Date:   Mon, 31 Jan 2022 12:56:55 +0100
From:   David Hildenbrand <david@...hat.com>
To:     "Nikunj A. Dadhania" <nikunj@....com>,
        Paolo Bonzini <pbonzini@...hat.com>
Cc:     Sean Christopherson <seanjc@...gle.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Jim Mattson <jmattson@...gle.com>,
        Joerg Roedel <joro@...tes.org>,
        Brijesh Singh <brijesh.singh@....com>,
        Tom Lendacky <thomas.lendacky@....com>,
        Peter Gonda <pgonda@...gle.com>, kvm@...r.kernel.org,
        linux-kernel@...r.kernel.org, Bharata B Rao <bharata@....com>
Subject: Re: [RFC PATCH 3/6] KVM: SVM: Implement demand page pinning

On 28.01.22 12:08, David Hildenbrand wrote:
> On 28.01.22 12:04, Nikunj A. Dadhania wrote:
>> On 1/28/2022 1:57 PM, David Hildenbrand wrote:
>>> On 28.01.22 07:57, Nikunj A. Dadhania wrote:
>>>> On 1/26/2022 4:16 PM, David Hildenbrand wrote:
>>>>> On 18.01.22 12:06, Nikunj A Dadhania wrote:
>>>>>> Use the memslot metadata to store the pinned data along with the pfns.
>>>>>> This improves the SEV guest startup time from O(n) to a constant by
>>>>>> deferring guest page pinning until the pages are used to satisfy nested
>>>>>> page faults. The page reference will be dropped in the memslot free
>>>>>> path.
>>>>>>
>>>>>> Remove the enc_region structure definition and the code which did
>>>>>> upfront pinning, as they are no longer needed in view of the demand
>>>>>> pinning support.
>>>>>>
>>>>>> Leave svm_register_enc_region() and svm_unregister_enc_region() as stubs
>>>>>> since qemu is dependent on this API.
>>>>>>
>>>>>> Signed-off-by: Nikunj A Dadhania <nikunj@....com>
>>>>>> ---
>>>>>>  arch/x86/kvm/svm/sev.c | 208 ++++++++++++++++-------------------------
>>>>>>  arch/x86/kvm/svm/svm.c |   1 +
>>>>>>  arch/x86/kvm/svm/svm.h |   3 +-
>>>>>>  3 files changed, 81 insertions(+), 131 deletions(-)
>>>>>>
>>>>>> diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
>>>>>> index d972ab4956d4..a962bed97a0b 100644
>>>>>> --- a/arch/x86/kvm/svm/sev.c
>>>>>> +++ b/arch/x86/kvm/svm/sev.c
>>>>>> @@ -66,14 +66,6 @@ static unsigned int nr_asids;
>>>>>>  static unsigned long *sev_asid_bitmap;
>>>>>>  static unsigned long *sev_reclaim_asid_bitmap;
>>>>>>  
>>>>>> -struct enc_region {
>>>>>> -	struct list_head list;
>>>>>> -	unsigned long npages;
>>>>>> -	struct page **pages;
>>>>>> -	unsigned long uaddr;
>>>>>> -	unsigned long size;
>>>>>> -};
>>>>>> -
>>>>>>  /* Called with the sev_bitmap_lock held, or on shutdown  */
>>>>>>  static int sev_flush_asids(int min_asid, int max_asid)
>>>>>>  {
>>>>>> @@ -257,8 +249,6 @@ static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
>>>>>>  	if (ret)
>>>>>>  		goto e_free;
>>>>>>  
>>>>>> -	INIT_LIST_HEAD(&sev->regions_list);
>>>>>> -
>>>>>>  	return 0;
>>>>>>  
>>>>>>  e_free:
>>>>>> @@ -1637,8 +1627,6 @@ static void sev_migrate_from(struct kvm_sev_info *dst,
>>>>>>  	src->handle = 0;
>>>>>>  	src->pages_locked = 0;
>>>>>>  	src->enc_context_owner = NULL;
>>>>>> -
>>>>>> -	list_cut_before(&dst->regions_list, &src->regions_list, &src->regions_list);
>>>>>>  }
>>>>>>  
>>>>>>  static int sev_es_migrate_from(struct kvm *dst, struct kvm *src)
>>>>>> @@ -1861,115 +1849,13 @@ int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
