Message-ID: <20160408151653.GC8961@cbox>
Date:	Fri, 8 Apr 2016 17:16:53 +0200
From:	Christoffer Dall <christoffer.dall@...aro.org>
To:	Marc Zyngier <marc.zyngier@....com>
Cc:	Suzuki K Poulose <suzuki.poulose@....com>,
	linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
	kvmarm@...ts.cs.columbia.edu, kvm@...r.kernel.org,
	mark.rutland@....com, will.deacon@....com, catalin.marinas@....com
Subject: Re: [PATCH 12/17] kvm-arm: Add explicit hyp page table modifiers

On Fri, Apr 08, 2016 at 04:09:11PM +0100, Marc Zyngier wrote:
> On 08/04/16 14:15, Christoffer Dall wrote:
> > On Mon, Apr 04, 2016 at 05:26:12PM +0100, Suzuki K Poulose wrote:
> >> We have common routines to modify hyp and stage2 page tables
> >> based on the 'kvm' parameter. For a smoother transition to
> >> using separate routines for each, duplicate the routines
> >> and modify the copy to work on hyp.
> >>
> >> Marks the forked routines with _hyp_ and gets rid of the
> >> kvm parameter which is no longer needed and is NULL for hyp.
> >> Also, gets rid of calls to kvm_tlb_flush_by_vmid_ipa() calls
> >> from the hyp versions. Uses explicit host page table accessors
> >> instead of the kvm_* page table helpers.
> >>
> >> Suggested-by: Christoffer Dall <christoffer.dall@...aro.org>
> >> Cc: Marc Zyngier <marc.zyngier@....com>
> >> Signed-off-by: Suzuki K Poulose <suzuki.poulose@....com>
> >> ---
> >>  arch/arm/kvm/mmu.c |  123 +++++++++++++++++++++++++++++++++++++++++++++++++---
> >>  1 file changed, 118 insertions(+), 5 deletions(-)
> >>
> >> diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
> >> index b46a337..2b491e5 100644
> >> --- a/arch/arm/kvm/mmu.c
> >> +++ b/arch/arm/kvm/mmu.c
> >> @@ -388,6 +388,119 @@ static void stage2_flush_vm(struct kvm *kvm)
> >>  	srcu_read_unlock(&kvm->srcu, idx);
> >>  }
> >>  
> >> +static void clear_hyp_pgd_entry(pgd_t *pgd)
> >> +{
> >> +	pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
> >> +	pgd_clear(pgd);
> >> +	pud_free(NULL, pud_table);
> >> +	put_page(virt_to_page(pgd));
> >> +}
> >> +
> >> +static void clear_hyp_pud_entry(pud_t *pud)
> >> +{
> >> +	pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
> >> +	VM_BUG_ON(pud_huge(*pud));
> >> +	pud_clear(pud);
> >> +	pmd_free(NULL, pmd_table);
> >> +	put_page(virt_to_page(pud));
> >> +}
> >> +
> >> +static void clear_hyp_pmd_entry(pmd_t *pmd)
> >> +{
> >> +	pte_t *pte_table = pte_offset_kernel(pmd, 0);
> >> +	VM_BUG_ON(pmd_thp_or_huge(*pmd));
> >> +	pmd_clear(pmd);
> >> +	pte_free_kernel(NULL, pte_table);
> >> +	put_page(virt_to_page(pmd));
> >> +}
> >> +
> >> +static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
> >> +{
> >> +	pte_t *pte, *start_pte;
> >> +
> >> +	start_pte = pte = pte_offset_kernel(pmd, addr);
> >> +	do {
> >> +		if (!pte_none(*pte)) {
> >> +			pte_t old_pte = *pte;
> >> +
> >> +			kvm_set_pte(pte, __pte(0));
> >> +
> >> +			/* XXX: Do we need to invalidate the cache for device mappings ? */
> > 
> > no, we will not be swapping out any pages mapped in Hyp mode so you can
> > get rid of both of the following two lines.
> > 
> >> +			if (!kvm_is_device_pfn(pte_pfn(old_pte)))
> >> +				kvm_flush_dcache_pte(old_pte);
> >> +
> >> +			put_page(virt_to_page(pte));
> >> +		}
> >> +	} while (pte++, addr += PAGE_SIZE, addr != end);
> >> +
> >> +	if (hyp_pte_table_empty(start_pte))
> >> +		clear_hyp_pmd_entry(pmd);
> >> +}
> >> +
