On 27/10/2023 16:32, Alexandre Ghiti wrote:
Those 2 patches fix the set_memory_XX() and set_direct_map_XX() APIs, which
in turn fix STRICT_KERNEL_RWX and memfd_secret(). Those were broken since the
permission changes were not applied to the linear mapping because the linear
mapping is mapped using hugepages and walk_page_range_novma() does not split
such mappings.
To fix that, patch 1 disables PGD mappings in the linear mapping as it is
hard to propagate changes at this level in *all* the page tables, this has the
downside of disabling PMD mapping for sv32 and PUD (1GB) mapping for sv39 in
the linear mapping (for specific kernels, we could add a Kconfig to enable
ARCH_HAS_SET_DIRECT_MAP and STRICT_KERNEL_RWX if needed, I'm pretty sure we'll
discuss that).
patch 2 implements the split of the huge linear mappings so that
walk_page_range_novma() can properly apply the permissions. The whole split is
protected with mmap_sem in write mode, but I'm wondering if that's enough,
any opinion on that is appreciated.
Changes in v2:
- Fix rv32 build
- Rebase on top of 6.6-rc6
- Move the flush_tlb_all outside the mmap_lock
Alexandre Ghiti (2):
riscv: Don't use PGD entries for the linear mapping
riscv: Fix set_memory_XX() and set_direct_map_XX() by splitting huge
linear mappings
arch/riscv/mm/init.c | 12 +-
arch/riscv/mm/pageattr.c | 270 +++++++++++++++++++++++++++++++++------
2 files changed, 236 insertions(+), 46 deletions(-)
Any feedback from anyone? That's a fix I'd like to see merged in 6.7, if
no objection, I'll ask for Palmer to take it.
Thanks,
Alex