>>>>>>  int svm_register_enc_region(struct kvm *kvm,
>>>>>>  			    struct kvm_enc_region *range)
>>>>>>  {
>>>>>> -	struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
>>>>>> -	struct enc_region *region;
>>>>>> -	int ret = 0;
>>>>>> -
>>>>>> -	if (!sev_guest(kvm))
>>>>>> -		return -ENOTTY;
>>>>>> -
>>>>>> -	/* If kvm is mirroring encryption context it isn't responsible for it */
>>>>>> -	if (is_mirroring_enc_context(kvm))
>>>>>> -		return -EINVAL;
>>>>>> -
>>>>>> -	if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
>>>>>> -		return -EINVAL;
>>>>>> -
>>>>>> -	region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
>>>>>> -	if (!region)
>>>>>> -		return -ENOMEM;
>>>>>> -
>>>>>> -	mutex_lock(&kvm->lock);
>>>>>> -	region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
>>>>>> -	if (IS_ERR(region->pages)) {
>>>>>> -		ret = PTR_ERR(region->pages);
>>>>>> -		mutex_unlock(&kvm->lock);
>>>>>> -		goto e_free;
>>>>>> -	}
>>>>>> -
>>>>>> -	region->uaddr = range->addr;
>>>>>> -	region->size = range->size;
>>>>>> -
>>>>>> -	list_add_tail(&region->list, &sev->regions_list);
>>>>>> -	mutex_unlock(&kvm->lock);
>>>>>> -
>>>>>> -	/*
>>>>>> -	 * The guest may change the memory encryption attribute from C=0 -> C=1
>>>>>> -	 * or vice versa for this memory range. Lets make sure caches are
>>>>>> -	 * flushed to ensure that guest data gets written into memory with
>>>>>> -	 * correct C-bit.
>>>>>> -	 */
>>>>>> -	sev_clflush_pages(region->pages, region->npages);
>>>>>> -
>>>>>> -	return ret;
>>>>>> -
>>>>>> -e_free:
>>>>>> -	kfree(region);
>>>>>> -	return ret;
>>>>>> -}
>>>>>> -
>>>>>> -static struct enc_region *
>>>>>> -find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
>>>>>> -{
>>>>>> -	struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
>>>>>> -	struct list_head *head = &sev->regions_list;
>>>>>> -	struct enc_region *i;
>>>>>> -
>>>>>> -	list_for_each_entry(i, head, list) {
>>>>>> -		if (i->uaddr == range->addr &&
>>>>>> -		    i->size == range->size)
>>>>>> -			return i;
>>>>>> -	}
>>>>>> -
>>>>>> -	return NULL;
>>>>>> -}
>>>>>> -
>>>>>> -static void __unregister_enc_region_locked(struct kvm *kvm,
>>>>>> -					   struct enc_region *region)
>>>>>> -{
>>>>>> -	sev_unpin_memory(kvm, region->pages, region->npages);
>>>>>> -	list_del(&region->list);
>>>>>> -	kfree(region);
>>>>>> +	return 0;
>>>>>>  }
>>>>>>  
>>>>>>  int svm_unregister_enc_region(struct kvm *kvm,
>>>>>>  			      struct kvm_enc_region *range)
>>>>>>  {
>>>>>> -	struct enc_region *region;
>>>>>> -	int ret;
>>>>>> -
>>>>>> -	/* If kvm is mirroring encryption context it isn't responsible for it */
>>>>>> -	if (is_mirroring_enc_context(kvm))
>>>>>> -		return -EINVAL;
>>>>>> -
>>>>>> -	mutex_lock(&kvm->lock);
>>>>>> -
>>>>>> -	if (!sev_guest(kvm)) {
>>>>>> -		ret = -ENOTTY;
>>>>>> -		goto failed;
>>>>>> -	}
>>>>>> -
>>>>>> -	region = find_enc_region(kvm, range);
>>>>>> -	if (!region) {
>>>>>> -		ret = -EINVAL;
>>>>>> -		goto failed;
>>>>>> -	}
>>>>>> -
>>>>>> -	/*
>>>>>> -	 * Ensure that all guest tagged cache entries are flushed before
>>>>>> -	 * releasing the pages back to the system for use. CLFLUSH will
>>>>>> -	 * not do this, so issue a WBINVD.
>>>>>> -	 */
