Hi all,
commit a6d60245 "Track which levels of the page tables have been cleared"
added cleared_(ptes|pmds|puds|p4ds) in struct mmu_gather, and the values
of them are set in some places. For example:
In include/asm-generic/tlb.h, pte_free_tlb() set the tlb->cleared_pmds:
---8<---
#ifndef pte_free_tlb
#define pte_free_tlb(tlb, ptep, address) \
do { \
__tlb_adjust_range(tlb, address, PAGE_SIZE); \
tlb->freed_tables = 1; \
tlb->cleared_pmds = 1; \
__pte_free_tlb(tlb, ptep, address); \
} while (0)
#endif
---8<---
However, in arch/s390/include/asm/tlb.h, pte_free_tlb() set the tlb->cleared_ptes:
---8<---
static inline void pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
unsigned long address)
{
__tlb_adjust_range(tlb, address, PAGE_SIZE);
tlb->mm->context.flush_mm = 1;
tlb->freed_tables = 1;
tlb->cleared_ptes = 1;
/*
* page_table_free_rcu takes care of the allocation bit masks
* of the 2K table fragments in the 4K page table page,
* then calls tlb_remove_table.
*/
page_table_free_rcu(tlb, (unsigned long *) pte, address);
}
---8<---
In my view, the cleared_(ptes|pmds|puds) and (pte|pmd|pud)_free_tlb
correspond one-to-one. So we should set cleared_ptes in pte_free_tlb(),
then use it when needed.
I'm very confused about this. Which is wrong? Or is there something
I understand wrong?
Thanks,
Zhenyu
On Sat, Mar 28, 2020 at 12:30:50PM +0800, Zhenyu Ye wrote:
> Hi all,
>
> commit a6d60245 "Track which levels of the page tables have been cleared"
> added cleared_(ptes|pmds|puds|p4ds) in struct mmu_gather, and the values
> of them are set in some places. For example:
>
> In include/asm-generic/tlb.h, pte_free_tlb() set the tlb->cleared_pmds:
> ---8<---
> #ifndef pte_free_tlb
> #define pte_free_tlb(tlb, ptep, address) \
> do { \
> __tlb_adjust_range(tlb, address, PAGE_SIZE); \
> tlb->freed_tables = 1; \
> tlb->cleared_pmds = 1; \
> __pte_free_tlb(tlb, ptep, address); \
> } while (0)
> #endif
> ---8<---
>
>
> However, in arch/s390/include/asm/tlb.h, pte_free_tlb() set the tlb->cleared_ptes:
> ---8<---
> static inline void pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
> unsigned long address)
> {
> __tlb_adjust_range(tlb, address, PAGE_SIZE);
> tlb->mm->context.flush_mm = 1;
> tlb->freed_tables = 1;
> tlb->cleared_ptes = 1;
> /*
> * page_table_free_rcu takes care of the allocation bit masks
> * of the 2K table fragments in the 4K page table page,
> * then calls tlb_remove_table.
> */
> page_table_free_rcu(tlb, (unsigned long *) pte, address);
> }
> ---8<---
>
>
> In my view, the cleared_(ptes|pmds|puds) and (pte|pmd|pud)_free_tlb
> correspond one-to-one. So we should set cleared_ptes in pte_free_tlb(),
> then use it when needed.
So pte_free_tlb() clears a table of PTE entries, or a PMD level entity,
also see free_pte_range(). So the generic code makes sense to me. The
PTE level invalidations will have happened on tlb_remove_tlb_entry().
> I'm very confused about this. Which is wrong? Or is there something
> I understand wrong?
I agree the s390 case is puzzling, Martin does s390 need a PTE level
invalidate for removing a PTE table or was this a mistake?
