[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel][PATCH] Intel EPT 1GB large page support



At 03:20 +0000 on 23 Feb (1266895248), Xu, Dongxiao wrote:
> EPT: 1GB large page support.
> 
> Alloc 1GB large page for EPT if possible. It also contains the logic to split
> large page into small ones (2M or 4k).
> 
> Signed-off-by: Dongxiao Xu <dongxiao.xu@xxxxxxxxx>
> Signed-off-by: Xiaohui Xin <xiaohui.xin@xxxxxxxxx>


> diff -r af2a0ed32ad5 xen/arch/x86/mm/hap/p2m-ept.c
> --- a/xen/arch/x86/mm/hap/p2m-ept.c   Sat Feb 20 19:25:07 2010 +0800
> +++ b/xen/arch/x86/mm/hap/p2m-ept.c   Sat Feb 20 19:43:49 2010 +0800
> @@ -183,17 +239,21 @@ ept_set_entry(struct domain *d, unsigned
>      int i;
>      int rv = 0;
>      int ret = 0;
> +    int split_level = 0;
>      int walk_level = order / EPT_TABLE_ORDER;
>      int direct_mmio = (p2mt == p2m_mmio_direct);
>      uint8_t ipat = 0;
>      int need_modify_vtd_table = 1;
>  
> -    /* We only support 4k and 2m pages now */
> -    BUG_ON(order && order != EPT_TABLE_ORDER);
> +    /* We support 4K, 2M, and 1G pages now */
> +    BUG_ON(order % EPT_TABLE_ORDER);

So order 27 allocations are OK? :)  Actually, now that I'm looking at
this BUG_ON, I can't see why it should be true even before this
change.  Is it enforced anywhere?

>      if (  order != 0 )
>          if ( (gfn & ((1UL << order) - 1)) )
>              return 1;
> +
> +    if ( order == 18 )
> +        printk(KERN_INFO "Got 1GB EPT entry.\n");

Definitely not. 

>      table = map_domain_page(mfn_x(pagetable_get_mfn(d->arch.phys_table)));
>  
> @@ -208,15 +268,21 @@ ept_set_entry(struct domain *d, unsigned
>              break;
>      }
>  
> -    /* If order == 9, we should never get SUPERPAGE or PoD.
> +    /* If order == 9 or 18, we should never get SUPERPAGE or PoD.

Why is that true?  Surely having introduced order-18 entries you can
now get SUPERPAGE when asking for an order-9 entry.

>       * If order == 0, we should only get POD if we have a POD superpage.
>       * If i > walk_level, we need to split the page; otherwise,
>       * just behave as normal. */
>      ASSERT(order == 0 || ret == GUEST_TABLE_NORMAL_PAGE);
>      ASSERT(ret != GUEST_TABLE_POD_PAGE || i != walk_level);
>  
> +    if ( i && i == walk_level && i <= cpu_vmx_ept_super_page_level_limit
> +           && ret == GUEST_TABLE_NORMAL_PAGE )
> +        ret = GUEST_TABLE_SUPER_PAGE;
> +

That needs a comment explaining it. 

>      index = gfn_remainder >> ( i ?  (i * EPT_TABLE_ORDER): order);
>      offset = (gfn_remainder & ( ((1 << (i*EPT_TABLE_ORDER)) - 1)));
> +
> +    split_level = i;
>  
>      ept_entry = table + index;
>  
> @@ -240,7 +306,7 @@ ept_set_entry(struct domain *d, unsigned
>  
>                  if ( (ept_entry->avail1 == p2m_ram_logdirty)
>                       && (p2mt == p2m_ram_rw) )
> -                    for ( i = 0; i < 512; i++ )
> +                    for ( i = 0; i < (1UL << order); i++ )

Gosh.  What's the performance of *that* like on live migration?