>>>>>> -	wbinvd_on_all_cpus();
>>>>>> -
>>>>>> -	__unregister_enc_region_locked(kvm, region);
>>>>>> -
>>>>>> -	mutex_unlock(&kvm->lock);
>>>>>>  	return 0;
>>>>>> -
>>>>>> -failed:
>>>>>> -	mutex_unlock(&kvm->lock);
>>>>>> -	return ret;
>>>>>>  }
>>>>>>  
>>>>>>  int svm_vm_copy_asid_from(struct kvm *kvm, unsigned int source_fd)
>>>>>> @@ -2018,7 +1904,6 @@ int svm_vm_copy_asid_from(struct kvm *kvm, unsigned int source_fd)
>>>>>>  	mirror_sev->fd = source_sev->fd;
>>>>>>  	mirror_sev->es_active = source_sev->es_active;
>>>>>>  	mirror_sev->handle = source_sev->handle;
>>>>>> -	INIT_LIST_HEAD(&mirror_sev->regions_list);
>>>>>>  	ret = 0;
>>>>>>  
>>>>>>  	/*
>>>>>> @@ -2038,8 +1923,6 @@ int svm_vm_copy_asid_from(struct kvm *kvm, unsigned int source_fd)
>>>>>>  void sev_vm_destroy(struct kvm *kvm)
>>>>>>  {
>>>>>>  	struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
>>>>>> -	struct list_head *head = &sev->regions_list;
>>>>>> -	struct list_head *pos, *q;
>>>>>>  
>>>>>>  	WARN_ON(sev->num_mirrored_vms);
>>>>>>  
>>>>>> @@ -2066,18 +1949,6 @@ void sev_vm_destroy(struct kvm *kvm)
>>>>>>  	 */
>>>>>>  	wbinvd_on_all_cpus();
>>>>>>  
>>>>>> -	/*
>>>>>> -	 * if userspace was terminated before unregistering the memory regions
>>>>>> -	 * then lets unpin all the registered memory.
>>>>>> -	 */
>>>>>> -	if (!list_empty(head)) {
>>>>>> -		list_for_each_safe(pos, q, head) {
>>>>>> -			__unregister_enc_region_locked(kvm,
>>>>>> -				list_entry(pos, struct enc_region, list));
>>>>>> -			cond_resched();
>>>>>> -		}
>>>>>> -	}
>>>>>> -
>>>>>>  	sev_unbind_asid(kvm, sev->handle);
>>>>>>  	sev_asid_free(sev);
>>>>>>  }
>>>>>> @@ -2946,13 +2817,90 @@ void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
>>>>>>  	ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, 1);
>>>>>>  }
>>>>>>  
>>>>>> +void sev_pin_spte(struct kvm *kvm, gfn_t gfn, enum pg_level level,
>>>>>> +		  kvm_pfn_t pfn)
>>>>>> +{
>>>>>> +	struct kvm_arch_memory_slot *aslot;
>>>>>> +	struct kvm_memory_slot *slot;
>>>>>> +	gfn_t rel_gfn, pin_pfn;
>>>>>> +	unsigned long npages;
>>>>>> +	kvm_pfn_t old_pfn;
>>>>>> +	int i;
>>>>>> +
>>>>>> +	if (!sev_guest(kvm))
>>>>>> +		return;
>>>>>> +
>>>>>> +	if (WARN_ON_ONCE(is_error_noslot_pfn(pfn) || kvm_is_reserved_pfn(pfn)))
>>>>>> +		return;
>>>>>> +
>>>>>> +	/* Tested till 1GB pages */
>>>>>> +	if (KVM_BUG_ON(level > PG_LEVEL_1G, kvm))
>>>>>> +		return;
>>>>>> +
>>>>>> +	slot = gfn_to_memslot(kvm, gfn);
>>>>>> +	if (!slot || !slot->arch.pfns)
>>>>>> +		return;
>>>>>> +
>>>>>> +	/*
>>>>>> +	 * Use relative gfn index within the memslot for the bitmap as well as
>>>>>> +	 * the pfns array
>>>>>> +	 */
>>>>>> +	rel_gfn = gfn - slot->base_gfn;
>>>>>> +	aslot = &slot->arch;
>>>>>> +	pin_pfn = pfn;
>>>>>> +	npages = KVM_PAGES_PER_HPAGE(level);
>>>>>> +
>>>>>> +	/* Pin the page, KVM doesn't yet support page migration. */
>>>>>> +	for (i = 0; i < npages; i++, rel_gfn++, pin_pfn++) {
>>>>>> +		if (test_bit(rel_gfn, aslot->pinned_bitmap)) {
>>>>>> +			old_pfn = aslot->pfns[rel_gfn];
>>>>>> +			if (old_pfn == pin_pfn)
>>>>>> +				continue;
>>>>>> +
>>>>>> +			put_page(pfn_to_page(old_pfn));
>>>>>> +		}
>>>>>> +
>>>>>> +		set_bit(rel_gfn, aslot->pinned_bitmap);
>>>>>> +		aslot->pfns[rel_gfn] = pin_pfn;
