On Thu, Apr 07, 2022, Vitaly Kuznetsov wrote:
> Currently, HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST{,EX} calls are handled
> the exact same way as HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE{,EX}: by
> flushing the whole VPID and this is sub-optimal. Switch to handling
> these requests with 'flush_tlb_gva()' hooks instead. Use the newly
> introduced TLB flush ring to queue the requests.
> 
> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
> ---
>  arch/x86/kvm/hyperv.c | 141 ++++++++++++++++++++++++++++++++++++------
>  1 file changed, 121 insertions(+), 20 deletions(-)
> 
> diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c
> index 81c44e0eadf9..a54d41656f30 100644
> --- a/arch/x86/kvm/hyperv.c
> +++ b/arch/x86/kvm/hyperv.c
> @@ -1792,6 +1792,35 @@ static u64 kvm_get_sparse_vp_set(struct kvm *kvm, struct kvm_hv_hcall *hc,
>  			      var_cnt * sizeof(*sparse_banks));
>  }
>  
> +static int kvm_hv_get_tlbflush_entries(struct kvm *kvm, struct kvm_hv_hcall *hc, u64 entries[],
> +				       u32 data_offset, int consumed_xmm_halves)

data_offset should be gpa_t, and the order of params should be consistent between
this and kvm_get_sparse_vp_set().

> +{
> +	int i;
> +
> +	if (hc->fast) {
> +		/*
> +		 * Each XMM holds two entries, but do not count halves that
> +		 * have already been consumed.
> +		 */
> +		if (hc->rep_cnt > (2 * HV_HYPERCALL_MAX_XMM_REGISTERS - consumed_xmm_halves))
> +			return -EINVAL;
> +
> +		for (i = 0; i < hc->rep_cnt; i++) {
> +			int j = i + consumed_xmm_halves;
> +
> +			if (j % 2)
> +				entries[i] = sse128_hi(hc->xmm[j / 2]);
> +			else
> +				entries[i] = sse128_lo(hc->xmm[j / 2]);
> +		}
> +
> +		return 0;
> +	}
> +
> +	return kvm_read_guest(kvm, hc->ingpa + data_offset,
> +			      entries, hc->rep_cnt * sizeof(entries[0]));

This is almost verbatim copy+pasted from kvm_get_sparse_vp_set().  If you slot in
the attached patched before this, then this function becomes:

static int kvm_hv_get_tlbflush_entries(struct kvm *kvm, struct kvm_hv_hcall *hc, u64 entries[],
				       int consumed_xmm_halves, gpa_t offset)
{
	return kvm_hv_get_hc_data(kvm, hc, hc->rep_cnt, hc->rep_cnt,
				  entries, consumed_xmm_halves, offset);
}


> +}

...

> @@ -1840,15 +1891,47 @@ void kvm_hv_vcpu_flush_tlb(struct kvm_vcpu *vcpu)
>  {
>  	struct kvm_vcpu_hv_tlbflush_ring *tlb_flush_ring;
>  	struct kvm_vcpu_hv *hv_vcpu = to_hv_vcpu(vcpu);
> -
> -	kvm_vcpu_flush_tlb_guest(vcpu);
> -
> -	if (!hv_vcpu)
> +	struct kvm_vcpu_hv_tlbflush_entry *entry;
> +	int read_idx, write_idx;
> +	u64 address;
> +	u32 count;
> +	int i, j;
> +
> +	if (!tdp_enabled || !hv_vcpu) {
> +		kvm_vcpu_flush_tlb_guest(vcpu);
>  		return;
> +	}
>  
>  	tlb_flush_ring = &hv_vcpu->tlb_flush_ring;
> +	read_idx = READ_ONCE(tlb_flush_ring->read_idx);
> +	write_idx = READ_ONCE(tlb_flush_ring->write_idx);
> +
> +	/* Pairs with smp_wmb() in hv_tlb_flush_ring_enqueue() */
> +	smp_rmb();
>  
> -	tlb_flush_ring->read_idx = tlb_flush_ring->write_idx;
> +	for (i = read_idx; i != write_idx; i = (i + 1) % KVM_HV_TLB_FLUSH_RING_SIZE) {
> +		entry = &tlb_flush_ring->entries[i];
> +
> +		if (entry->flush_all)
> +			goto out_flush_all;
> +
> +		/*
> +		 * Lower 12 bits of 'address' encode the number of additional
> +		 * pages to flush.
> +		 */
> +		address = entry->addr & PAGE_MASK;
> +		count = (entry->addr & ~PAGE_MASK) + 1;
> +		for (j = 0; j < count; j++)
> +			static_call(kvm_x86_flush_tlb_gva)(vcpu, address + j * PAGE_SIZE);
> +	}
> +	++vcpu->stat.tlb_flush;
> +	goto out_empty_ring;
> +
> +out_flush_all:
> +	kvm_vcpu_flush_tlb_guest(vcpu);
> +
> +out_empty_ring:
> +	tlb_flush_ring->read_idx = write_idx;
>  }
>  
>  static u64 kvm_hv_flush_tlb(struct kvm_vcpu *vcpu, struct kvm_hv_hcall *hc)
> @@ -1857,12 +1940,13 @@ static u64 kvm_hv_flush_tlb(struct kvm_vcpu *vcpu, struct kvm_hv_hcall *hc)
>  	struct hv_tlb_flush_ex flush_ex;
>  	struct hv_tlb_flush flush;
>  	DECLARE_BITMAP(vcpu_mask, KVM_MAX_VCPUS);
> +	u64 entries[KVM_HV_TLB_FLUSH_RING_SIZE - 2];

