arm64: mm: fix pmd_write CoW brokenness
authorWill Deacon <will.deacon@arm.com>
Tue, 27 May 2014 18:11:58 +0000 (19:11 +0100)
committerMark Brown <broonie@kernel.org>
Fri, 24 Oct 2014 10:54:42 +0000 (11:54 +0100)
commit263d7734264145c7517194cf06dc84cebb15a2a7
tree6d424bd1418a86ec0186a9c8b3b6ac1548807648
parentae635c76a615535a1c080f5e1e49b8c3a032c1cd
arm64: mm: fix pmd_write CoW brokenness

Commit 9c7e535fcc17 ("arm64: mm: Route pmd thp functions through pte
equivalents") changed the pmd manipulator and accessor functions to
convert the target pmd to a pte, process it with the pte functions, then
convert it back. Along the way, we gained support for PTE_WRITE, however
this is completely ignored by set_pmd_at, and so we fail to set the
PMD_SECT_RDONLY for PMDs, resulting in all sorts of lovely failures (like
CoW not working).

Partially reverting the offending commit (by making use of
PMD_SECT_RDONLY explicitly for pmd_{write,wrprotect,mkwrite} functions)
leads to further issues because pmd_write can then return potentially
incorrect values for page table entries marked as RDONLY, leading to
BUG_ON(pmd_write(entry)) tripping under some THP workloads.

This patch fixes the issue by routing set_pmd_at through set_pte_at,
which correctly takes the PTE_WRITE flag into account. Given that
THP mappings are always anonymous, the additional cache-flushing code
in __sync_icache_dcache won't impose any significant overhead as the
flush will be skipped.

Cc: Catalin Marinas <catalin.marinas@arm.com>
Acked-by: Steve Capper <steve.capper@arm.com>
Tested-by: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit ceb218359de22e70980801d4fa04fffbfc44adb8)
Signed-off-by: Mark Brown <broonie@kernel.org>
arch/arm64/include/asm/pgtable.h