Re: [PATCH v2 3/3] xen: use idle vcpus to scrub pages

[Bob Liu <bob.liu@oracle.com>]

On 07/24/2014 02:24 PM, Jan Beulich wrote:
>>>> On 24.07.14 at 04:08, <bob.liu@oracle.com> wrote:
> 
>> On 07/23/2014 03:28 PM, Jan Beulich wrote:
>>>>>> On 15.07.14 at 11:16, <bob.liu@oracle.com> wrote:
>>>> After so many days I haven't make a workable solution if don't remove
>>>> pages temporarily. The hardest part is iterating the heap free list
>>>> without holding heap_lock because if holding the lock it might be heavy
>>>> lock contention.
>>>> So do you think it's acceptable if fixed all other concerns about this
>>>> patch?
>>>
>>> No, I don't think so. Instead I'm of the opinion that you may have
>>> worked in the wrong direction: Rather than not taking the heap lock
>>> at all, it may also be sufficient to shrink the lock holding time (i.e.
>>> avoid long loops with the lock held).
>>>
>>
>> But I still think have to drop pages from heap list temporarily else
>> heap lock must be taken for a long time to get rid of E.g. below race
>> condition.
>>
>> A: alloc path            B: idle loop
>>
>>                         spin_lock(&heap_lock)
>>                       page_list_for_each( pg, &heap(node, zone, order) )
>>                         if _PGC_need_scrub is set, break;
>>                         spin_unlock(&heap_lock)
>>
>>                         if ( test_bit(_PGC_need_scrub, pg)
>>
>> ^^^^
>> spin_lock(&heap_lock)
>> delist page
>> spin_unlock(&heap_lock)
>>
>> write data to this page
>>
>>                          scrub_one_page(pg)
>>                          ^^^ will clean useful data
> 
> No (and I'm sure I said so before): The only problem is with the
> linked list itself; the page contents are not a problem - the
> allocation path can simply wait for the already suggested
> _PGC_scrubbing flag to clear before returning. And as already

The page contents are a problem if the race condition I mentioned in
previous email happen.

Because there is a time window between checking the PGC_need_scrub flag
and doing the real scrub in idle thread, the idle thread will still
scrub a page after that page have been allocated by allocation path and
been used(and have been written some useful data).

> said (see above), by avoiding page_list_for_each() within the
> locked region you already significantly reduce lock contention.
> I.e. you need another means to find pages awaiting to be

Right, but I don't have better ideas beside using page_list_for_each()
or delist pages temporarily from heap list.

> scrubbed. You may want to leverage that the allocation path
> does the scrubbing if needed (e.g. by not scrubbing the first
> of any set of contiguous free pages on the idle path, linking up
> all other ones recognizing that their link fields are unused while
> on an order-greater-than-zero free list).

It sounds like another list have to be introduced, but I don't think
this can help to get rid of lock contention or the similar race condition.

-- 
Regards,
-Bob