On 10/29/19 4:38 AM, Dan Williams wrote:

On Mon, Oct 28, 2019 at 2:48 AM Aneesh Kumar K.V
[off-list ref] wrote:

quoted

The page size used to map the namespace is arch dependent. For example
architectures like ppc64 use 16MB page size for direct-mapping. If the namespace
size is not aligned to the mapping page size, we can observe kernel crash
during namespace init and destroy.

This is due to kernel doing partial map/unmap of the resource range

BUG: Unable to handle kernel data access at 0xc001000406000000
Faulting instruction address: 0xc000000000090790
NIP [c000000000090790] arch_add_memory+0xc0/0x130
LR [c000000000090744] arch_add_memory+0x74/0x130
Call Trace:
  arch_add_memory+0x74/0x130 (unreliable)
  memremap_pages+0x74c/0xa30
  devm_memremap_pages+0x3c/0xa0
  pmem_attach_disk+0x188/0x770
  nvdimm_bus_probe+0xd8/0x470
  really_probe+0x148/0x570
  driver_probe_device+0x19c/0x1d0
  device_driver_attach+0xcc/0x100
  bind_store+0x134/0x1c0
  drv_attr_store+0x44/0x60
  sysfs_kf_write+0x74/0xc0
  kernfs_fop_write+0x1b4/0x290
  __vfs_write+0x3c/0x70
  vfs_write+0xd0/0x260
  ksys_write+0xdc/0x130
  system_call+0x5c/0x68

Signed-off-by: Aneesh Kumar K.V <redacted>
---
  arch/arm64/mm/flush.c     | 11 +++++++++++
  arch/powerpc/lib/pmem.c   | 21 +++++++++++++++++++--
  arch/x86/mm/pageattr.c    | 12 ++++++++++++
  include/linux/libnvdimm.h |  1 +
  4 files changed, 43 insertions(+), 2 deletions(-)

diff --git a/arch/arm64/mm/flush.c b/arch/arm64/mm/flush.c
index ac485163a4a7..90c54c600023 100644
--- a/arch/arm64/mm/flush.c
+++ b/arch/arm64/mm/flush.c

@@ -91,4 +91,15 @@ void arch_invalidate_pmem(void *addr, size_t size)
         __inval_dcache_area(addr, size);
  }
  EXPORT_SYMBOL_GPL(arch_invalidate_pmem);
+
+unsigned long arch_validate_namespace_size(unsigned int ndr_mappings, unsigned long size)
+{
+       u32 remainder;
+
+       div_u64_rem(size, PAGE_SIZE * ndr_mappings, &remainder);
+       if (remainder)
+               return PAGE_SIZE * ndr_mappings;
+       return 0;
+}
+EXPORT_SYMBOL_GPL(arch_validate_namespace_size);
  #endif

diff --git a/arch/powerpc/lib/pmem.c b/arch/powerpc/lib/pmem.c
index 377712e85605..2e661a08dae5 100644
--- a/arch/powerpc/lib/pmem.c
+++ b/arch/powerpc/lib/pmem.c

@@ -17,14 +17,31 @@ void arch_wb_cache_pmem(void *addr, size_t size)
         unsigned long start = (unsigned long) addr;
         flush_dcache_range(start, start + size);
  }
-EXPORT_SYMBOL(arch_wb_cache_pmem);
+EXPORT_SYMBOL_GPL(arch_wb_cache_pmem);

  void arch_invalidate_pmem(void *addr, size_t size)
  {
         unsigned long start = (unsigned long) addr;
         flush_dcache_range(start, start + size);
  }
-EXPORT_SYMBOL(arch_invalidate_pmem);
+EXPORT_SYMBOL_GPL(arch_invalidate_pmem);
+
+unsigned long arch_validate_namespace_size(unsigned int ndr_mappings, unsigned long size)
+{
+       u32 remainder;
+       unsigned long linear_map_size;
+
+       if (radix_enabled())
+               linear_map_size = PAGE_SIZE;
+       else
+               linear_map_size = (1UL << mmu_psize_defs[mmu_linear_psize].shift);

This seems more a "supported_alignments" problem, and less a namespace
size or PAGE_SIZE problem, because if the starting address is
misaligned this size validation can still succeed when it shouldn't.

Isn't supported_alignments an indication of how user want the namespace 
to be mapped to applications?  Ie, with the above restrictions we can 
still do both 64K and 16M mapping of the namespace to userspace.

Also for supported alignment the huge page mapping is further  dependent 
on the THP feature.

The restrictions here are mostly w.r.t the direct-mapping page size used 
by some architecture.

One problem is that __size_store() does not validate the size against
the namespace alignment.

However, the next problem is that alignment is a property of the pfn
device, but not the raw namespace. I think this alignment constraint
should be captured by exposing "align" and "supported_alignments" at
the namespace level as the minimum alignment. The pfn level alignment
could then be an additional alignment constraint, but ndctl would
likely set them to the same value.

The "* ndr_mappings" constraint should be left out of the interface,
because that's a side effect of supporting namespace-type-blk aliasing
which no platform seems to do in practice. If for some strange reason
it's need in the future I'd rather expose the "aliasing" property
rather than fold it into the align / supported_aligns interface.

-aneesh