Re: [PATCH] x86/memshr: fix preemption in relinquish_shared_pages()

[Andrew Cooper <andrew.cooper3@citrix.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1763 bytes --]

On 17/12/13 09:07, Jan Beulich wrote:
> For one, should hypercall_preempt_check() return false the first time
> it gets called, it would never have got called again (because count,
> being checked for equality, didn't get reset to zero).
>
> And then, if there were a huge range of unshared pages, with count not
> getting incremented at all in that case there would also not be any
> preemption.
>
> Fix this by using a biased increment (ratio 1:16 for unshared vs shared
> pages), and check for certain bits in count to be clear rather than for
> a specific value (the alternative would be to do a greater-or-equal
> comparison and flush the value to zero in case of a "false" return from
> hypercall_preempt_check()).
>
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>

>
> --- a/xen/arch/x86/mm/mem_sharing.c
> +++ b/xen/arch/x86/mm/mem_sharing.c
> @@ -1290,11 +1290,13 @@ int relinquish_shared_pages(struct domai
>              set_rc = p2m->set_entry(p2m, gfn, _mfn(0), PAGE_ORDER_4K,
>                                      p2m_invalid, p2m_access_rwx);
>              ASSERT(set_rc != 0);
> -            count++;
> +            count += 0x10;
>          }
> +        else
> +            ++count;
>  
> -        /* Preempt every 2MiB. Arbitrary */
> -        if ( (count == 512) && hypercall_preempt_check() )
> +        /* Preempt every 2MiB (shared) or 32MiB (unshared) - arbitrary. */
> +        if ( !(count & 0x1ff0) && hypercall_preempt_check() )
>          {
>              p2m->next_shared_gfn_to_relinquish = gfn + 1;
>              rc = -EAGAIN;
>
>
>
>
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel


[-- Attachment #1.2: Type: text/html, Size: 2645 bytes --]

[-- Attachment #2: Type: text/plain, Size: 126 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel