Re: [patch 1/7] cpusets: add dirty map to struct address_space

[Christoph Lameter <cl@linux-foundation.org>]

On Tue, 4 Nov 2008, Andrew Morton wrote:

>> +#ifdef CONFIG_CPUSETS
>> +#if MAX_NUMNODES <= BITS_PER_LONG
>> +	nodemask_t		dirty_nodes;	/* nodes with dirty pages */
>> +#else
>> +	nodemask_t		*dirty_nodes;	/* pointer to mask, if dirty */
>> +#endif
>> +#endif
>>  } __attribute__((aligned(sizeof(long))));
>
> eek.  Increasing the size of the address_space (and hence of the inode)
> is a moderately big deal - there can be millions of these in memory.

Well this is adding only a single word to the inode structure.

>> @@ -72,6 +72,8 @@ struct writeback_control {
>>  	 * so we use a single control to update them
>>  	 */
>>  	unsigned no_nrwrite_index_update:1;
>> +
>> +	nodemask_t *nodes;		/* Nodemask to writeback */
>
> This one doesn't get ifdefs?

The structure is typically allocated temporarily on the stack.

>> +	nodemask_t *nodes = mapping->dirty_nodes;
>> +	int node = page_to_nid(page);
>> +
>> +	if (!nodes) {
>> +		nodes = kmalloc(sizeof(nodemask_t), GFP_ATOMIC);
>
> erk, OK, called from __set_page_dirty, needs to be atomic.
>
> What are the consequences when this allocation fails?

Dirty tracking will not occur. All nodes are assumed to be dirty.

We discussed this earlier 
http://www.ussg.iu.edu/hypermail/linux/kernel/0709.1/2291.html