Lina reports random oopsen originating from the fast GUP code when 16K pages are used with 4-level page-tables, the fourth level being folded at runtime due to lack of LPA2. In this configuration, the generic implementation of p4d_offset_lockless() will return a 'p4d_t *' corresponding to the 'pgd_t' allocated on the stack of the caller, gup_fast_pgd_range(). This is normally fine, but when the fourth level of page-table is folded at runtime, pud_offset_lockless() will offset from the address of the 'p4d_t' to calculate the address of the PUD in the same page-table page. This results in a stray stack read when the 'p4d_t' has been allocated on the stack and can send the walker into the weeds. Fix the problem by providing our own definition of p4d_offset_lockless() when CONFIG_PGTABLE_LEVELS <= 4 which returns the real page-table pointer rather than the address of the local stack variable. Cc: Catalin Marinas <catalin.marinas@xxxxxxx> Cc: Ard Biesheuvel <ardb@xxxxxxxxxx> Cc: <stable@xxxxxxxxxxxxxxx> Link: https://lore.kernel.org/r/50360968-13fb-4e6f-8f52-1725b3177215@xxxxxxxxxxxxx Fixes: 0dd4f60a2c76 ("arm64: mm: Add support for folding PUDs at runtime") Reported-by: Asahi Lina <lina@xxxxxxxxxxxxx> Signed-off-by: Will Deacon <will@xxxxxxxxxx> --- arch/arm64/include/asm/pgtable.h | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h index f8efbc128446..7a4f5604be3f 100644 --- a/arch/arm64/include/asm/pgtable.h +++ b/arch/arm64/include/asm/pgtable.h @@ -1065,6 +1065,28 @@ static inline bool pgtable_l5_enabled(void) { return false; } #define p4d_offset_kimg(dir,addr) ((p4d_t *)dir) +static inline +p4d_t *p4d_offset_lockless_folded(pgd_t *pgdp, pgd_t pgd, unsigned long addr) +{ + /* + * With runtime folding of the pud, pud_offset_lockless() passes + * the 'pgd_t *' we return here to p4d_to_folded_pud(), which + * will offset the pointer assuming that it points into + * a page-table page. However, the fast GUP path passes us a + * pgd_t allocated on the stack and so we must use the original + * pointer in 'pgdp' to construct the p4d pointer instead of + * using the generic p4d_offset_lockless() implementation. + * + * Note: reusing the original pointer means that we may + * dereference the same (live) page-table entry multiple times. + * This is safe because it is still only loaded once in the + * context of each level and the CPU guarantees same-address + * read-after-read ordering. + */ + return p4d_offset(pgdp, addr); +} +#define p4d_offset_lockless p4d_offset_lockless_folded + #endif /* CONFIG_PGTABLE_LEVELS > 4 */ #define pgd_ERROR(e) \ -- 2.45.2.1089.g2a221341d9-goog
