Re: [PATCH v3 01/10] KVM: arm64: Add snap shooting the host stage-2 pagetables

[Sebastian Ene]

On Tue, Nov 21, 2023 at 02:12:50PM +0000, Vincent Donnefort wrote:

Hi,

> On Wed, Nov 15, 2023 at 05:16:31PM +0000, Sebastian Ene wrote:
> > Introduce a new HVC that allows the caller to snap shoot the stage-2
> 
> "snapshot" same in the title of this patch.
> 

I will update the name, thanks.

> > pagetables under NVHE debug configuration. The caller specifies the
> > location where the pagetables are copied and must ensure that the memory
> > is accessible by the hypervisor. The memory where the pagetables are
> > copied has to be allocated by the caller and shared with the hypervisor.
> > 
> > Signed-off-by: Sebastian Ene <sebastianene@xxxxxxxxxx>
> > ---
> >  arch/arm64/include/asm/kvm_asm.h              |   1 +
> >  arch/arm64/include/asm/kvm_pgtable.h          |  36 +++++++
> >  arch/arm64/kvm/hyp/include/nvhe/mem_protect.h |   1 +
> >  arch/arm64/kvm/hyp/nvhe/hyp-main.c            |  20 ++++
> >  arch/arm64/kvm/hyp/nvhe/mem_protect.c         | 102 ++++++++++++++++++
> >  arch/arm64/kvm/hyp/pgtable.c                  |  56 ++++++++++
> >  6 files changed, 216 insertions(+)
> > 
> > diff --git a/arch/arm64/include/asm/kvm_asm.h b/arch/arm64/include/asm/kvm_asm.h
> > index 24b5e6b23417..99145a24c0f6 100644
> > --- a/arch/arm64/include/asm/kvm_asm.h
> > +++ b/arch/arm64/include/asm/kvm_asm.h
> > @@ -81,6 +81,7 @@ enum __kvm_host_smccc_func {
> >  	__KVM_HOST_SMCCC_FUNC___pkvm_init_vm,
> >  	__KVM_HOST_SMCCC_FUNC___pkvm_init_vcpu,
> >  	__KVM_HOST_SMCCC_FUNC___pkvm_teardown_vm,
> > +	__KVM_HOST_SMCCC_FUNC___pkvm_copy_host_stage2,
> >  };
> >  
> >  #define DECLARE_KVM_VHE_SYM(sym)	extern char sym[]
> > diff --git a/arch/arm64/include/asm/kvm_pgtable.h b/arch/arm64/include/asm/kvm_pgtable.h
> > index d3e354bb8351..be615700f8ac 100644
> > --- a/arch/arm64/include/asm/kvm_pgtable.h
> > +++ b/arch/arm64/include/asm/kvm_pgtable.h
> > @@ -10,6 +10,7 @@
> >  #include <linux/bits.h>
> >  #include <linux/kvm_host.h>
> >  #include <linux/types.h>
> > +#include <asm/kvm_host.h>
> >  
> >  #define KVM_PGTABLE_MAX_LEVELS		4U
> >  
> > @@ -351,6 +352,21 @@ struct kvm_pgtable {
> >  	kvm_pgtable_force_pte_cb_t		force_pte_cb;
> >  };
> >  
> > +/**
> > + * struct kvm_pgtable_snapshot - Snapshot page-table wrapper.
> > + * @pgtable:		The page-table configuration.
> > + * @mc:			Memcache used for pagetable pages allocation.
> > + * @pgd_hva:		Host virtual address of a physically contiguous buffer
> > + *			used for storing the PGD.
> > + * @pgd_len:		The size of the phyisically contiguous buffer in bytes.
> > + */
> > +struct kvm_pgtable_snapshot {
> > +	struct kvm_pgtable			pgtable;
> > +	struct kvm_hyp_memcache			mc;
> > +	void					*pgd_hva;
> > +	size_t					pgd_len;
> > +};
> > +
> >  /**
> >   * kvm_pgtable_hyp_init() - Initialise a hypervisor stage-1 page-table.
> >   * @pgt:	Uninitialised page-table structure to initialise.
> > @@ -756,4 +772,24 @@ enum kvm_pgtable_prot kvm_pgtable_hyp_pte_prot(kvm_pte_t pte);
> >   */
> >  void kvm_tlb_flush_vmid_range(struct kvm_s2_mmu *mmu,
> >  				phys_addr_t addr, size_t size);
> > +
> > +#ifdef CONFIG_NVHE_EL2_DEBUG
> > +/**
> > + * kvm_pgtable_stage2_copy() - Snapshot the pagetable
> > + *
> > + * @to_pgt:	Destination pagetable
> > + * @from_pgt:	Source pagetable. The caller must lock the pagetables first
> > + * @mc:		The memcache where we allocate the destination pagetables from
> > + */
> > +int kvm_pgtable_stage2_copy(struct kvm_pgtable *to_pgt,
> > +			    const struct kvm_pgtable *from_pgt,
> > +			    void *mc);
> > +#else
> > +static inline int kvm_pgtable_stage2_copy(struct kvm_pgtable *to_pgt,
> > +					  const struct kvm_pgtable *from_pgt,
> > +					  void *mc)
> > +{
> > +	return -EPERM;
> > +}
> > +#endif	/* CONFIG_NVHE_EL2_DEBUG */
> >  #endif	/* __ARM64_KVM_PGTABLE_H__ */
> > diff --git a/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h b/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h
> > index 0972faccc2af..9cfb35d68850 100644
> > --- a/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h
> > +++ b/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h
> > @@ -69,6 +69,7 @@ int __pkvm_host_donate_hyp(u64 pfn, u64 nr_pages);
> >  int __pkvm_hyp_donate_host(u64 pfn, u64 nr_pages);
> >  int __pkvm_host_share_ffa(u64 pfn, u64 nr_pages);
> >  int __pkvm_host_unshare_ffa(u64 pfn, u64 nr_pages);
> > +int __pkvm_host_stage2_prepare_copy(struct kvm_pgtable_snapshot *snapshot);
> 
> nit: the struct being a snapshot and that function doing the actual snapshot,
> maybe __pkvm_host_stage2_snapshot() would be a better name?
> 

