Re: [RFC PATCH V4 01/13] netback: page pool version 1

[Eric Dumazet <eric.dumazet@gmail.com>]

Le jeudi 02 février 2012 à 16:49 +0000, Wei Liu a écrit :
> A global page pool. Since we are moving to 1:1 model netback, it is
> better to limit total RAM consumed by all the vifs.
> 
> With this patch, each vif gets page from the pool and puts the page
> back when it is finished with the page.
> 
> This pool is only meant to access via exported interfaces. Internals
> are subject to change when we discover new requirements for the pool.
> 
> Current exported interfaces include:
> 
> page_pool_init: pool init
> page_pool_destroy: pool destruction
> page_pool_get: get a page from pool
> page_pool_put: put page back to pool
> is_in_pool: tell whether a page belongs to the pool
> 
> Current implementation has following defects:
>  - Global locking
>  - No starve prevention mechanism / reservation logic
> 
> Global locking tends to cause contention on the pool. No reservation
> logic may cause vif to starve. A possible solution to these two
> problems will be each vif maintains its local cache and claims a
> portion of the pool. However the implementation will be tricky when
> coming to pool management, so let's worry about that later.
> 
> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> Tested-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> Signed-off-by: Wei Liu <wei.liu2@citrix.com>
> ---

Hmm, this kind of stuff should be discussed on lkml.

I doubt we want yet another memory allocator, with a global lock
(contended), and no NUMA properties.


> +int page_pool_init()
> +{
> +	int cpus = 0;
> +	int i;
> +
> +	cpus = num_online_cpus();
> +	pool_size = cpus * ENTRIES_PER_CPU;
> +
> +	pool = vzalloc(sizeof(struct page_pool_entry) * pool_size);
> +
> +	if (!pool)
> +		return -ENOMEM;
> +
> +	for (i = 0; i < pool_size - 1; i++)
> +		pool[i].u.fl = i+1;
> +	pool[pool_size-1].u.fl = INVALID_ENTRY;
> +	free_count = pool_size;
> +	free_head = 0;
> +
> +	return 0;
> +}
> +

num_online_cpus() disease once again.

code depending on num_online_cpus() is always suspicious.