Message-ID: <20230314184123.000022ee@gmail.com>
Date:   Tue, 14 Mar 2023 18:41:23 +0200
From:   Zhi Wang <zhi.wang.linux@...il.com>
To:     Steven Price <steven.price@....com>
Cc:     kvm@...r.kernel.org, kvmarm@...ts.linux.dev,
        Catalin Marinas <catalin.marinas@....com>,
        Marc Zyngier <maz@...nel.org>, Will Deacon <will@...nel.org>,
        James Morse <james.morse@....com>,
        Oliver Upton <oliver.upton@...ux.dev>,
        Suzuki K Poulose <suzuki.poulose@....com>,
        Zenghui Yu <yuzenghui@...wei.com>,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
        Joey Gouly <joey.gouly@....com>,
        Alexandru Elisei <alexandru.elisei@....com>,
        Christoffer Dall <christoffer.dall@....com>,
        Fuad Tabba <tabba@...gle.com>, linux-coco@...ts.linux.dev
Subject: Re: [RFC PATCH 17/28] arm64: RME: Runtime faulting of memory

On Fri, 10 Mar 2023 15:47:19 +0000
Steven Price <steven.price@....com> wrote:

> On 06/03/2023 18:20, Zhi Wang wrote:
> > On Fri, 27 Jan 2023 11:29:21 +0000
> > Steven Price <steven.price@....com> wrote:
> >   
> >> At runtime if the realm guest accesses memory which hasn't yet been
> >> mapped then KVM needs to either populate the region or fault the guest.
> >>
> >> For memory in the lower (protected) region of IPA a fresh page is
> >> provided to the RMM which will zero the contents. For memory in the
> >> upper (shared) region of IPA, the memory from the memslot is mapped
> >> into the realm VM non secure.
> >>
> >> Signed-off-by: Steven Price <steven.price@....com>
> >> ---
> >>  arch/arm64/include/asm/kvm_emulate.h | 10 +++++
> >>  arch/arm64/include/asm/kvm_rme.h     | 12 ++++++
> >>  arch/arm64/kvm/mmu.c                 | 64 +++++++++++++++++++++++++---
> >>  arch/arm64/kvm/rme.c                 | 48 +++++++++++++++++++++
> >>  4 files changed, 128 insertions(+), 6 deletions(-)
> >>
> >> diff --git a/arch/arm64/include/asm/kvm_emulate.h b/arch/arm64/include/asm/kvm_emulate.h
> >> index 285e62914ca4..3a71b3d2e10a 100644
> >> --- a/arch/arm64/include/asm/kvm_emulate.h
> >> +++ b/arch/arm64/include/asm/kvm_emulate.h
> >> @@ -502,6 +502,16 @@ static inline enum realm_state kvm_realm_state(struct kvm *kvm)
> >>  	return READ_ONCE(kvm->arch.realm.state);
> >>  }
> >>  
> >> +static inline gpa_t kvm_gpa_stolen_bits(struct kvm *kvm)
> >> +{
> >> +	if (kvm_is_realm(kvm)) {
> >> +		struct realm *realm = &kvm->arch.realm;
> >> +
> >> +		return BIT(realm->ia_bits - 1);
> >> +	}
> >> +	return 0;
> >> +}
> >> +  
> > 
> > "stolen" seems a little bit vague. Maybe "shared" bit would be better as
> > SEV-SNP has C bit and TDX has shared bit. It would be nice to align with
> > the common knowledge.  
> 
> The Arm CCA term is the "protected" bit[1] - although the bit is
> backwards as it's cleared to indicate protected... so not ideal naming! ;)
> 
> But it's termed 'stolen' here as it's effectively removed from the set
> of value address bits. And this function is returning a mask of the bits
> that are not available as address bits. The naming was meant to be
> generic that this could encompass other features that need to reserve
> IPA bits.
> 
> But it's possible this is too generic and perhaps we should just deal
> with a single bit rather than potential masks. Alternatively we could
> invert this and return a set of valid bits:
> 
> static inline gpa_t kvm_gpa_valid_bits(struct kvm *kvm)
> {
> 	if (kvm_is_realm(kvm)) {
> 		struct realm *realm = &kvm->arch.realm;
> 
> 		return ~BIT(realm->ia_bits - 1);
> 	}
> 	return ~(gpa_t)0;
> }
> 
> That would at least match the current usage where the inverse is what we
> need.
> 
> So SEV-SNP or TDX have a concept of a mask to apply to addresses from
> the guest? Can we steal any existing terms?
> 

In a general level, they are using "shared"/"private". TDX is using a
function kvm_gfn_shared_mask() to get the mask and three other marcos to
apply the mask on GPA (IPA)[1]. SEV-SNP is re-using SME macros e.g.
__sme_clr() to apply the mask[2].

