On 15/01/2020 11:09, Jürgen Groß wrote:
> On 15.01.20 11:54, Sergey Dyasli wrote:
>> Hi Juergen,
>>
>> On 08/01/2020 15:20, Sergey Dyasli wrote:
>>> It is incorrect to call pmd_populate_kernel() multiple times for the
>>> same page table. Xen notices it during kasan_populate_early_shadow():
>>>
>>> (XEN) mm.c:3222:d155v0 mfn 3704b already pinned
>>>
>>> This happens for kasan_early_shadow_pte when USE_SPLIT_PTE_PTLOCKS is
>>> enabled. Fix this by introducing set_pmd_early_shadow() which calls
>>> pmd_populate_kernel() only once and uses set_pmd() afterwards.
>>>
>>> Signed-off-by: Sergey Dyasli <sergey.dyasli@xxxxxxxxxx>
>>
>> Looks like the plan to use set_pmd() directly has failed: it's an
>> arch-specific function and can't be used in arch-independent code
>> (as kbuild test robot has proven).
>>
>> Do you see any way out of this other than disabling SPLIT_PTE_PTLOCKS
>> for PV KASAN?
>
> Change set_pmd_early_shadow() like the following:
>
> #ifdef CONFIG_XEN_PV
> static inline void set_pmd_early_shadow(pmd_t *pmd, pte_t *early_shadow)
> {
> static bool pmd_populated = false;
>
> if (likely(pmd_populated)) {
> set_pmd(pmd, __pmd(__pa(early_shadow) | _PAGE_TABLE));
> } else {
> pmd_populate_kernel(&init_mm, pmd, early_shadow);
> pmd_populated = true;
> }
> }
> #else
> static inline void set_pmd_early_shadow(pmd_t *pmd, pte_t *early_shadow)
> {
> pmd_populate_kernel(&init_mm, pmd, early_shadow);
> }
> #endif
>
> ... and move it to include/xen/xen-ops.h and call it with
> lm_alias(kasan_early_shadow_pte) as the second parameter.
Your suggestion to use ifdef is really good, especially now when I
figured out that CONFIG_XEN_PV implies X86. But I don't like the idea
of kasan code calling a non-empty function from xen-ops.h when
CONFIG_XEN_PV is not defined. I'd prefer to keep set_pmd_early_shadow()
in mm/kasan/init.c with the suggested ifdef.
--
Thanks,
Sergey
