On Tuesday 13 October 2015 09:37 PM, Kirill A. Shutemov wrote:
> On Tue, Oct 13, 2015 at 01:58:39PM +0000, Vineet Gupta wrote:
>> Hi Kirill,
>>
>> I'm running LTP tests on the new ARC THP code and thp03 seems to be triggering mm
>> spew.
>>
>> --------------->8---------------------
>> [ARCLinux]# ./ltp-thp03-extract
>> PID 60
>> bad pmd bf1c4600 be600231
>> ../mm/pgtable-generic.c:34: bad pgd be600231.
>> bad pmd bf1c4604 bd800231
>> ../mm/pgtable-generic.c:34: bad pgd bd800231.
>> BUG: Bad rss-counter state mm:bf12e900 idx:1 val:512
>> BUG: non-zero nr_ptes on freeing mm: 2
>> --------------->8---------------------
>>
>> I know what exactly is happening and the likely fix, but would want to get some
>> thoughts from you if possible.
>>
>> background: ARC is software page walked with PGD -> PTE -> page for normal and PMD
>> -> page for THP case. A vanilla PGD doesn't have any flags - only pointer to PTE
>>
>> A reduced version of thp03 allocates a THP, dirties it, followed by
>> mprotect(PROT_NONE).
>> At the time of mprotect() -> change_huge_pmd() -> pmd_modify() needs to change
>> some of the bits.
>>
>> The issue is ARC implementation of pmd_modify() based on pte variant, which
>> retains the soft pte bits (dirty and accessed).
>>
>> static inline pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
>> {
>> return pte_pmd(pte_modify(pmd_pte(pmd), newprot));
>> }
>>
>> Obvious fix is to rewrite pmd_modify() so that it clears out all pte type flags
>> but that assumes PMD is becoming PGD (a vanilla PGD on ARC doesn't have any
>> flags). Can we have pmd_modify() ever be called for NOT splitting pmd e.g.
>> mprotect Write to Read which won't split the THP like it does now and simply
>> changes the prot flags. My proposed version of pmd_modify() will loose the dirty bit.
> Hm? pmd_modify() is nothing to do with splitting. The mprotect() codepath
> you've mentioned above calls pmd_modify() only if the THP is fully in
> mprotect range.
Indeed my mental picture of this was messed up - specially because behind the
back, pmd_modify() for ARC (based on pte_modify()) was buggered to clear the huge
page bit itself :-) So we had a THP PMD which would start failing for
pmd_trans_huge() and thus treated like a normal PGD. But it had the leftover PMD
soft bits, which triggered the MM spew.
The localized fix is below, while better fix is to make pte_modify() only clear
R-W-X bits (currently it clears everything except soft accessed/dirty bits)
static inline pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
{
- return pte_pmd(pte_modify(pmd_pte(pmd), newprot));
+ /*
+ * open-coded pte_modify() additionally retaining HW_SZ bit
+ * otherwise, pmd_trans_huge() checks start failing
+ */
+ return __pmd((pmd_val(pmd) & (_PAGE_CHG_MASK | _PAGE_HW_SZ)) |
pgprot_val(newprot));
}
>
>> In short, what are the semantics of pmd_modify() - essentially does it imply pmd
>> is being split so are free to make it like PGD.
> No, pmd_modify() cannot make such assumption. That's just not true -- we
> don't split PMD in such codepath. And even if we do, we construct new PMD
> entry from scratch instead of modifying existing one.
>
> So the semantics of pmd_modify(): you can assume that the entry is
> pmd_large(), going to stay this way and you need to touch only
> protection-related bit.
Thx !
-Vineet
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href