Guess we can take them as an reference: using an inline function to get
the protected_bit_mask, like kvm_ipa_protected_mask() with the spec text
you pasted in the comment of the function. The name echoes the spec
description.

Then another necessary functions like kvm_gpa_{is, to}_{shared, private}
, which applies the mask to a GPA(IPA), to echo the terms in the generic
KVM knowledge. (Guess we can refine that with realm_is_addr_protected()).

[1] https://www.spinics.net/lists/kernel/msg4718104.html
[2] https://lore.kernel.org/lkml/20230220183847.59159-25-michael.roth@amd.com/

> 
> [1] Technically the spec only states: "Software in a Realm should treat
> the most significant bit of an IPA as a protection attribute." I don't
> think the bit is directly referred to in the spec as anything other than
> "the most significant bit". Although that in itself is confusing as it
> is the most significant *active* bit (i.e the configured IPA size
> changes which bit is used).
> 
> > Also, it would be nice to change the name of gpa_stolen_mask accordingly.
> >   
> >>  static inline bool vcpu_is_rec(struct kvm_vcpu *vcpu)
> >>  {
> >>  	if (static_branch_unlikely(&kvm_rme_is_available))
> >> diff --git a/arch/arm64/include/asm/kvm_rme.h b/arch/arm64/include/asm/kvm_rme.h
> >> index 9d1583c44a99..303e4a5e5704 100644
> >> --- a/arch/arm64/include/asm/kvm_rme.h
> >> +++ b/arch/arm64/include/asm/kvm_rme.h
> >> @@ -50,6 +50,18 @@ void kvm_destroy_rec(struct kvm_vcpu *vcpu);
> >>  int kvm_rec_enter(struct kvm_vcpu *vcpu);
> >>  int handle_rme_exit(struct kvm_vcpu *vcpu, int rec_run_status);
> >>  
> >> +void kvm_realm_unmap_range(struct kvm *kvm, unsigned long ipa, u64 size);
> >> +int realm_map_protected(struct realm *realm,
> >> +			unsigned long hva,
> >> +			unsigned long base_ipa,
> >> +			struct page *dst_page,
> >> +			unsigned long map_size,
> >> +			struct kvm_mmu_memory_cache *memcache);
> >> +int realm_map_non_secure(struct realm *realm,
> >> +			 unsigned long ipa,
> >> +			 struct page *page,
> >> +			 unsigned long map_size,
> >> +			 struct kvm_mmu_memory_cache *memcache);
> >>  int realm_set_ipa_state(struct kvm_vcpu *vcpu,
> >>  			unsigned long addr, unsigned long end,
> >>  			unsigned long ripas);
> >> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> >> index f29558c5dcbc..5417c273861b 100644
> >> --- a/arch/arm64/kvm/mmu.c
> >> +++ b/arch/arm64/kvm/mmu.c
> >> @@ -235,8 +235,13 @@ static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64
> >>  
> >>  	lockdep_assert_held_write(&kvm->mmu_lock);
> >>  	WARN_ON(size & ~PAGE_MASK);
> >> -	WARN_ON(stage2_apply_range(kvm, start, end, kvm_pgtable_stage2_unmap,
> >> -				   may_block));
> >> +
> >> +	if (kvm_is_realm(kvm))
> >> +		kvm_realm_unmap_range(kvm, start, size);
> >> +	else
> >> +		WARN_ON(stage2_apply_range(kvm, start, end,
> >> +					   kvm_pgtable_stage2_unmap,
> >> +					   may_block));
> >>  }
> >>  
> >>  static void unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size)
> >> @@ -250,7 +255,11 @@ static void stage2_flush_memslot(struct kvm *kvm,
> >>  	phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
> >>  	phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