Sure, we can use __pkvm_host_stage2_snapshot.

> >  
> >  bool addr_is_memory(phys_addr_t phys);
> >  int host_stage2_idmap_locked(phys_addr_t addr, u64 size, enum kvm_pgtable_prot prot);
> > diff --git a/arch/arm64/kvm/hyp/nvhe/hyp-main.c b/arch/arm64/kvm/hyp/nvhe/hyp-main.c
> > index 2385fd03ed87..98646cc67497 100644
> > --- a/arch/arm64/kvm/hyp/nvhe/hyp-main.c
> > +++ b/arch/arm64/kvm/hyp/nvhe/hyp-main.c
> > @@ -314,6 +314,25 @@ static void handle___pkvm_teardown_vm(struct kvm_cpu_context *host_ctxt)
> >  	cpu_reg(host_ctxt, 1) = __pkvm_teardown_vm(handle);
> >  }
> >  
> > +static void handle___pkvm_copy_host_stage2(struct kvm_cpu_context *host_ctxt)
> > +{
> > +#ifdef CONFIG_NVHE_EL2_DEBUG
> > +	int ret = -EPERM;
> > +	DECLARE_REG(struct kvm_pgtable_snapshot *, snapshot, host_ctxt, 1);
> > +	kvm_pteref_t pgd;
> > +
> > +	snapshot = kern_hyp_va(snapshot);
> > +	ret = __pkvm_host_stage2_prepare_copy(snapshot);
> > +	if (!ret) {
> > +		pgd = snapshot->pgtable.pgd;
> > +		snapshot->pgtable.pgd = (kvm_pteref_t)__hyp_pa(pgd);
> > +	}
> > +	cpu_reg(host_ctxt, 1) = ret;
> > +#else
> > +	cpu_reg(host_ctxt, 0) = SMCCC_RET_NOT_SUPPORTED;
> > +#endif
> > +}
> > +
> >  typedef void (*hcall_t)(struct kvm_cpu_context *);
> >  
> >  #define HANDLE_FUNC(x)	[__KVM_HOST_SMCCC_FUNC_##x] = (hcall_t)handle_##x
> > @@ -348,6 +367,7 @@ static const hcall_t host_hcall[] = {
> >  	HANDLE_FUNC(__pkvm_init_vm),
> >  	HANDLE_FUNC(__pkvm_init_vcpu),
> >  	HANDLE_FUNC(__pkvm_teardown_vm),
> > +	HANDLE_FUNC(__pkvm_copy_host_stage2),
> >  };
> >  
> >  static void handle_host_hcall(struct kvm_cpu_context *host_ctxt)
> > diff --git a/arch/arm64/kvm/hyp/nvhe/mem_protect.c b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
> > index 8d0a5834e883..1c3ab5ac9110 100644
> > --- a/arch/arm64/kvm/hyp/nvhe/mem_protect.c
> > +++ b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
> > @@ -266,6 +266,108 @@ int kvm_guest_prepare_stage2(struct pkvm_hyp_vm *vm, void *pgd)
> >  	return 0;
> >  }
> >  
> > +#ifdef CONFIG_NVHE_EL2_DEBUG
> > +static struct hyp_pool snapshot_pool = {0};
> > +static DEFINE_HYP_SPINLOCK(snapshot_pool_lock);
> > +
> > +static void *snapshot_zalloc_pages_exact(size_t size)
> > +{
> > +	void *addr = hyp_alloc_pages(&snapshot_pool, get_order(size));
> > +
> > +	hyp_split_page(hyp_virt_to_page(addr));
> > +
> > +	/*
> > +	 * The size of concatenated PGDs is always a power of two of PAGE_SIZE,
> > +	 * so there should be no need to free any of the tail pages to make the
> > +	 * allocation exact.
> > +	 */
> > +	WARN_ON(size != (PAGE_SIZE << get_order(size)));
> > +
> > +	return addr;
> > +}
> > +
> > +static void snapshot_get_page(void *addr)
> > +{
> > +	hyp_get_page(&snapshot_pool, addr);
> > +}
> > +
> > +static void *snapshot_zalloc_page(void *mc)
> > +{
> > +	struct hyp_page *p;
> > +	void *addr;
> > +
> > +	addr = hyp_alloc_pages(&snapshot_pool, 0);
> > +	if (addr)
> > +		return addr;
> > +
> > +	addr = pop_hyp_memcache(mc, hyp_phys_to_virt);
> > +	if (!addr)
> > +		return addr;
> > +
> > +	memset(addr, 0, PAGE_SIZE);
> > +	p = hyp_virt_to_page(addr);
> > +	memset(p, 0, sizeof(*p));
> 
> Why do you need to rease this entry of the vmemmap? That sounds odd.
> 