> >> +static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
> >> +{
> >> +	phys_addr_t next;
> >> +	pmd_t *pmd, *start_pmd;
> >> +
> >> +	start_pmd = pmd = pmd_offset(pud, addr);
> >> +	do {
> >> +		next = pmd_addr_end(addr, end);
> >> +		if (!pmd_none(*pmd)) {
> >> +			if (pmd_thp_or_huge(*pmd)) {
> > 
> > do we ever actually map anything with section mappings in the Hyp
> > mappings?
> 
> No, this is purely a page mapping so far. On my system, the HYP text is
> just over 4 pages big (4k pages), so the incentive is pretty low, unless
> we can demonstrate some big gains due to the reduced TLB impact.
> 
> >> +				pmd_t old_pmd = *pmd;
> >> +
> >> +				pmd_clear(pmd);
> >> +				kvm_flush_dcache_pmd(old_pmd);
> >> +				put_page(virt_to_page(pmd));
> >> +			} else {
> >> +				unmap_hyp_ptes(pmd, addr, next);
> >> +			}
> >> +		}
> >> +	} while (pmd++, addr = next, addr != end);
> >> +
> >> +	if (hyp_pmd_table_empty(start_pmd))
> >> +		clear_hyp_pud_entry(pud);
> >> +}
> >> +
> >> +static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
> >> +{
> >> +	phys_addr_t next;
> >> +	pud_t *pud, *start_pud;
> >> +
> >> +	start_pud = pud = pud_offset(pgd, addr);
> >> +	do {
> >> +		next = pud_addr_end(addr, end);
> >> +		if (!pud_none(*pud)) {
> >> +			if (pud_huge(*pud)) {
> > 
> > do we ever actually map anything with huge pud
> > mappings for the Hyp space?
> 
> Same thing. Looks like there is some potential simplification here.
> 
> > 
> >> +				pud_t old_pud = *pud;
> >> +
> >> +				pud_clear(pud);
> >> +				kvm_flush_dcache_pud(old_pud);
> >> +				put_page(virt_to_page(pud));
> >> +			} else {
> >> +				unmap_hyp_pmds(pud, addr, next);
> >> +			}
> >> +		}
> >> +	} while (pud++, addr = next, addr != end);
> >> +
> >> +	if (hyp_pud_table_empty(start_pud))
> >> +		clear_hyp_pgd_entry(pgd);
> >> +}
> >> +
> >> +static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
> >> +{
> >> +	pgd_t *pgd;
> >> +	phys_addr_t addr = start, end = start + size;
> >> +	phys_addr_t next;
> >> +
> >> +	pgd = pgdp + pgd_index(addr);
> >> +	do {
> >> +		next = pgd_addr_end(addr, end);
> >> +		if (!pgd_none(*pgd))
> >> +			unmap_hyp_puds(pgd, addr, next);
> >> +	} while (pgd++, addr = next, addr != end);
> > 
> > shouldn't we flush the EL2 (hyp) TLB here, strictly speaking?
> > 
> > Or do we rely on all mappings ever created/torn down here to always have
> > the same VA/PA relationship?  Since we didn't flush the EL2 TLB in the
> > existing code, that indeed does seem to be the case.
> 
> Actually, we never unmap anything from HYP. Once a structure (kvm, vcpu)
> is mapped there, it stays forever, whatever happens to the VM (that's
> because we'd otherwise have to refcount the number of objects in a page,
> and I'm lazy...).
> 
> > That, in turn, raises the question why we don't simply map all pages
> > that could be referenced by a kmalloc() in Hyp mode during the init
> > phase and be done with all this hyp mapping/unmapping stuff?
> > 
> > In any case, that behavior doesn't have to change now, but if we don't
> > add a TLB flush here, I'd like a comment to explain why that's not
> > needed.
> 
> Hope you have your answer above... ;-)
> 
Not quite: Could we just map the linearly mapped region in Hyp mode from
the beginning and be done with all this?

Otherwise yes, I have the answer, and we should add a comment too.

Thanks,
-Christoffer

Powered by blists - more mailing lists