Re: [PATCH 2/6] qemu-kvm: Modify and introduce wrapper functions to access phys_ram_dirty.

[Avi Kivity <avi@redhat.com>]

On 03/16/2010 12:53 PM, Yoshiaki Tamura wrote:
> Modifies wrapper functions for byte-based phys_ram_dirty bitmap to
> bit-based phys_ram_dirty bitmap, and adds more wrapper functions to prevent
> direct access to the phys_ram_dirty bitmap.
>    

> +
> +static inline int cpu_physical_memory_get_dirty_flags(ram_addr_t addr)
> +{
> +    unsigned long mask;
> +    int index = (addr>>  TARGET_PAGE_BITS) / HOST_LONG_BITS;
> +    int offset = (addr>>  TARGET_PAGE_BITS)&  (HOST_LONG_BITS - 1);
> +    int ret = 0;
> +
> +    mask = 1UL<<  offset;
> +    if (phys_ram_vga_dirty[index]&  mask)
> +        ret |= VGA_DIRTY_FLAG;
> +    if (phys_ram_code_dirty[index]&  mask)
> +        ret |=  CODE_DIRTY_FLAG;
> +    if (phys_ram_migration_dirty[index]&  mask)
> +        ret |= MIGRATION_DIRTY_FLAG;
> +
> +    return ret;
>   }
>
>   static inline int cpu_physical_memory_get_dirty(ram_addr_t addr,
>                                                   int dirty_flags)
>   {
> -    return phys_ram_dirty[addr>>  TARGET_PAGE_BITS]&  dirty_flags;
> +    return cpu_physical_memory_get_dirty_flags(addr)&  dirty_flags;
>   }
>    

This turns one cacheline access into three.  If the dirty bitmaps were 
in an array, you could do

   return dirty_bitmaps[dirty_index][addr >> (TARGET_PAGE_BITS + 
BITS_IN_LONG)] & mask;

with one cacheline access.

>
>   static inline void cpu_physical_memory_set_dirty(ram_addr_t addr)
>   {
> -    phys_ram_dirty[addr>>  TARGET_PAGE_BITS] = 0xff;
> +    unsigned long mask;
> +    int index = (addr>>  TARGET_PAGE_BITS) / HOST_LONG_BITS;
> +    int offset = (addr>>  TARGET_PAGE_BITS)&  (HOST_LONG_BITS - 1);
> +
> +    mask = 1UL<<  offset;
> +    phys_ram_vga_dirty[index] |= mask;
> +    phys_ram_code_dirty[index] |= mask;
> +    phys_ram_migration_dirty[index] |= mask;
> +}
>    

This is also three cacheline accesses.  I think we should have a master 
bitmap which is updated by set_dirty(), and which is or'ed into the 
other bitmaps when they are accessed.  At least the vga and migration 
bitmaps are only read periodically, not randomly, so this would be very 
fast.  In a way, this is similar to how the qemu bitmap is updated from 
the kvm bitmap today.

I am not sure about the code bitmap though.


-- 
error compiling committee.c: too many arguments to function