When we last left this debate [1] it was becoming clear that the 'page-less' approach left too many I/O scenarios off the table. The page-less enabling is still useful for avoiding the overhead of struct page where it is not needed, but in the end, page-backed persistent memory seems to be a requirement. With that assumption in place the next debate was where to allocate the storage for the memmap array, or otherwise reduce the overhead of 'struct page' with a fancier object like variable length pages. This series takes the position of mapping persistent memory with standard 'struct page' and pushes the policy decision of allocating the storage for the memmap array, from RAM or PMEM, to userspace. It turns out the best place to allocate 64-bytes per 4K page will be platform specific. If PMEM capacities are low then mapping in RAM is a good choice. Otherwise, for very large capacities storing the memmap in PMEM might be a better choice. Yet again, PMEM might not have the performance characteristics favorable to a high rate of change object like 'struct page'. The kernel can make a reasonable guess, but it seems we will need to maintain the ability to override any default. Outside of the new libvdimm sysfs mechanisms to specify the memmap allocation policy for a given PMEM device, the core of this implementation is 'struct vmem_altmap'. 'vmem_altmap' alters the memory hotplug code to optionally use a reserved PMEM-pfn range rather than dynamic allocation for the memmap. Only lightly tested so far to confirm valid pfn_to_page() and page_address() conversions across a range of persistent memory specified by 'memmap=ss!nn' (kernel command line option to simulate a PMEM range). [1]: https://lists.01.org/pipermail/linux-nvdimm/2015-May/000748.html --- Dan Williams (7): x86, mm: ZONE_DEVICE for "device memory" x86, mm: introduce struct vmem_altmap x86, mm: arch_add_dev_memory() mm: register_dev_memmap() libnvdimm, e820: make CONFIG_X86_PMEM_LEGACY a tristate option libnvdimm, pfn: 'struct page' provider infrastructure libnvdimm, pmem: 'struct page' for pmem arch/powerpc/mm/init_64.c | 7 + arch/x86/Kconfig | 19 ++ arch/x86/include/uapi/asm/e820.h | 2 arch/x86/kernel/Makefile | 2 arch/x86/kernel/pmem.c | 79 +-------- arch/x86/mm/init_64.c | 160 +++++++++++++----- drivers/nvdimm/Kconfig | 26 +++ drivers/nvdimm/Makefile | 5 + drivers/nvdimm/btt.c | 8 - drivers/nvdimm/btt_devs.c | 172 +------------------ drivers/nvdimm/claim.c | 201 ++++++++++++++++++++++ drivers/nvdimm/e820.c | 86 ++++++++++ drivers/nvdimm/namespace_devs.c | 34 +++- drivers/nvdimm/nd-core.h | 9 + drivers/nvdimm/nd.h | 59 ++++++- drivers/nvdimm/pfn.h | 35 ++++ drivers/nvdimm/pfn_devs.c | 334 +++++++++++++++++++++++++++++++++++++ drivers/nvdimm/pmem.c | 213 +++++++++++++++++++++++- drivers/nvdimm/region.c | 2 drivers/nvdimm/region_devs.c | 19 ++ include/linux/kmap_pfn.h | 33 ++++ include/linux/memory_hotplug.h | 21 ++ include/linux/mm.h | 53 ++++++ include/linux/mmzone.h | 23 +++ mm/kmap_pfn.c | 195 ++++++++++++++++++++++ mm/memory_hotplug.c | 84 ++++++--- mm/page_alloc.c | 18 ++ mm/sparse-vmemmap.c | 60 ++++++- mm/sparse.c | 44 +++-- tools/testing/nvdimm/Kbuild | 7 + tools/testing/nvdimm/test/iomap.c | 13 + 31 files changed, 1673 insertions(+), 350 deletions(-) create mode 100644 drivers/nvdimm/claim.c create mode 100644 drivers/nvdimm/e820.c create mode 100644 drivers/nvdimm/pfn.h create mode 100644 drivers/nvdimm/pfn_devs.c -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>