> On Sep 25, 2021, at 1:54 PM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>
> From: Nadav Amit <namit@xxxxxxxxxx>
>
> change_pXX_range() currently does not use mmu_gather, but instead
> implements its own deferred TLB flushes scheme. This both complicates
> the code, as developers need to be aware of different invalidation
> schemes, and prevents opportunities to avoid TLB flushes or perform them
> in finer granularity.
>
> Use mmu_gather in change_pXX_range(). As the pages are not released,
> only record the flushed range using tlb_flush_pXX_range().
Andrea pointed out that I do not take care of THP. Actually, there is
indeed a missing TLB flush on THP, but it is not required due to the
pmdp_invalidate(). Anyhow, the patch needs to address it cleanly, and
to try to avoid the flush on pmdp_invalidate(), which at least on x86
does not appear to be necessary.
There is an additional bug, as tlb_change_page_size() needs to be
called.
-- Jerome,
While I am reviewing my (bad) code, I wanted to understand whether
update of migration entries requires a TLB flush, because I do not
think I got that right either.
I thought they should not, but I now am not very sure. I am very
confused by the following code in migrate_vma_collect_pmd():
pte_unmap_unlock(ptep - 1, ptl);
/* Only flush the TLB if we actually modified any entries */
if (unmapped)
flush_tlb_range(walk->vma, start, end);
According to this code flush_tlb_range() is called without the ptl.
So theoretically there is a possible race:
CPU0 CPU1
---- ----
migrate_vma_collect_pmd()
set_pte_at() [ present->
non-present]
pte_unmap_unlock()
madvise(MADV_DONTNEED)
zap_pte_range()
[ PTE non-present =>
no flush ]
So my questions:
1. Is there a reason the above scenario is invalid?
2. Does one need to flush a migration entry he updates it?
Thanks,
Nadav
