Re: [PATCH 8/9] KVM: X86: Optimize pte_list_desc with per-array counter

[Peter Xu]

On Thu, Jun 24, 2021 at 02:15:20PM -0400, Peter Xu wrote:
> Add a counter field into pte_list_desc, so as to simplify the add/remove/loop
> logic.  E.g., we don't need to loop over the array any more for most reasons.
> 
> This will make more sense after we've switched the array size to be larger
> otherwise the counter will be a waste.
> 
> Initially I wanted to store a tail pointer at the head of the array list so we
> don't need to traverse the list at least for pushing new ones (if without the
> counter we traverse both the list and the array).  However that'll need
> slightly more change without a huge lot benefit, e.g., after we grow entry
> numbers per array the list traversing is not so expensive.
> 
> So let's be simple but still try to get as much benefit as we can with just
> these extra few lines of changes (not to mention the code looks easier too
> without looping over arrays).
> 
> I used the same a test case to fork 500 child and recycle them ("./rmap_fork
> 500" [1]), this patch further speeds up the total fork time of about 14%, which
> is a total of 38% of vanilla kernel:
> 
>         Vanilla:      367.20 (+-4.58%)
>         3->15 slots:  302.00 (+-5.30%)
>         Add counter:  265.20 (+-9.88%)
> 
> [1] https://github.com/xzpeter/clibs/commit/825436f825453de2ea5aaee4bdb1c92281efe5b3
> 
> Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> ---
>  arch/x86/kvm/mmu/mmu.c | 26 ++++++++++++++++----------
>  1 file changed, 16 insertions(+), 10 deletions(-)
> 
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index 8888ae291cb9..b21e52dfc27b 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -136,10 +136,15 @@ module_param(dbg, bool, 0644);
>  #include <trace/events/kvm.h>
>  
>  /* make pte_list_desc fit well in cache lines */
> -#define PTE_LIST_EXT 15
> +#define PTE_LIST_EXT 14
>  
>  struct pte_list_desc {
>  	u64 *sptes[PTE_LIST_EXT];
> +	/*
> +	 * Stores number of entries stored in the pte_list_desc.  No need to be
> +	 * u64 but just for easier alignment.  When PTE_LIST_EXT, means full.
> +	 */
> +	u64 spte_count;
>  	struct pte_list_desc *more;
>  };
>  
> @@ -830,7 +835,7 @@ static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
>  			struct kvm_rmap_head *rmap_head)
>  {
>  	struct pte_list_desc *desc;
> -	int i, count = 0;
> +	int count = 0;
>  
>  	if (!rmap_head->val) {
>  		rmap_printk("%p %llx 0->1\n", spte, *spte);
> @@ -840,24 +845,24 @@ static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
>  		desc = mmu_alloc_pte_list_desc(vcpu);
>  		desc->sptes[0] = (u64 *)rmap_head->val;
>  		desc->sptes[1] = spte;
> +		desc->spte_count = 2;
>  		rmap_head->val = (unsigned long)desc | 1;
>  		++count;
>  	} else {
>  		rmap_printk("%p %llx many->many\n", spte, *spte);
>  		desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
> -		while (desc->sptes[PTE_LIST_EXT-1]) {
> +		while (desc->spte_count == PTE_LIST_EXT) {
>  			count += PTE_LIST_EXT;
> -
>  			if (!desc->more) {
>  				desc->more = mmu_alloc_pte_list_desc(vcpu);
>  				desc = desc->more;
> +				desc->spte_count = 0;
>  				break;
>  			}
>  			desc = desc->more;
>  		}
> -		for (i = 0; desc->sptes[i]; ++i)
> -			++count;
> -		desc->sptes[i] = spte;
> +		count += desc->spte_count;
> +		desc->sptes[desc->spte_count++] = spte;
>  	}
>  	return count;
>  }
> @@ -873,8 +878,10 @@ pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
>  		;
>  	desc->sptes[i] = desc->sptes[j];
>  	desc->sptes[j] = NULL;
> +	desc->spte_count--;
>  	if (j != 0)
>  		return;
> +	WARN_ON_ONCE(desc->spte_count);
>  	if (!prev_desc && !desc->more)
>  		rmap_head->val = 0;
>  	else
> @@ -930,7 +937,7 @@ static void pte_list_remove(struct kvm_rmap_head *rmap_head, u64 *sptep)
>  unsigned int pte_list_count(struct kvm_rmap_head *rmap_head)
>  {
>  	struct pte_list_desc *desc;
> -	unsigned int i, count = 0;
> +	unsigned int count = 0;
>  
>  	if (!rmap_head->val)
>  		return 0;
> @@ -940,8 +947,7 @@ unsigned int pte_list_count(struct kvm_rmap_head *rmap_head)
>  	desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
>  
>  	while (desc) {
> -		for (i = 0; (i < PTE_LIST_EXT) && desc->sptes[i]; i++)
> -			count++;
> +		count += desc->spte_count;
>  		desc = desc->more;
>  	}

I think I still missed another loop in pte_list_desc_remove_entry() that we can
drop.  With some other cleanups, I plan to squash below into this patch too..

---8<---
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index 719fb6fd0aa0..2d8c56eb36f8 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -872,16 +872,13 @@ pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
                           struct pte_list_desc *desc, int i,
                           struct pte_list_desc *prev_desc)
 {
-       int j;
+       int j = desc->spte_count - 1;
 
-       for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
-               ;
        desc->sptes[i] = desc->sptes[j];
        desc->sptes[j] = NULL;
        desc->spte_count--;
-       if (j != 0)
+       if (desc->spte_count)
                return;
-       WARN_ON_ONCE(desc->spte_count);
        if (!prev_desc && !desc->more)
                rmap_head->val = 0;
        else
@@ -913,7 +910,7 @@ static void __pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head)
                desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
                prev_desc = NULL;
                while (desc) {
-                       for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
+                       for (i = 0; i < desc->spte_count; ++i) {
                                if (desc->sptes[i] == spte) {
                                        pte_list_desc_remove_entry(rmap_head,
                                                        desc, i, prev_desc);
---8<---

-- 
Peter Xu