What's up with the -2?  And given the multitude of things going on in this code,
I'd strongly prefer this be tlbflush_entries.

Actually, if you do:

	u64 __tlbflush_entries[KVM_HV_TLB_FLUSH_RING_SIZE - 2];
	u64 *tlbflush_entries;

and drop all_addr, the code to get entries can be

	if (hc->code == HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE ||
	    hc->code == HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX ||
	    hc->rep_cnt > ARRAY_SIZE(tlbflush_entries)) {
		tlbfluish_entries = NULL;
	} else {
		if (kvm_hv_get_tlbflush_entries(kvm, hc, __tlbflush_entries,
						consumed_xmm_halves, data_offset))
			return HV_STATUS_INVALID_HYPERCALL_INPUT;
		tlbfluish_entries = __tlbflush_entries;
	}

and the calls to queue flushes becomes

			hv_tlb_flush_ring_enqueue(v, tlbflush_entries, hc->rep_cnt);

That way a bug will "just" be a NULL pointer dereference and not consumption of
uninitialized data (though such a bug might be caught be caught by the compiler).

>  	u64 valid_bank_mask;
>  	u64 sparse_banks[KVM_HV_MAX_SPARSE_VCPU_SET_BITS];
>  	struct kvm_vcpu *v;
>  	unsigned long i;
> -	bool all_cpus;
> -
> +	bool all_cpus, all_addr;
> +	int data_offset = 0, consumed_xmm_halves = 0;

data_offset should be a gpa_t.

>  	/*
>  	 * The Hyper-V TLFS doesn't allow more than 64 sparse banks, e.g. the
>  	 * valid mask is a u64.  Fail the build if KVM's max allowed number of

...

> +read_flush_entries:
> +	if (hc->code == HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE ||
> +	    hc->code == HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX ||
> +	    hc->rep_cnt > (KVM_HV_TLB_FLUSH_RING_SIZE - 2)) {

Rather than duplicate the -2 magic, it's far better to do:


> +		all_addr = true;
> +	} else {
> +		if (kvm_hv_get_tlbflush_entries(kvm, hc, entries,
> +						data_offset, consumed_xmm_halves))

As mentioned, the order for this call should match kvm_get_sparse_vp_set().

>  			return HV_STATUS_INVALID_HYPERCALL_INPUT;
> +		all_addr = false;
>  	}
>  
> -do_flush:
> +
>  	/*
>  	 * vcpu->arch.cr3 may not be up-to-date for running vCPUs so we can't
>  	 * analyze it here, flush TLB regardless of the specified address space.
>  	 */
>  	if (all_cpus) {
>  		kvm_for_each_vcpu(i, v, kvm)
> -			hv_tlb_flush_ring_enqueue(v);
> +			hv_tlb_flush_ring_enqueue(v, all_addr, entries, hc->rep_cnt);
>  
>  		kvm_make_all_cpus_request(kvm, KVM_REQ_HV_TLB_FLUSH);
>  	} else {
> @@ -1951,7 +2052,7 @@ static u64 kvm_hv_flush_tlb(struct kvm_vcpu *vcpu, struct kvm_hv_hcall *hc)
>  			v = kvm_get_vcpu(kvm, i);
>  			if (!v)
>  				continue;
> -			hv_tlb_flush_ring_enqueue(v);
> +			hv_tlb_flush_ring_enqueue(v, all_addr, entries, hc->rep_cnt);
>  		}
>  
>  		kvm_make_vcpus_request_mask(kvm, KVM_REQ_HV_TLB_FLUSH, vcpu_mask);
> -- 
> 2.35.1
>