Re: [PATCH v2 4/4] x86/shadow: refactor shadow_vram_{get,put}_l1e()

[Jan Beulich <jbeulich@suse.com>]

On 10.10.2020 09:45, Roger Pau Monné wrote:
> On Thu, Oct 08, 2020 at 04:36:47PM +0100, Andrew Cooper wrote:
>> On 08/10/2020 16:15, Roger Pau Monné wrote:
>>> On Wed, Sep 16, 2020 at 03:08:40PM +0200, Jan Beulich wrote:
>>>> +void shadow_vram_put_mfn(mfn_t mfn, unsigned int l1f,
>>>> +                         mfn_t sl1mfn, const void *sl1e,
>>>> +                         const struct domain *d)
>>>> +{
>>>> +    unsigned long gfn;
>>>> +    struct sh_dirty_vram *dirty_vram = d->arch.hvm.dirty_vram;
>>>> +
>>>> +    ASSERT(is_hvm_domain(d));
>>>> +
>>>> +    if ( !dirty_vram /* tracking disabled? */ ||
>>>> +         !(l1f & _PAGE_RW) /* read-only mapping? */ ||
>>>> +         !mfn_valid(mfn) /* mfn can be invalid in mmio_direct */)
>>>> +        return;
>>>> +
>>>> +    gfn = gfn_x(mfn_to_gfn(d, mfn));
>>>> +    /* Page sharing not supported on shadow PTs */
>>>> +    BUG_ON(SHARED_M2P(gfn));
>>>> +
>>>> +    if ( (gfn >= dirty_vram->begin_pfn) && (gfn < dirty_vram->end_pfn) )
>>>> +    {
>>>> +        unsigned long i = gfn - dirty_vram->begin_pfn;
>>>> +        const struct page_info *page = mfn_to_page(mfn);
>>>> +        bool dirty = false;
>>>> +        paddr_t sl1ma = mfn_to_maddr(sl1mfn) | PAGE_OFFSET(sl1e);
>>>> +
>>>> +        if ( (page->u.inuse.type_info & PGT_count_mask) == 1 )
>>>> +        {
>>>> +            /* Last reference */
>>>> +            if ( dirty_vram->sl1ma[i] == INVALID_PADDR )
>>>> +            {
>>>> +                /* We didn't know it was that one, let's say it is dirty */
>>>> +                dirty = true;
>>>> +            }
>>>> +            else
>>>> +            {
>>>> +                ASSERT(dirty_vram->sl1ma[i] == sl1ma);
>>>> +                dirty_vram->sl1ma[i] = INVALID_PADDR;
>>>> +                if ( l1f & _PAGE_DIRTY )
>>>> +                    dirty = true;
>>>> +            }
>>>> +        }
>>>> +        else
>>>> +        {
>>>> +            /* We had more than one reference, just consider the page dirty. */
>>>> +            dirty = true;
>>>> +            /* Check that it's not the one we recorded. */
>>>> +            if ( dirty_vram->sl1ma[i] == sl1ma )
>>>> +            {
>>>> +                /* Too bad, we remembered the wrong one... */
>>>> +                dirty_vram->sl1ma[i] = INVALID_PADDR;
>>>> +            }
>>>> +            else
>>>> +            {
>>>> +                /*
>>>> +                 * Ok, our recorded sl1e is still pointing to this page, let's
>>>> +                 * just hope it will remain.
>>>> +                 */
>>>> +            }
>>>> +        }
>>>> +
>>>> +        if ( dirty )
>>>> +        {
>>>> +            dirty_vram->dirty_bitmap[i / 8] |= 1 << (i % 8);
>>> Could you use _set_bit here?
>>
>> __set_bit() uses 4-byte accesses.  This uses 1-byte accesses.
> 
> Right, this is allocated using alloc directly, not the bitmap helper,
> and the size if rounded to byte level, not unsigned int.
> 
>> Last I checked, there is a boundary issue at the end of the dirty_bitmap.
>>
>> Both Julien and I have considered changing our bit infrastructure to use
>> byte accesses, which would make them more generally useful.
> 
> Does indeed seem useful to me, as we could safely expand the usage of
> the bitmap ops without risking introducing bugs.

Aren't there architectures being handicapped when it comes to sub-word
accesses? At least common code may better not make assumptions towards
more fine grained accesses ...

As to x86, couldn't we make the macros evaluate alignof(*(addr)) and
choose byte-based accesses when it's less than 4?

Jan