>>>>>> +		get_page(pfn_to_page(pin_pfn));
>>>>>
>>>>>
>>>>> I assume this is to replace KVM_MEMORY_ENCRYPT_REG_REGION, which ends up
>>>>> calling svm_register_enc_region()->sev_pin_memory(), correct?
>>>>
>>>> Yes, that is correct.
>>>>>
>>>>> sev_pin_memory() correctly checks the RLIMIT_MEMLOCK and uses
>>>>> pin_user_pages_fast().
>>>>>
>>>>> I have to strongly assume that sev_pin_memory() is *wrong* as is because
>>>>> it's supposed to supply FOLL_LONGTERM -- after all we're pinning these
>>>>> pages possibly forever.
>>>>>
>>>>>
>>>>> I might be wrong but
>>>>>
>>>>> 1. You are missing the RLIMIT_MEMLOCK check
>>>>
>>>> Yes, I will add this check during the enc_region registration.
>>>>
>>>>> 2. get_page() is the wong way of long-term pinning a page. You would
>>>>> have to mimic what pin_user_pages_fast(FOLL_LONGTERM) does to eventually
>>>>> get it right (e.g., migrate the page off of MIGRATE_CMA or ZONE_MOVABLE).
>>>>
>>>> Let me go through this and I will come back. Thanks for pointing this out.
>>>
>>> I asusme the "issue" is that KVM uses mmu notifier and does a simple
>>> get_user_pages() to obtain the references, to drop the reference when
>>> the entry is invalidated via a mmu notifier call. So once you intent to
>>> long-term pin, it's already to late.
>>>
>>> If you could teach KVM to do a long-term pin when stumbling over these
>>> special encrypted memory regions (requires a proper matching
>>> unpin_user_pages() call from KVM), then you could "take over" that pin
>>> by get_page(), and let KVM do the ordinary put_page(), while you would
>>> do the unpin_user_pages().
>>>
>>
>> The fault path looks like this in KVM x86 mmu code:
>>
>> direct_page_fault()
>> -> kvm_faultin_pfn()
>>    -> __gfn_to_pfn_memslot()
>>       -> hva_to_pfn()
>>          -> hva_to_pfn_{slow,fast}()
>>             -> get_user_pages_*()      <<<<==== This is where the
>>                                                 reference is taken
>>
>> Next step is to create the mappings which is done in below functions:
>>
>> -> kvm_tdp_mmu_map() / __direct_map()
>>
>>    -> Within this function (patch 1/6), I call sev_pin_spte to take an extra 
>>       reference to pin it using get_page. 
>>
>>       Is it possible to use pin_user_pages(FOLL_LONGTERM) here? Wouldn't that 
>>       be equivalent to "take over" solution that you are suggesting?
>>
> 
> The issue is that pin_user_pages(FOLL_LONGTERM) might have to migrate
> the page, which will fail if there is already an additional reference
> from get_user_pages_*().
> 

Minor addition: hva_to_pfn_{slow,fast}() *don't* take a reference,
because we neither supply FOLL_GET nor FOLL_PIN. GUP users that rely on
memory notifiers don't require refernces.

I don't know what the implications would be if you FOLL_PIN |
FOLL_LONGTERM after already having a reference via
hva_to_pfn_{slow,fast}() in your hand in the callpath. Migration code
would effectively want to unmap the old page and call mmu notifiers to
properly invalidate the KVM MMU ...

In an ideal word, you'd really do a FOLL_PIN | FOLL_LONGTERM right away,
not doing the  get_user_pages_*()  first.

-- 
Thanks,

David / dhildenb

Powered by blists - more mailing lists