Oren Berman <oren(a)lightbitslabs.com&gt; wrote: