Re: [PATCH v12 64/84] KVM: LoongArch: Mark "struct page" pfns dirty only in "slow" page fault path

[Sean Christopherson <seanjc@google.com>]

On Sat, Aug 03, 2024, maobibo wrote:
> On 2024/8/3 上午3:32, Sean Christopherson wrote:
> > On Fri, Aug 02, 2024, maobibo wrote:
> > > On 2024/7/27 上午7:52, Sean Christopherson wrote:
> > > > Mark pages/folios dirty only the slow page fault path, i.e. only when
> > > > mmu_lock is held and the operation is mmu_notifier-protected, as marking a
> > > > page/folio dirty after it has been written back can make some filesystems
> > > > unhappy (backing KVM guests will such filesystem files is uncommon, and
> > > > the race is minuscule, hence the lack of complaints).
> > > > 
> > > > See the link below for details.
> > > > 
> > > > Link: https://lore.kernel.org/all/cover.1683044162.git.lstoakes@gmail.com
> > > > Signed-off-by: Sean Christopherson <seanjc@google.com>
> > > > ---
> > > >    arch/loongarch/kvm/mmu.c | 18 ++++++++++--------
> > > >    1 file changed, 10 insertions(+), 8 deletions(-)
> > > > 
> > > > diff --git a/arch/loongarch/kvm/mmu.c b/arch/loongarch/kvm/mmu.c
> > > > index 2634a9e8d82c..364dd35e0557 100644
> > > > --- a/arch/loongarch/kvm/mmu.c
> > > > +++ b/arch/loongarch/kvm/mmu.c
> > > > @@ -608,13 +608,13 @@ static int kvm_map_page_fast(struct kvm_vcpu *vcpu, unsigned long gpa, bool writ
> > > >    		if (kvm_pte_young(changed))
> > > >    			kvm_set_pfn_accessed(pfn);
> > > > -		if (kvm_pte_dirty(changed)) {
> > > > -			mark_page_dirty(kvm, gfn);
> > > > -			kvm_set_pfn_dirty(pfn);
> > > > -		}
> > > >    		if (page)
> > > >    			put_page(page);
> > > >    	}
> > > > +
> > > > +	if (kvm_pte_dirty(changed))
> > > > +		mark_page_dirty(kvm, gfn);
> > > > +
> > > >    	return ret;
> > > >    out:
> > > >    	spin_unlock(&kvm->mmu_lock);
> > > > @@ -915,12 +915,14 @@ static int kvm_map_page(struct kvm_vcpu *vcpu, unsigned long gpa, bool write)
> > > >    	else
> > > >    		++kvm->stat.pages;
> > > >    	kvm_set_pte(ptep, new_pte);
> > > > -	spin_unlock(&kvm->mmu_lock);
> > > > -	if (prot_bits & _PAGE_DIRTY) {
> > > > -		mark_page_dirty_in_slot(kvm, memslot, gfn);
> > > > +	if (writeable)
> > > Is it better to use write or (prot_bits & _PAGE_DIRTY) here?  writable is
> > > pte permission from function hva_to_pfn_slow(), write is fault action.
> > 
> > Marking folios dirty in the slow/full path basically necessitates marking the
> > folio dirty if KVM creates a writable SPTE, as KVM won't mark the folio dirty
> > if/when _PAGE_DIRTY is set.
> > 
> > Practically speaking, I'm 99.9% certain it doesn't matter.  The folio is marked
> > dirty by core MM when the folio is made writable, and cleaning the folio triggers
> > an mmu_notifier invalidation.  I.e. if the page is mapped writable in KVM's
> yes, it is. Thanks for the explanation. kvm_set_pfn_dirty() can be put only
> in slow page fault path. I only concern with fault type, read fault type can
> set pte entry writable however not _PAGE_DIRTY at stage-2 mmu table.
> 
> > stage-2 PTEs, then its folio has already been marked dirty.
> Considering one condition although I do not know whether it exists actually.
> user mode VMM writes the folio with hva address firstly, then VCPU thread
> *reads* the folio. With primary mmu table, pte entry is writable and
> _PAGE_DIRTY is set, with secondary mmu table(state-2 PTE table), it is
> pte_none since the filio is accessed at first time, so there will be slow
> page fault path for stage-2 mmu page table filling.
> 
> Since it is read fault, stage-2 PTE will be created with _PAGE_WRITE(coming
> from function hva_to_pfn_slow()), however _PAGE_DIRTY is not set. Do we need
> call kvm_set_pfn_dirty() at this situation?

If KVM doesn't mark the folio dirty when the stage-2 _PAGE_DIRTY flag is set,
i.e. as proposed in this series, then yes, KVM needs to call kvm_set_pfn_dirty()
even though the VM hasn't (yet) written to the memory.  In practice, KVM calling
kvm_set_pfn_dirty() is redundant the majority of the time, as the stage-1 PTE
will have _PAGE_DIRTY set, and that will get propagated to the folio when the
primary MMU does anything relevant with the PTE.  And for file systems that care
about writeback, odds are very good that the folio was marked dirty even earlier,
when MM invoked vm_operations_struct.page_mkwrite().

The reason I am pushing to have all architectures mark pages/folios dirty in the
slow page fault path is that a false positive (marking a folio dirty without the
folio ever being written in _any_ context since the last pte_mkclean()) is rare,
and at worst results an unnecessary writeback.  On the other hand, marking folios
dirty in fast page fault handlers (or anywhere else that isn't protected by
mmu_notifiers) is technically unsafe.

In other words, the intent is to sacrifice accuracy to improve stability/robustness,
because the vast majority of time the loss in accuracy has no effect, and the worst
case scenario is that the kernel does I/O that wasn't necessary.