Changes since v1 [1]:
* Fix the fact that devmem_is_allowed() == 2 does not prevent
mmap access (Kees)
* Rather than teach devmem_is_allowed() == 2 to map zero pages in the
mmap case, just fail (Nikolay)
[1]: http://lore.kernel.org/67f5b75c37143_71fe2949b@xxxxxxxxxxxxxxxxxxxxxxxxx.notmuch
---
The story starts with Nikolay reporting an SEPT violation due to
mismatched encrypted/non-encrypted mappings of the BIOS data space [2].
An initial suggestion to just make sure that the BIOS data space is
mapped consistently [3] ran into another issue that TDX and SEV-SNP
disagree about when that space can be mapped as encrypted.
Then, in response to a partial patch to allow SEV-SNP to block BIOS data
space for other reasons [4], Dave asked why not just give up on /dev/mem
access entirely in the confidential VM case [5].
Enter this series to:
1/ Close a subtle hole whereby /dev/mem that is supposed return zeros in
lieu of access only enforces that for read()/write()
2/ Use that new closed hole to reliably disable all /dev/mem access for
confidential x86 VMs
[2]: http://lore.kernel.org/20250318113604.297726-1-nik.borisov@xxxxxxxx
[3]: http://lore.kernel.org/174346288005.2166708.14425674491111625620.stgit@xxxxxxxxxxxxxxxxxxxxxxxxx
[4]: http://lore.kernel.org/20250403120228.2344377-1-naveen@xxxxxxxxxx
[5]: http://lore.kernel.org/fd683daa-d953-48ca-8c5d-6f4688ad442c@xxxxxxxxx
---
Dan Williams (3):
x86/devmem: Remove duplicate range_is_allowed() definition
devmem: Block mmap access when read/write access is restricted
x86/devmem: Restrict /dev/mem access for potentially unaccepted memory by default
arch/x86/Kconfig | 2 ++
arch/x86/include/asm/x86_init.h | 2 ++
arch/x86/kernel/x86_init.c | 6 ++++++
arch/x86/mm/init.c | 23 +++++++++++++++++------
arch/x86/mm/pat/memtype.c | 31 ++++---------------------------
drivers/char/mem.c | 18 ------------------
include/linux/io.h | 26 ++++++++++++++++++++++++++
7 files changed, 57 insertions(+), 51 deletions(-)
base-commit: 0af2f6be1b4281385b618cb86ad946eded089ac8