> >>  
> >> -	stage2_apply_range_resched(kvm, addr, end, kvm_pgtable_stage2_flush);
> >> +	if (kvm_is_realm(kvm))
> >> +		kvm_realm_unmap_range(kvm, addr, end - addr);
> >> +	else
> >> +		stage2_apply_range_resched(kvm, addr, end,
> >> +					   kvm_pgtable_stage2_flush);
> >>  }
> >>  
> >>  /**
> >> @@ -818,6 +827,10 @@ void stage2_unmap_vm(struct kvm *kvm)
> >>  	struct kvm_memory_slot *memslot;
> >>  	int idx, bkt;
> >>  
> >> +	/* For realms this is handled by the RMM so nothing to do here */
> >> +	if (kvm_is_realm(kvm))
> >> +		return;
> >> +
> >>  	idx = srcu_read_lock(&kvm->srcu);
> >>  	mmap_read_lock(current->mm);
> >>  	write_lock(&kvm->mmu_lock);
> >> @@ -840,6 +853,7 @@ void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
> >>  	pgt = mmu->pgt;
> >>  	if (kvm_is_realm(kvm) &&
> >>  	    kvm_realm_state(kvm) != REALM_STATE_DYING) {
> >> +		unmap_stage2_range(mmu, 0, (~0ULL) & PAGE_MASK);
> >>  		write_unlock(&kvm->mmu_lock);
> >>  		kvm_realm_destroy_rtts(&kvm->arch.realm, pgt->ia_bits,
> >>  				       pgt->start_level);
> >> @@ -1190,6 +1204,24 @@ static bool kvm_vma_mte_allowed(struct vm_area_struct *vma)
> >>  	return vma->vm_flags & VM_MTE_ALLOWED;
> >>  }
> >>  
> >> +static int realm_map_ipa(struct kvm *kvm, phys_addr_t ipa, unsigned long hva,
> >> +			 kvm_pfn_t pfn, unsigned long map_size,
> >> +			 enum kvm_pgtable_prot prot,
> >> +			 struct kvm_mmu_memory_cache *memcache)
> >> +{
> >> +	struct realm *realm = &kvm->arch.realm;
> >> +	struct page *page = pfn_to_page(pfn);
> >> +
> >> +	if (WARN_ON(!(prot & KVM_PGTABLE_PROT_W)))
> >> +		return -EFAULT;
> >> +
> >> +	if (!realm_is_addr_protected(realm, ipa))
> >> +		return realm_map_non_secure(realm, ipa, page, map_size,
> >> +					    memcache);
> >> +
> >> +	return realm_map_protected(realm, hva, ipa, page, map_size, memcache);
> >> +}
> >> +
> >>  static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> >>  			  struct kvm_memory_slot *memslot, unsigned long hva,
> >>  			  unsigned long fault_status)
> >> @@ -1210,9 +1242,15 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> >>  	unsigned long vma_pagesize, fault_granule;
> >>  	enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
> >>  	struct kvm_pgtable *pgt;
> >> +	gpa_t gpa_stolen_mask = kvm_gpa_stolen_bits(vcpu->kvm);
> >>  
> >>  	fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
> >>  	write_fault = kvm_is_write_fault(vcpu);
> >> +
> >> +	/* Realms cannot map read-only */  
> > 
> > Out of curiosity, why? It would be nice to have more explanation in the
> > comment.  
> 
> The RMM specification doesn't support mapping protected memory read
> only. I don't believe there is any reason why it couldn't, but equally I
> don't think there any use cases for a guest needing read-only pages so
> this just isn't supported by the RMM. Since the page is necessarily
> taken away from the host it's fairly irrelevant (from the host's
> perspective) whether it is actually read only or not.
> 
> However, this is technically wrong for the case of unprotected (shared)
> pages - it should be possible to map those read only. But I need to have
> a think about how to fix that up.