Hi Peter,
On 2020/3/30 20:16, Peter Zijlstra wrote:
> On Sat, Mar 28, 2020 at 12:30:50PM +0800, Zhenyu Ye wrote:
>> Hi all,
>>
>> commit a6d60245 "Track which levels of the page tables have been cleared"
>> added cleared_(ptes|pmds|puds|p4ds) in struct mmu_gather, and the values
>> of them are set in some places. For example:
>>
>> In include/asm-generic/tlb.h, pte_free_tlb() set the tlb->cleared_pmds:
>> ---8<---
>> #ifndef pte_free_tlb
>> #define pte_free_tlb(tlb, ptep, address) \
>> do { \
>> __tlb_adjust_range(tlb, address, PAGE_SIZE); \
>> tlb->freed_tables = 1; \
>> tlb->cleared_pmds = 1; \
>> __pte_free_tlb(tlb, ptep, address); \
>> } while (0)
>> #endif
>> ---8<---
>>
>>
>> However, in arch/s390/include/asm/tlb.h, pte_free_tlb() set the tlb->cleared_ptes:
>> ---8<---
>> static inline void pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
>> unsigned long address)
>> {
>> __tlb_adjust_range(tlb, address, PAGE_SIZE);
>> tlb->mm->context.flush_mm = 1;
>> tlb->freed_tables = 1;
>> tlb->cleared_ptes = 1;
>> /*
>> * page_table_free_rcu takes care of the allocation bit masks
>> * of the 2K table fragments in the 4K page table page,
>> * then calls tlb_remove_table.
>> */
>> page_table_free_rcu(tlb, (unsigned long *) pte, address);
>> }
>> ---8<---
>>
>>
>> In my view, the cleared_(ptes|pmds|puds) and (pte|pmd|pud)_free_tlb
>> correspond one-to-one. So we should set cleared_ptes in pte_free_tlb(),
>> then use it when needed.
>
> So pte_free_tlb() clears a table of PTE entries, or a PMD level entity,
> also see free_pte_range(). So the generic code makes sense to me. The
> PTE level invalidations will have happened on tlb_remove_tlb_entry().
>
Thanks for your explanation. I can understand now.
>> I'm very confused about this. Which is wrong? Or is there something
>> I understand wrong?
>
> I agree the s390 case is puzzling, Martin does s390 need a PTE level
> invalidate for removing a PTE table or was this a mistake?
>
Then we should wait for @ Martin's reply. Though s390 has never used
this value, I think we still should correct it if this is a mistake.
Thanks,
Zhenyu
Sorry, just saw that now..
On 30.03.20 14:16, Peter Zijlstra wrote:
> On Sat, Mar 28, 2020 at 12:30:50PM +0800, Zhenyu Ye wrote:
>> Hi all,
>>
>> commit a6d60245 "Track which levels of the page tables have been cleared"
>> added cleared_(ptes|pmds|puds|p4ds) in struct mmu_gather, and the values
>> of them are set in some places. For example:
>>
>> In include/asm-generic/tlb.h, pte_free_tlb() set the tlb->cleared_pmds:
>> ---8<---
>> #ifndef pte_free_tlb
>> #define pte_free_tlb(tlb, ptep, address) \
>> do { \
>> __tlb_adjust_range(tlb, address, PAGE_SIZE); \
>> tlb->freed_tables = 1; \
>> tlb->cleared_pmds = 1; \
>> __pte_free_tlb(tlb, ptep, address); \
>> } while (0)
>> #endif
>> ---8<---
>>
>>
>> However, in arch/s390/include/asm/tlb.h, pte_free_tlb() set the tlb->cleared_ptes:
>> ---8<---
>> static inline void pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
>> unsigned long address)
>> {
>> __tlb_adjust_range(tlb, address, PAGE_SIZE);
>> tlb->mm->context.flush_mm = 1;
>> tlb->freed_tables = 1;
>> tlb->cleared_ptes = 1;
>> /*
>> * page_table_free_rcu takes care of the allocation bit masks
>> * of the 2K table fragments in the 4K page table page,
>> * then calls tlb_remove_table.
>> */
>> page_table_free_rcu(tlb, (unsigned long *) pte, address);
>> }
>> ---8<---
adding Gerald and Vasily. Gerald can you have a look?
>>
>>
>> In my view, the cleared_(ptes|pmds|puds) and (pte|pmd|pud)_free_tlb
>> correspond one-to-one. So we should set cleared_ptes in pte_free_tlb(),
>> then use it when needed.
>
> So pte_free_tlb() clears a table of PTE entries, or a PMD level entity,
> also see free_pte_range(). So the generic code makes sense to me. The
> PTE level invalidations will have happened on tlb_remove_tlb_entry().
>
>> I'm very confused about this. Which is wrong? Or is there something
>> I understand wrong?
>
> I agree the s390 case is puzzling, Martin does s390 need a PTE level
> invalidate for removing a PTE table or was this a mistake?
>
Gerald,
can you have a look?
