Re: [PATCHv6 3/3] xen/privcmd: improve performance of MMAPBATCH_V2

[David Vrabel <david.vrabel@citrix.com>]

On 09/03/15 12:25, Stefano Stabellini wrote:
> On Fri, 6 Mar 2015, David Vrabel wrote:
>> 
>> +/*
>> + * Similar to traverse_pages, but use each page as a "block" of
>> + * data to be processed as one unit.
>> + */
>> +static int traverse_pages_block(unsigned nelem, size_t size,
>> +				struct list_head *pos,
>> +				int (*fn)(void *data, int nr, void *state),
>> +				void *state)
>> +{
>> +	void *pagedata;
>> +	unsigned pageidx;
>> +	int ret = 0;
>> +
>> +	BUG_ON(size > PAGE_SIZE);
> 
> I looks like that PAGE_SIZE needs to be a multiple of size. Maybe we can
> add a BUG_ON for that too.

There is no such requirement because...

>> +	pageidx = PAGE_SIZE;
>> +
>> +	while (nelem) {
>> +		int nr = (PAGE_SIZE/size);

...the number of elements per pages is rounded down here.

>> +		struct page *page;
>> +		if (nr > nelem)
>> +			nr = nelem;
>> +		pos = pos->next;
>> +		page = list_entry(pos, struct page, lru);
>> +		pagedata = page_address(page);
>> +		ret = (*fn)(pagedata, nr, state);
>> +		if (ret)
>> +			break;
>> +		nelem -= nr;
>> +	}
>> +
>> +	return ret;
>> +}

David