If the fault IPA carries the protected bit, can't we do like:

if (vcpu_is_rec(vcpu) && fault_ipa_is_protected)
	write_fault = true

Are there still other gaps?
>
> >> +	if (vcpu_is_rec(vcpu))
> >> +		write_fault = true;
> >> +
> >>  	exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
> >>  	VM_BUG_ON(write_fault && exec_fault);
> >>  
> >> @@ -1272,7 +1310,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> >>  	if (vma_pagesize == PMD_SIZE || vma_pagesize == PUD_SIZE)
> >>  		fault_ipa &= ~(vma_pagesize - 1);
> >>  
> >> -	gfn = fault_ipa >> PAGE_SHIFT;
> >> +	gfn = (fault_ipa & ~gpa_stolen_mask) >> PAGE_SHIFT;
> >>  	mmap_read_unlock(current->mm);
> >>  
> >>  	/*
> >> @@ -1345,7 +1383,8 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> >>  	 * If we are not forced to use page mapping, check if we are
> >>  	 * backed by a THP and thus use block mapping if possible.
> >>  	 */
> >> -	if (vma_pagesize == PAGE_SIZE && !(force_pte || device)) {
> >> +	/* FIXME: We shouldn't need to disable this for realms */
> >> +	if (vma_pagesize == PAGE_SIZE && !(force_pte || device || kvm_is_realm(kvm))) {  
> > 
> > Why do we have to disable this temporarily?  
> 
> The current uABI (not using memfd) has some serious issues regarding
> huge page support. KVM normally follows the user space mappings of the
> memslot - so if user space has a huge page (transparent or hugetlbs)
> then stage 2 for the guest also gets one.
> 
> However realms sometimes require that the stage 2 differs. The main
> examples are:
> 
>  * RIPAS - if part of a huge page is RIPAS_RAM and part RIPAS_EMPTY then
> the huge page would have to be split.
> 
>  * Initially populated memory: basically the same as above - if the
> populated memory doesn't perfectly align with huge pages, then the
> head/tail pages would need to be broken up.
> 
> Removing this hack allows the huge pages to be created in stage 2, but
> then causes overmapping of the initial contents, then later on when the
> VMM (or guest) attempts to change the properties of the misaligned tail
> it gets an error because the pages are already present in stage 2.
> 
> The planned solution to all this is to stop following the user space
> page tables and create huge pages opportunistically from the memfd that
> backs the protected range. For now this hack exists to avoid things
> "randomly" failing when e.g. the initial kernel image isn't huge page
> aligned. In theory it should be possible to make this work with the
> current uABI, but it's not worth it when we know we're replacing it.