On 30.03.20 14:16, Peter Zijlstra wrote:
> On Sat, Mar 28, 2020 at 12:30:50PM +0800, Zhenyu Ye wrote:
>> Hi all,
>>
>> commit a6d60245 "Track which levels of the page tables have been cleared"
>> added cleared_(ptes|pmds|puds|p4ds) in struct mmu_gather, and the values
>> of them are set in some places. For example:
>>
>> In include/asm-generic/tlb.h, pte_free_tlb() set the tlb->cleared_pmds:
>> ---8<---
>> #ifndef pte_free_tlb
>> #define pte_free_tlb(tlb, ptep, address) \
>> do { \
>> __tlb_adjust_range(tlb, address, PAGE_SIZE); \
>> tlb->freed_tables = 1; \
>> tlb->cleared_pmds = 1; \
>> __pte_free_tlb(tlb, ptep, address); \
>> } while (0)
>> #endif
>> ---8<---
>>
>>
>> However, in arch/s390/include/asm/tlb.h, pte_free_tlb() set the tlb->cleared_ptes:
>> ---8<---
>> static inline void pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
>> unsigned long address)
>> {
>> __tlb_adjust_range(tlb, address, PAGE_SIZE);
>> tlb->mm->context.flush_mm = 1;
>> tlb->freed_tables = 1;
>> tlb->cleared_ptes = 1;
>> /*
>> * page_table_free_rcu takes care of the allocation bit masks
>> * of the 2K table fragments in the 4K page table page,
>> * then calls tlb_remove_table.
>> */
>> page_table_free_rcu(tlb, (unsigned long *) pte, address);
>> }
>> ---8<---
>>
>>
>> In my view, the cleared_(ptes|pmds|puds) and (pte|pmd|pud)_free_tlb
>> correspond one-to-one. So we should set cleared_ptes in pte_free_tlb(),
>> then use it when needed.
>
> So pte_free_tlb() clears a table of PTE entries, or a PMD level entity,
> also see free_pte_range(). So the generic code makes sense to me. The
> PTE level invalidations will have happened on tlb_remove_tlb_entry().
>
>> I'm very confused about this. Which is wrong? Or is there something
>> I understand wrong?
>
> I agree the s390 case is puzzling, Martin does s390 need a PTE level
> invalidate for removing a PTE table or was this a mistake?
>
On Wed, 8 Apr 2020 10:51:59 +0200
Christian Borntraeger <[email protected]> wrote:
[...]
>
> adding Gerald and Vasily. Gerald can you have a look?
>
> >>
> >>
> >> In my view, the cleared_(ptes|pmds|puds) and (pte|pmd|pud)_free_tlb
> >> correspond one-to-one. So we should set cleared_ptes in pte_free_tlb(),
> >> then use it when needed.
> >
> > So pte_free_tlb() clears a table of PTE entries, or a PMD level entity,
> > also see free_pte_range(). So the generic code makes sense to me. The
> > PTE level invalidations will have happened on tlb_remove_tlb_entry().
> >
> >> I'm very confused about this. Which is wrong? Or is there something
> >> I understand wrong?
> >
> > I agree the s390 case is puzzling, Martin does s390 need a PTE level
> > invalidate for removing a PTE table or was this a mistake?
> >
Peter is right, the PTE level invalidations will happen before. For
s390, not exactly at the tlb_remove_tlb_entry() itself, since
__tlb_remove_tlb_entry() is not defined, but rather directly at the
preceding ptep_get_and_clear(). I think this also the reason why we
cannot easily optimize for larger granularity.
Anyway, pte_free_tlb() will then later only take care of freeing
the page table page, no further PTE level clearing/invalidation
needed. I see no reason why s390 should behave differently from
the generic code, wrt to cleared_pxds setting in pxd_free_tlb().
So I guess this was an "off-by-one" mistake in commit 9de7d833e3708
("s390/tlb: Convert to generic mmu_gather"), since the other
pxd_free_tlb() functions also show similar puzzling behavior.
Not consistently off-by-one though, as pmd_free_tlb() seems
to behave correctly, setting tlb->cleared_puds = 1, similar to
generic code.
That was a very nice catch, Zhenyu, thanks for reporting!
We are not yet making use of the tlb->cleared_pxds for s390, but
we would certainly have stumbled over this if we ever tried.
Will send a patch to make s390 behave like generic code here.
Regards,
Gerald