I guess I did in the same way as the guest does for
`guest_s2_zalloc_page`. Do you see any issues with this ?

> > +	p->refcount = 1;
> > +
> > +	return addr;
> > +}
> > +
> > +static void snapshot_s2_free_pages_exact(void *addr, unsigned long size)
> > +{
> > +	u8 order = get_order(size);
> > +	unsigned int i;
> > +	struct hyp_page *p;
> > +
> > +	for (i = 0; i < (1 << order); i++) {
> > +		p = hyp_virt_to_page(addr + (i * PAGE_SIZE));
> > +		hyp_page_ref_dec_and_test(p);
> 
> hyp_page_ref_dec() would be enough here.
> 

Yes, we can use hyp_page_ref_dec here, thanks.

> > +	}
> > +}
> > +
> > +int __pkvm_host_stage2_prepare_copy(struct kvm_pgtable_snapshot *snapshot)
> > +{
> > +	size_t required_pgd_len;
> > +	struct kvm_pgtable_mm_ops mm_ops = {0};
> > +	struct kvm_s2_mmu *mmu = &host_mmu.arch.mmu;
> > +	struct kvm_pgtable *to_pgt, *from_pgt = &host_mmu.pgt;
> > +	struct kvm_hyp_memcache *memcache = &snapshot->mc;
> > +	int ret;
> > +	void *pgd;
> 
> The shared memory between host and hyp should probably be pinned here even if it
> is behind a DEBUG option. That would also ensure that the memory has actually
> been shared beforehand and we'll not crash while running that. 
> 
> Pinning that memory would also make sure that the right amount has been shared
> for the pgd and we'll not overflow.
> 

After having the discussion offline I think it's better to change this
to the donation API from host -> hypervisor as the host doesn't need
share access while the snapshot is beeing done. 

> > +
> > +	required_pgd_len = kvm_pgtable_stage2_pgd_size(mmu->vtcr);
> > +	if (snapshot->pgd_len < required_pgd_len)
> > +		return -ENOMEM;
> > +
> > +	to_pgt = &snapshot->pgtable;
> > +	pgd = kern_hyp_va(snapshot->pgd_hva);
> > +
> > +	hyp_spin_lock(&snapshot_pool_lock);
> > +	hyp_pool_init(&snapshot_pool, hyp_virt_to_pfn(pgd),
> > +		      required_pgd_len / PAGE_SIZE, 0);
> > +
> > +	mm_ops.zalloc_pages_exact	= snapshot_zalloc_pages_exact;
> > +	mm_ops.zalloc_page		= snapshot_zalloc_page;
> > +	mm_ops.free_pages_exact		= snapshot_s2_free_pages_exact;
> > +	mm_ops.get_page			= snapshot_get_page;
> > +	mm_ops.phys_to_virt		= hyp_phys_to_virt;
> > +	mm_ops.virt_to_phys		= hyp_virt_to_phys;
> > +	mm_ops.page_count		= hyp_page_count;
> 
> It looks a bit odd to me to have to handle refcount while the lifespan of that
> newly create page-table only the lifespace of that function, where everything is
> behind a lock.
> 

I doubled checked on this and I don't think we need it so it can be
dropped, thanks.

> Is the need for the pool to handle the physically contiguous mem for the PGD? If
> that is the case, maybe you could avoid the pool by just passing as a separated
> argument the memory for the PGD and in another one the memcache for the other
> levels?
> 

I found it convinient to reuse it as we are doing the same thing when we
build the host stage-2.