>                          paging_mark_dirty(d, mfn_x(mfn) - offset + i);
>              }
>              else
> @@ -261,51 +327,22 @@ ept_set_entry(struct domain *d, unsigned
>      }
>      else
>      {
> -        /* 
> -         * It's super page before, now set one of the 4k pages, so
> -         * we should split the 2m page to 4k pages now.
> -         */
> -        /* Pointers to / into new (split) middle-level table */
> -        ept_entry_t *split_table = NULL;
> -        ept_entry_t *split_ept_entry = NULL;
> -        /* Info about old (superpage) table */
> -        unsigned long super_mfn = ept_entry->mfn;
> -        p2m_type_t super_p2mt = ept_entry->avail1;
> -        /* The new l2 entry which we'll write after we've build the new l1 
> table */
> -        ept_entry_t l2_ept_entry;
> -
> -        /* 
> -         * Allocate new page for new ept middle level entry which is
> -         * before a leaf super entry
> -         */
> -        if ( !ept_set_middle_entry(d, &l2_ept_entry) )
> -            goto out;
> -
> -        /* Split the super page before to 4k pages */
> -        split_table = map_domain_page(l2_ept_entry.mfn);
> -        offset = gfn & ((1 << EPT_TABLE_ORDER) - 1);
> -
> -        for ( i = 0; i < 512; i++ )
> -        {
> -            split_ept_entry = split_table + i;
> -            split_ept_entry->emt = epte_get_entry_emt(d, gfn - offset + i,
> -                                                      _mfn(super_mfn + i),
> -                                                      &ipat, direct_mmio);
> -            split_ept_entry->ipat = ipat;
> -            split_ept_entry->sp_avail =  0;
> -            /* Don't increment mfn if it's a PoD mfn */
> -            if ( super_p2mt != p2m_populate_on_demand )
> -                split_ept_entry->mfn = super_mfn + i;
> -            else
> -                split_ept_entry->mfn = super_mfn; 
> -            split_ept_entry->avail1 = super_p2mt;
> -            split_ept_entry->avail2 = 0;
> -
> -            ept_p2m_type_to_flags(split_ept_entry, super_p2mt);
> -        }
> -
> -        /* Set the destinated 4k page as normal */
> -        split_ept_entry = split_table + offset;
> +        int num = order / EPT_TABLE_ORDER;
> +        int level;
> +        ept_entry_t *split_ept_entry;
> +    
> +        if ( num >= cpu_vmx_ept_super_page_level_limit )
> +            num = cpu_vmx_ept_super_page_level_limit;
> +        for ( level = split_level; level > num ; level-- )
> +        {
> +            rv = ept_split_large_page(d, &table, &index, gfn, level);
> +            if ( !rv )
> +                goto out;
> +        }
> +
> +        split_ept_entry = table + index;
> +        split_ept_entry->avail1 = p2mt;
> +        ept_p2m_type_to_flags(split_ept_entry, p2mt);
>          split_ept_entry->emt = epte_get_entry_emt(d, gfn, mfn, &ipat,
>                                                    direct_mmio);
>          split_ept_entry->ipat = ipat;
> @@ -314,12 +351,6 @@ ept_set_entry(struct domain *d, unsigned
>              need_modify_vtd_table = 0;
>          else
>              split_ept_entry->mfn = mfn_x(mfn);
> -
> -        split_ept_entry->avail1 = p2mt;
> -        ept_p2m_type_to_flags(split_ept_entry, p2mt);
> -
> -        unmap_domain_page(split_table);
> -        *ept_entry = l2_ept_entry;
>      }
>  
>      /* Track the highest gfn for which we have ever had a valid mapping */
> @@ -336,7 +367,7 @@ out:
>      ept_sync_domain(d);
>  
>      /* Now the p2m table is not shared with vt-d page table */
> -    if ( iommu_enabled && need_iommu(d) && need_modify_vtd_table )
> +    if ( rv && iommu_enabled && need_iommu(d) && need_modify_vtd_table )
>      {
>          if ( p2mt == p2m_ram_rw )
>          {
> @@ -459,7 +490,7 @@ out:
>  /* WARNING: Only caller doesn't care about PoD pages.  So this function will
>   * always return 0 for PoD pages, not populate them.  If that becomes 
> necessary,
>   * pass a p2m_query_t type along to distinguish. */
> -static ept_entry_t ept_get_entry_content(struct domain *d, unsigned long gfn)
> +static ept_entry_t ept_get_entry_content(struct domain *d, unsigned long 
> gfn, int *level)
>  {
>      ept_entry_t *table =
>          map_domain_page(mfn_x(pagetable_get_mfn(d->arch.phys_table)));
> @@ -487,6 +518,7 @@ static ept_entry_t ept_get_entry_content
>      index = gfn_remainder >> (i * EPT_TABLE_ORDER);
>      ept_entry = table + index;
>      content = *ept_entry;
> +    *level = i;
>  
>   out:
>      unmap_domain_page(table);
> @@ -579,7 +611,10 @@ void ept_change_entry_emt_with_range(str
>      p2m_lock(d->arch.p2m);
>      for ( gfn = start_gfn; gfn <= end_gfn; gfn++ )
>      {
> -        e = ept_get_entry_content(d, gfn);
> +        int level = 0;
> +        uint64_t trunk = 0;
> +
> +        e = ept_get_entry_content(d, gfn, &level);
>          if ( !p2m_has_emt(e.avail1) )
>              continue;
>  
> @@ -588,25 +623,23 @@ void ept_change_entry_emt_with_range(str
>  
>          if ( e.sp_avail )
>          {
> -            if ( !(gfn & ((1 << EPT_TABLE_ORDER) - 1)) &&
> -                 ((gfn + 0x1FF) <= end_gfn) )
> +            while ( level )
>              {
> -                /* 
> -                 * gfn assigned with 2M, and the end covers more than 2m 
> areas.
> -                 * Set emt for super page.
> -                 */
> -                order = EPT_TABLE_ORDER;
> -                if ( need_modify_ept_entry(d, gfn, mfn, e.ipat, e.emt, 
> e.avail1) )
> -                    ept_set_entry(d, gfn, mfn, order, e.avail1);
> -                gfn += 0x1FF;
> -            }
> -            else
> -            {
> -                /* Change emt for partial entries of the 2m area. */
> -                if ( need_modify_ept_entry(d, gfn, mfn, e.ipat, e.emt, 
> e.avail1) )
> -                    ept_set_entry(d, gfn, mfn, order, e.avail1);
> -                gfn = ((gfn >> EPT_TABLE_ORDER) << EPT_TABLE_ORDER) + 0x1FF;
> -            }
> +                trunk = (1UL << (level * EPT_TABLE_ORDER)) - 1;
> +                if ( !(gfn & trunk) && (gfn + trunk <= end_gfn) )
> +                {
> +                    /* gfn assigned with 2M, and the end covers more than 2m 
> areas.

This code covers 2M and 1G superpages; please fix the comment too.

> +                     * Set emt for super page.
> +                     */
> +                    order = level * EPT_TABLE_ORDER;
> +                    if ( need_modify_ept_entry(d, gfn, mfn, 
> +                          e.ipat, e.emt, e.avail1) )
> +                        ept_set_entry(d, gfn, mfn, order, e.avail1);
> +                    gfn += trunk;
> +                    break;
> +                }
> +                level--;
> +             }
>          }
>          else /* gfn assigned with 4k */
>          {

Cheers,

Tim.

-- 
Tim Deegan <Tim.Deegan@xxxxxxxxxx>
Principal Software Engineer, XenServer Engineering
Citrix Systems UK Ltd.  (Company #02937203, SL9 0BG)

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.