I see. Will dig it and see if there is any idea come to my mind.
> 
> >>  		if (fault_status == FSC_PERM && fault_granule > PAGE_SIZE)
> >>  			vma_pagesize = fault_granule;
> >>  		else
> >> @@ -1382,6 +1421,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> >>  	 */
> >>  	if (fault_status == FSC_PERM && vma_pagesize == fault_granule)
> >>  		ret = kvm_pgtable_stage2_relax_perms(pgt, fault_ipa, prot);
> >> +	else if (kvm_is_realm(kvm))
> >> +		ret = realm_map_ipa(kvm, fault_ipa, hva, pfn, vma_pagesize,
> >> +				    prot, memcache);
> >>  	else
> >>  		ret = kvm_pgtable_stage2_map(pgt, fault_ipa, vma_pagesize,
> >>  					     __pfn_to_phys(pfn), prot,
> >> @@ -1437,6 +1479,7 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
> >>  	struct kvm_memory_slot *memslot;
> >>  	unsigned long hva;
> >>  	bool is_iabt, write_fault, writable;
> >> +	gpa_t gpa_stolen_mask = kvm_gpa_stolen_bits(vcpu->kvm);
> >>  	gfn_t gfn;
> >>  	int ret, idx;
> >>  
> >> @@ -1491,7 +1534,7 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
> >>  
> >>  	idx = srcu_read_lock(&vcpu->kvm->srcu);
> >>  
> >> -	gfn = fault_ipa >> PAGE_SHIFT;
> >> +	gfn = (fault_ipa & ~gpa_stolen_mask) >> PAGE_SHIFT;
> >>  	memslot = gfn_to_memslot(vcpu->kvm, gfn);
> >>  	hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
> >>  	write_fault = kvm_is_write_fault(vcpu);
> >> @@ -1536,6 +1579,7 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
> >>  		 * of the page size.
> >>  		 */
> >>  		fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
> >> +		fault_ipa &= ~gpa_stolen_mask;
> >>  		ret = io_mem_abort(vcpu, fault_ipa);
> >>  		goto out_unlock;
> >>  	}
> >> @@ -1617,6 +1661,10 @@ bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
> >>  	if (!kvm->arch.mmu.pgt)
> >>  		return false;
> >>  
> > 
> > Does the unprotected (shared) region of a realm support aging?  
> 
> In theory this should be possible to support by unmapping the NS entry
> and handling the fault. But the hardware access flag optimisation isn't
> available with the RMM, and the overhead of RMI calls to unmap/map could
> be significant.
> 
> For now this isn't something we've looked at, but I guess it might be
> worth trying out when we have some real hardware to benchmark on.
> 
> >> +	/* We don't support aging for Realms */
> >> +	if (kvm_is_realm(kvm))
> >> +		return true;
> >> +
> >>  	WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
> >>  
> >>  	kpte = kvm_pgtable_stage2_mkold(kvm->arch.mmu.pgt,
> >> @@ -1630,6 +1678,10 @@ bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
> >>  	if (!kvm->arch.mmu.pgt)
> >>  		return false;
> >>  
> >> +	/* We don't support aging for Realms */
> >> +	if (kvm_is_realm(kvm))
> >> +		return true;
> >> +
> >>  	return kvm_pgtable_stage2_is_young(kvm->arch.mmu.pgt,
> >>  					   range->start << PAGE_SHIFT);
> >>  }
> >> diff --git a/arch/arm64/kvm/rme.c b/arch/arm64/kvm/rme.c
> >> index 3405b43e1421..3d46191798e5 100644
> >> --- a/arch/arm64/kvm/rme.c
> >> +++ b/arch/arm64/kvm/rme.c
> >> @@ -608,6 +608,54 @@ int realm_map_protected(struct realm *realm,
> >>  	return -ENXIO;
> >>  }
> >>  
> >> +int realm_map_non_secure(struct realm *realm,
> >> +			 unsigned long ipa,
> >> +			 struct page *page,
> >> +			 unsigned long map_size,
> >> +			 struct kvm_mmu_memory_cache *memcache)
> >> +{
> >> +	phys_addr_t rd = virt_to_phys(realm->rd);
> >> +	int map_level;
> >> +	int ret = 0;
> >> +	unsigned long desc = page_to_phys(page) |
> >> +			     PTE_S2_MEMATTR(MT_S2_FWB_NORMAL) |
> >> +			     /* FIXME: Read+Write permissions for now */  
> > Why can't we handle the prot from the realm_map_ipa()? Working in progress? :)  
> 
> Yes, work in progress - this comes from the "Realms cannot map
> read-only" in user_mem_abort() above. Since all faults are treated as
> write faults we need to upgrade to read/write here too.
> 
> The prot in realm_map_ipa isn't actually useful currently because we
> simply WARN_ON and return if it doesn't have PROT_W. Again this needs to
> be fixed! It's on my todo list ;)
> 
> Steve
> 
> >> +			     (3 << 6) |
> >> +			     PTE_SHARED;
> >> +
> >> +	if (WARN_ON(!IS_ALIGNED(ipa, map_size)))
> >> +		return -EINVAL;
> >> +
> >> +	switch (map_size) {
> >> +	case PAGE_SIZE:
> >> +		map_level = 3;
> >> +		break;
> >> +	case RME_L2_BLOCK_SIZE:
> >> +		map_level = 2;
> >> +		break;
> >> +	default:
> >> +		return -EINVAL;
> >> +	}
> >> +
> >> +	ret = rmi_rtt_map_unprotected(rd, ipa, map_level, desc);
> >> +
> >> +	if (RMI_RETURN_STATUS(ret) == RMI_ERROR_RTT) {
> >> +		/* Create missing RTTs and retry */
> >> +		int level = RMI_RETURN_INDEX(ret);
> >> +
> >> +		ret = realm_create_rtt_levels(realm, ipa, level, map_level,
> >> +					      memcache);
> >> +		if (WARN_ON(ret))
> >> +			return -ENXIO;
> >> +
> >> +		ret = rmi_rtt_map_unprotected(rd, ipa, map_level, desc);
> >> +	}
> >> +	if (WARN_ON(ret))
> >> +		return -ENXIO;
> >> +
> >> +	return 0;
> >> +}
> >> +
> >>  static int populate_par_region(struct kvm *kvm,
> >>  			       phys_addr_t ipa_base,
> >>  			       phys_addr_t ipa_end)  
> >   
> 

Powered by blists - more mailing lists