> Especially IIUC, the pool would be used here only for the PGD allocation, as you
> add only the size of the PGD to the pool?
>

Yes, we only populate the pool with memory for the PGD. Actually I
wonder if we can just allocate everything from the pool and drop the
memcache.

> >+
> > +	to_pgt->ia_bits		= from_pgt->ia_bits;
> > +	to_pgt->start_level	= from_pgt->start_level;
> > +	to_pgt->flags		= from_pgt->flags;
> > +	to_pgt->mm_ops		= &mm_ops;
> > +
> > +	host_lock_component();
> > +	ret = kvm_pgtable_stage2_copy(to_pgt, from_pgt, memcache);
> > +	host_unlock_component();
> > +
> > +	hyp_spin_unlock(&snapshot_pool_lock);
> > +
> > +	return ret;
> > +}
> > +#endif /* CONFIG_NVHE_EL2_DEBUG */
> > +
> >  void reclaim_guest_pages(struct pkvm_hyp_vm *vm, struct kvm_hyp_memcache *mc)
> >  {
> >  	void *addr;
> > diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
> > index 1966fdee740e..46b15d74118f 100644
> > --- a/arch/arm64/kvm/hyp/pgtable.c
> > +++ b/arch/arm64/kvm/hyp/pgtable.c
> > @@ -1598,3 +1598,59 @@ void kvm_pgtable_stage2_free_unlinked(struct kvm_pgtable_mm_ops *mm_ops, void *p
> >  	WARN_ON(mm_ops->page_count(pgtable) != 1);
> >  	mm_ops->put_page(pgtable);
> >  }
> > +
> > +#ifdef CONFIG_NVHE_EL2_DEBUG
> > +static int stage2_copy_walker(const struct kvm_pgtable_visit_ctx *ctx,
> > +			      enum kvm_pgtable_walk_flags visit)
> > +{
> > +	struct kvm_pgtable_mm_ops *mm_ops = ctx->mm_ops;
> > +	void *copy_table, *original_addr;
> > +	kvm_pte_t new = ctx->old;
> > +
> > +	if (!stage2_pte_is_counted(ctx->old))
> > +		return 0;
> > +
> > +	if (kvm_pte_table(ctx->old, ctx->level)) {
> > +		copy_table = mm_ops->zalloc_page(ctx->arg);
> > +		if (!copy_table)
> > +			return -ENOMEM;
> > +
> > +		original_addr = kvm_pte_follow(ctx->old, mm_ops);
> > +
> > +		memcpy(copy_table, original_addr, PAGE_SIZE);
> > +		new = kvm_init_table_pte(copy_table, mm_ops);
> > +	}
> > +
> > +	*ctx->ptep = new;
> > +
> > +	return 0;
> > +}
> > +
> > +int kvm_pgtable_stage2_copy(struct kvm_pgtable *to_pgt,
> > +			    const struct kvm_pgtable *from_pgt,
> > +			    void *mc)
> > +{
> > +	int ret;
> > +	size_t pgd_sz;
> > +	struct kvm_pgtable_mm_ops *mm_ops = to_pgt->mm_ops;
> > +	struct kvm_pgtable_walker walker = {
> > +		.cb	= stage2_copy_walker,
> > +		.flags	= KVM_PGTABLE_WALK_LEAF |
> > +			  KVM_PGTABLE_WALK_TABLE_PRE,
> 
> Do you need the WALK_LEAF? Entire tables are copied and you don't need to change
> anything at the leaf level. Actually, stage2_copy_walker() writes new =
> ctx->old in that case, doesn't it?
>

We need it because we invoke the callback on the _PRE flag if it is a
table and we invoke it on the _LEAF if it is a block mapping.

> > +		.arg = mc
> > +	};
> > +
> > +	pgd_sz = kvm_pgd_pages(to_pgt->ia_bits, to_pgt->start_level) *
> > +		PAGE_SIZE;
> > +	to_pgt->pgd = (kvm_pteref_t)mm_ops->zalloc_pages_exact(pgd_sz);
> > +	if (!to_pgt->pgd)
> > +		return -ENOMEM;
> > +
> > +	memcpy(to_pgt->pgd, from_pgt->pgd, pgd_sz);
> > +
> > +	ret = kvm_pgtable_walk(to_pgt, 0, BIT(to_pgt->ia_bits), &walker);
> > +	mm_ops->free_pages_exact(to_pgt->pgd, pgd_sz);
> 
> > +
> > +	return ret;
> > +}
> > +#endif /* CONFIG_NVHE_EL2_DEBUG */
> > -- 
> > 2.43.0.rc0.421.g78406f8d94-goog
> >